NONLINEAR FINITE ELEMENT METHODS

Carlos A. Felippa
Department of Aerospace Engineering Sciences and
Center for Space Structures and Controls
University of Colorado
Boulder, Colorado 80309-0429, USA

August 2001

Material assembled from lecture notes for the course Nonlinear Finite Elements Methods, offered
since 1987 to date at the Aerospace Engineering Sciences Department of the University of Colorado
at Boulder.
 Preface
This textbook presents an Introduction to the computer-based simulation of nonlinear structures by
the Finite Element Method (FEM). It assembles the still “unconverged” lecture notes of Nonlinear
Finite Element Methods or NFEM. This is an advanced graduate course offered in the Aerospace
Engineering Sciences of the University of Colorado at Boulder.
NFEM was first taught on the Spring Semester 1986 and has been repeated every two or three years.
Unlike the Introduction to Finite Element Methods (IFEM), NFEM is not a core course. It is
typicall taken by second year graduate students that are interested in the topic of nonlinear simulation
of mechanical systems.
Prerequisites for the course are an introductory course in finite elements such as IFEM, graduate-level
calculus, linear algebra, knowledge of structural mechanics at the Mechanics of Materials level, and
ability to program in a higher level language such as Matlab or Mathematica.
The course originally used Fortran 77 as computer implementation language. This has been gradually
changed to Mathematica since 1995. Unlike IFEM the changeover is not yet complete since the course
has been offered only twice since.

Book Objectives

(To be completed)

Book Organization

(To be completed)

Exercises

Each Chapter is followed by a list of homework exercises that pose problems of varying difficulty.
Each exercise is labeled by a tag of the form
[type:rating]
The type is indicated by letters A, C, D or N for exercises to be answered primarily by analytical
work, computer programming, descriptive narration, and numerical calculations, respectively. Some
exercises involve a combination of these traits, in which case a combination of letters separated
by + is used, e.g., A+N. For some problems heavy analytical work may be helped by the use of a
computer-algebra system, in which case the type is identified as A/C.
The rating is a number between 5 and 50 that estimates the degree of difficulty of an Exercise, in the
following “logarithmic” scale:
5 A simple question that can be answered in seconds, or is already answered in the text if the
student has read and understood the material.
10 A straightforward question that can be answered in minutes.
15 A relatively simple question that requires some thinking, and may take on the order of half to
one hour to answer.

i
20 Either a problem of moderate difficulty, or a straightforward one requiring lengthy computations
or some programming, normally taking one to six hours of work.
25 A scaled up version of the above, estimated to require six hours to one day of work.
30 A problem of moderate difficulty that normally requires on the order of one or two days of work.
Arriving at the answer may involve a combination of techniques, some background or reference
material, or lenghty but straightforward programming.
40 A difficult problem that may be solvable only by gifted and well prepared individual students,
or a team. Difficulties may be due to the need of correct formulation, advanced mathematics,
or high level programming. With the proper preparation, background and tools these problems
may be solved in hours or days, while remaining inaccessible to unprepared or average students.
50 A research problem, worthy of publication if solved.
Most Exercises have a rating of 15 or 20. Assigning three or four per week puts a load of roughly 5-10
hours of solution work, plus the time needed to prepare the answer material. Assignments of difficulty
25 or 30 are better handled by groups, or given in take-home exams. Assignments of difficulty over
30 are never assigned in the course, but provided as a challenge for an elite group.
Occasionally an Exercise has two or more distinct but related parts identified as items. In that case a
rating may be given for each item. For example: [A/C:15+20]. This does not mean that the exercise
as a whole has a difficulty of 35, because the scale is roughly logarithmic; the numbers simply rate
the expected effort per item.

Selecting Course Material

(To be completed)

Acknowledgements

Thanks are due to students and colleagues who have provided valuable feedback on the original course
Notes, and helped metamorphosis into a textbook. Two invigorating sabbaticals in 1993 and 2001
provided blocks of time to develop, reformat and integrate material. The hospitality of Dr. P. G.
Bergan of Det Norske Veritas at Oslo, Norway and Professor E. Oñate of CIMNE/UPC at Barcelona,
Spain, during those sabbaticals is gratefully acknowleged.

ii
iii
 Chapter Contents
Section
1 . . . . . . . . . . . . . . . . . . 1-1
2 . . . . . . . . . . . . . . . . . . 2-1
3 . . . . . . . . . . . . . . . . . . 3-1
4 . . . . . . . . . . . . . . . . . . 4-1
5 . . . . . . . . . . . . . . . . . . 5-1
6 . . . . . . . . . . . . . . . . . . 6-1
7 . . . . . . . . . . . . . . . . . . 7-1
8 . . . . . . . . . . . . . . . . . . 8-1
9 . . . . . . . . . . . . . . . . . . 9-1
10 . . . . . . . . . . . . . . . . . . 10-1
11 . . . . . . . . . . . . . . . . . . 11-1
12 . . . . . . . . . . . . . . . . . . 12-1
13 . . . . . . . . . . . . . . . . . . 13-1
14 . . . . . . . . . . . . . . . . . . 14-1
15 . . . . . . . . . . . . . . . . . . 15-1
16 . . . . . . . . . . . . . . . . . . 16-1
17 . . . . . . . . . . . . . . . . . . 17-1
18 . . . . . . . . . . . . . . . . . . 18-1
19 . . . . . . . . . . . . . . . . . . 19-1
20 . . . . . . . . . . . . . . . . . . 20-1
21 . . . . . . . . . . . . . . . . . . 21-1
22 . . . . . . . . . . . . . . . . . . 22-1
23 . . . . . . . . . . . . . . . . . . 23-1
24 . . . . . . . . . . . . . . . . . . 24-1
23 . . . . . . . . . . . . . . . . . . 23-1
24 . . . . . . . . . . . . . . . . . . 24-1
25 . . . . . . . . . . . . . . . . . . 25-1
26 . . . . . . . . . . . . . . . . . . 26-1
27 . . . . . . . . . . . . . . . . . . 27-1
28 . . . . . . . . . . . . . . . . . . 28-1
Appendices
A . . . . . . . . . . . . . . . . . . A-1
B . . . . . . . . . . . . . . . . . . B-1
C . . . . . . . . . . . . . . . . . . C-1
D . . . . . . . . . . . . . . . . . . D-1
H . . . . . . . . . . . . . . . . . . H-1

iv
v
 14
.

The CR Formulation:
Space Bar

14–1
Chapter 14: THE CR FORMULATION: SPACE BAR 14–2

TABLE OF CONTENTS

Page
§14.1. Introduction 14–3
§14.2. Line Segment Moving in 3D 14–3
§14.2.1. Line Segment Derivatives . . . . . . . . . . . . . 14–4
§14.2.2. Derivatives of Length Functions . . . . . . . . . . . 14–5
§14.2.3. Mathematica Implementation and FD Verification . . . . . 14–6
§14.3. The CR Bar Element 14–9
§14.3.1. Internal Energy, Force and Stiffness . . . . . . . . . 14–9
§14.3.2. Matrices for Specific Strain Measures . . . . . . . . . 14–10
§14.3.3. Mathematica Implementation . . . . . . . . . . . 14–11
§14. Exercises . . . . . . . . . . . . . . . . . . . . . . 14–14

14–2
14–3 §14.2 LINE SEGMENT MOVING IN 3D

§14.1. Introduction

The expressions provided in the foregoing two Chapters and in Appendix R address the formation of
incremental equations for an arbitrary CR element. The main restriction is the assumption of small
local deformations, which permit the use of the linearized equations for the deformational energy.
Closed form of the incremental expressions become intractable for elements of arbitrary geometry,
such as curved shells or beams, because c and R are complicated functions of the element displacement
field, which is turn determined by u. Fortunately the CR approach is often used with elements of
simple geometry in which rotational freedoms, if any, may be ignored in defining the corotated
configuration. Under those conditions one can work out the base-to-deformed transformation arrays
directly from geometric arguments. The transformations may be then systematically applied to
existing linear elements through a modular interface, as illustrated in Figure 12.7.
In particular, for a simplex element (a constant strain element without rotational freedoms) it is
possible to work out all transformation from the intrinsic geometry of a line segment, triangle or
tetrahedron moving in 3D space. (By “intrinsic” is meant changes in edge lengths, face areas and
volume dimensions.) For the line segment modeling a space bar the formulas are worked out and
collected in the next sections. These apply to several types of finite elements, such as bars and cables.
The results may be specialized to two dimensions, if desired, by setting the third coordinate to zero.
It is important that the kinematic analysis be exact so that arbitrary rigid body motions can be
accomodated. Restrictions on local deformations in the motion from C R to C D can then be made
when considering specific elements, particularly those endowed with rotational DOFs.
Several of the following results are new. Their closed form derivation was made possible because of
the use of Mathematica to synthesize abtruse algebraic expressions containing symbolic terms.

§14.2. Line Segment Moving in 3D

uX2
Deformed uZ2
configuration CD 2 (x 2 ,y 2 ,z 2 )
1 (x1 ,y1 ,z 1 ) deformed length a
uX1 uZ1
Y, y uY2
uY1
10 (X1 ,Y1 ,Z 1 ) base length a 0
X, x
Z, z Global system
Base configuration C 0 2 0 (X2 ,Y2 ,Z2 )

Figure 14.1. A line segment moving in 3D space.

Consider the line segment shown in Figure 14.1, defined by the end nodes 1-2. The segment moves
in three-dimensional space. The global axes will be denoted by {X, Y, Z } instead of {X 1 , X 2 , X 3 } so

14–3
Chapter 14: THE CR FORMULATION: SPACE BAR 14–4

a z = z21
a 2= a x2 + a y2 + a2z 2
a y = y21

1 a x = x21
Y, y

aX0 = X21
10
X, x aY0 = Y21
Z, z Global system a02 = 2
aX0 + 2
a Y0+ 2
a Z0
20
aZ0 = Z 21

Figure 14.2. Line segment components.

that {X n , Yn , Z n } can be used for the coordinates of node n. Global axes {x, y, z}, which are used
for the deformed configuration, coalesce with {X, Y, Z },
The base line configuration C0 is specified by the coordinates (X 1 , Y1 , Z 1 ) and (X 2 , Y2 , Z 2 ) of the
line end nodes. The line moves to the deformed (current) configuration C D of length a defined by
coordinates {x1 = X 1 +u X 1 , y1 = Y1 +u Y 1 , z 1 = Z 1 +u Z 1 } and {x2 = X 2 +u X 2 , y2 = Y2 +u Y 2 , z 2 =
Z 2 + u Z 2 }, where u X 1 through u Z 2 are the node displacements. Node coordinate and displacement
differences are abbreviated by X 21 = X 2 − X 1 , x21 = x2 − x1 , u X 21 = u X 2 − u X 1 , etc. As illustrated
in Figure 14.2, the line lengths are given by
a02 = a 2X 0 + aY2 0 + a 2Z 0 = X 21
2
+ Y21
2
+ Z 21
2
, a 2 = ax2 + a y2 + az2 = x21
2
+ y21
2
+ z 21
2
. (14.1)
For further use define the following vectors
         
aX0 1 aX0 ax 1 ax u X 21
a0 = aY 0 , â0 = aY 0 , a = a y , â = a y , u21 = a − a0 = u Y 21 , (14.2)
aZ 0 a0 a a a a u Z 21
Z0 x z

Here â0 = a0 /a0 and â = a/a denote the direction cosine vectors of the base and current line
segment, respectively. This “hat convention” will be used to identify direction vectors normalized to
unit length. It is important not to confuse the 3-vector u21 with the 6-vector of node displacements
u = [ u X1 uY 1 u Z1 u X2 uY 2 u Z 2 ]T (14.3)
In fact, ∂u21 /∂u is the 3 × 6 matrix [ −I I ], where I is the 3 × 3 identity matrix.
§14.2.1. Line Segment Derivatives
Suppose that the node displacements are functions of two variables, generically denoted by  and
: u X 1 = u X 1 (, ), etc. The partial derivative of a with respect to  is
 ∂u 
X 21

∂a ax ∂u X 21 a y ∂u Y 21 az ∂u Z 21  ∂ 
 
= + + = â u21, , in which u21, =  ∂u Y 21  . (14.4)
T
∂ a ∂ a ∂ a ∂  ∂ 
∂u Z 21
∂
14–4
14–5 §14.2 LINE SEGMENT MOVING IN 3D

Note that the base configuration is only “remembered” through the displacements since the initial
length a0 does not appear explicitly.
To obtain the second derivative we take the partial of (14.4) with respect to the generic variable ,
which yields âT ∂ 2 u21 /(∂∂ ) + (∂ âT /∂ ) (∂u21 /∂). While the first term is easy, the second one
requires the derivatives of â. Since this involves the variation of a fixed-length (unit) vector, it can
be expected to involve the orthogonal projector associated with a. The final result can be presented
in the compact matrix form

∂ 2a
= u21,
T
H u21, + âT u21, , (14.5)
∂∂
in which
 
∂ 2 u X 21
 
 ∂∂  a y2 + az2 −ax a y −ax az
 2  1  1 1
u21, = ∂ u Y 21 , H= −a y ax az2 + ax2 −a y az  = (I − ââ ) = Pa . (14.6)
T
 ∂∂  a 3 a a
  −az ax −az a y ax2 + a y2
∂ 2 u Z 21
∂∂
Here I denotes the identity matrix of order 3 whereas Pa = I − ââT is the orthogonal projector
associated with the direction a. (To show this, square Pa and verify that Pa2 = Pa .)
The results (14.4) and (14.5) can be specialized to various choices. For example, if the displacements
are viewed as functions of real time (or a time-like parameter) t, we have  ≡ ≡ t, and

ȧ = âT u̇21 , ä = âT ü21 + u̇21

T
H u̇21 (14.7)

where a superposed dot denotes derivative with respect to t. Of more interest for element derivation
is to set  and in turn to the six entries of the node displacement vector u arranged as (14.3). In
this case the term u21, in (14.5) vanishes, and we obtain


∂a 1 def ∂ 2a H −H def
= [ −ax −a y −az ax a y az ]T = [ −âT âT ]T = h, = = G.
∂u a ∂u ∂u −H H
(14.8)
Vector h and matrix G will appear in the derivation of the space bar element in §14..3. In addition
the following definition and relations are useful there:


def I −I 1
J = = aG + h hT , G = J − h hT . (14.9)
−I I a

§14.2.2. Derivatives of Length Functions

Often the derivatives of a function F(a) are needed, for example for bar strain measures other than
engineering strains. The function F(a) is assumed twice differentiable respect to a. The a-derivative
abbreviations are
∂ F(a) ∂ 2 F(a)
F,a = , F,aa = , (14.10)
∂a ∂a 2

14–5
Chapter 14: THE CR FORMULATION: SPACE BAR 14–6

(Partials are used since F could be a function of other variables in addition to a.) Using the previous
results and the chain rule we obtain
∂ F(a) ∂ F(a) ∂a
= = F,a âT u21, ,
∂ ∂a ∂
∂ 2 F(a) ∂ F,a ∂a T ∂ 2a 
= â u21, + F,a = u21,
T
F,aa ââT + F,a H u21, + F,a âT u21,
∂∂ ∂a ∂ ∂∂
(14.11)
For F(a) = a the previous results are recovered. The case of powers of a: F(a) = a n has immediate
applications. For n = 2, with  and specialized first to t and then to u, we obtain the squared-length
derivatives:
˙ ¨
a 2 = 2a â u̇21 , a 2 = 2u̇T u̇21 + 2a ü21 (14.12)


∂a 2
∂ a
2 2
I −I
= 2a [ −âT âT ]T = 2a h, =2 = 2J. (14.13)
∂u ∂u ∂u −I I
The simplicity of the Hessian of a 2 will be a cogent argument for the use of the Green measure
of strain in bars. For n = −1, with  and specialized as above, we obtain the inverse-length
derivatives:
˙ = −(1/a 2 ) â u̇,
1/a ¨ = (1/a 3 ) u̇T (3â âT − I) u̇ − (1/a 2 ) ü
1/a (14.14)
21 21 21


∂(1/a) ∂ 2 (1/a) 3â âT − I −3â âT + I
= −(1/a ) [ −â â ] ,
2
= (1/a )3
(14.15)
∂u ∂u ∂u −3â âT + I 3â âT − I
The foregoing equations may be specialized to two-dimensional motions of a segment moving in the
{X, Y } plane by setting the Z component to zero and then removing that component from vectors
and matrices.
All of the results given so far are geometrically exact and pose no limit on how much the segment
stretches or contracts. Restrictions in the form of small deformations will appear when the formulas
are applied to a bar element in §14.3.

§14.2.3. Mathematica Implementation and FD Verification

The foregoing formulas for the first and second derivatives of an arbitrary function F(a), given in
(14.11), have been implemented in Mathematica in the form of two modules. Results are numerically
verified by finite differences.
Scripts for the computation of first partial derivatives of F(a) with respect to  are shown in Fig-
ure 14.3. This is done by module LengthFunctionFirstDerivatives. The results when  is
identified with the node displacement vector are verified with central finite differences with module
LengthFunctionFirstDerivativesByFD. The driver code is shown at the bottom of the figure.
√
The computations are exercised for F(a) set to a, a 2 , 1/a, 1/a 2 , a and log(a) in the loop shown at
the bottom of Figure 14.3. The coordinate and displacement values used for numerical verification
are

(a X , aY , a Z ) = (11, 10, 2), (u X 1 , u Y 1 , u Z 1 ) = (3, −5, −5), (u X 2 , u Y 2 , u Z 2 ) = (8, −4, 1),

(ax , a y , az ) = (a X , aY , a Z )+(u X 2 , u Y 2 , u Z 2 ) − (u X 1 , u Y 1 , u Z 1 ) = (16, 11, 8), (14.16)
   
a0 = a 2X +aY2 +a 2Z = 112 +102 +22 = 15, a= ax2 +a y2 +az2 = 162 +112 +82 = 21.

14–6
14–7 §14.2 LINE SEGMENT MOVING IN 3D

LengthFunctionFirstDerivative[{fa_,a_},aXYZ_,u21_,Φ_]:=Module[
{nu=Length[u21],nΦ=Length[Φ],axyz,u21Φ}, axyz=aXYZ+u21;
u21Φ=Table[D[u21[[i]],Φ[[j]]],{i,1,nu},{j,1,nΦ}];
Return[(D[fa,a]*axyz/a).u21Φ] ];

LengthFunctionFirstDerivativeByFD[{fa_,a_},aXYZ_,u21_,δ _]:=
Module[{ax,ay,az,axyz,axyzp,axyzm,axp,ayp,azp,axm,aym,azm,
i,inc,dfada,da,d=Table[0,{6}]}, {ax,ay,az}=axyz=aXYZ+u21;
dfada=D[fa,a]/.a->Sqrt[ax^2+ay^2+az^2];
inc= δ*{{-1,0,0},{0,-1,0},{0,0,-1},{1,0,0},{0,1,0},{0,0,1}};
For [i=1,i<=6,i++, axyz=aXYZ+u21;
axyzp=axyz+inc[[i]]; axyzm=axyz-inc[[i]];
{axp,ayp,azp}=axyzp; {axm,aym,azm}=axyzm;
da=Sqrt[axp^2+ayp^2+azp^2]-Sqrt[axm^2+aym^2+azm^2];
d[[i]]=dfada*da/(2*δ );
]; Return[d]];

ClearAll[a,axyz,a,aX,aY,aZ,uX1,uY1,uZ1,uX2,uY2,uZ2,Φ,δ]; δ =1/100;
rep={aX->11,aY->10,aZ->2,uX1->3,uY1->-5,uZ1->-5,uX2->8,uY2->-4,
uZ2->1,a0->15,a->21};
aXYZ={aX,aY,aZ}; u21={uX2-uX1,uY2-uY1,uZ2-uZ1};
axyz=aXYZ+u21; {ax,ay,az}=axyz;
Print["aXYZ=",aXYZ/.rep," u21=",u21/.rep," axyz=",axyz/.rep,
" a0=",Sqrt[aX^2+aY^2+aZ^2]/.rep," a=",Sqrt[ax^2+ay^2+az^2]/.rep];
u={uX1,uY1,uZ1,uX2,uY2,uZ2};
For [if=1,if<=6,if++, fa={a,a^2,1/a,1/a^2,Sqrt[a],Log[a]}[[if]];
d=Simplify[LengthFunctionFirstDerivative[{fa,a},aXYZ,u21,u]];
Print["fa= ",fa,", d=",d,"\n = ",N[d/.rep]];
dFD=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ/.rep,u21/.rep, δ];
Print["d by FD= ", N[dFD/.rep]];
];

Figure 14.3. Mathematica implementation of first-derivative computations in (14.11) for F(a) and
numerical check by central finite differences.

fa= a , d=
aX − uX1 + uX2 aY − uY1 + uY2 aZ − uZ1 + uZ2 aX − uX1 + uX2 aY − uY1 + uY2 aZ − uZ1 + uZ2
− ,− ,− , , ,
2 a3/ 2 2 a3/2 2 a3/2 2 a3/2 2 a3/2 2 a3/2
= − 0.0831306, − 0.0571523, − 0.0415653, 0.0831306, 0.0571523, 0.0415653

d by FD= − 0.0831306, − 0.0571523, − 0.0415653, 0.0831306, 0.0571523, 0.0415653

√
Figure 14.4. Partial results from script of Figure 14.3 for F(a) = a and numerical values (14.16).
Displacement increment δ = 0.01 used for verification by central finite differences.

√
Figure 14.4 shows results for the case F(a) = a. Numerical results correspond to the data (14.16).
It can be seen that the finite difference values, obtained with a displacement increment of δ = 0.01,
agree to all places shown with the analytical results.
Scripts for the computation of second partial derivative of F(a) with respect to  and are shown
in Figure 14.5. This is done by module LengthFunctionSecondDerivatives. The results when
both  and are identified with the node displacement vector are verified through finite differences
with module LengthFunctionSecondDerivativesByFD. The driver code is√ shown at the bottom
of the figure. The computations are exercised for F(a) set to a, a 2 , 1/a, 1/a 2 , a and log(a) in the
loop shown at the bottom of Figure 14.5.
√
Figure 14.6 shows numerical results for F(a) = a and the data (14.16). Symbolic results are
omitted to reduce clutter. It can be seen that the finite difference values, obtained with a displacement
increment of δ = 0.01, agree to at least 5 places with the analytical results.

14–7
Chapter 14: THE CR FORMULATION: SPACE BAR 14–8

LengthFunctionSecondDerivative[{fa_,a_},aXYZ_,u21_,{Φ_,Ψ_}]:=Module[
{nc=Length[aXYZ],nu=Length[u21],nΦ=Length[Φ],nΨ=Length[Ψ],
u21Φ,u21Ψ,J,H,P,S,Q,dfda,d2fdada}, axyz=aXYZ+u21; S=Q=Table[0,{6},{6}];
u21Φ=Table[D[u21[[i]],Φ[[j]]],{i,1,nu},{j,1,nΦ}];
u21Ψ=Table[D[u21[[i]],Ψ[[j]]],{i,1,nu},{j,1,nΨ}];
J=Table[axyz[[i]]*axyz[[j]]/a^2,{i,1,nc},{j,1,nc}];
H=(IdentityMatrix[nc]-J)/a; Q=Table[0,{nΦ},{nΨ}];
For [i=1,i<=nc,i++, Q=Q+(axyz[[i]]/a)*
Table[D[u21[[i]],Φ[[j]],Ψ[[k]]],{j,1,nΦ},{k,1,nΨ}] ];
dfda=D[fa,a]; d2fdada=D[dfda,a];
S=Transpose[u21Φ].(d2fdada*J+dfda*H).u21Ψ;
Return[{S,Q}];
];

LengthFunctionSecondDerivativeByFD[{fa_,a_},aXYZ_,u21_,δ _]:=
Module[{aX,aY,aZ,uX21,uY21,uZ21,ax,ay,az,
i,inc,du,S=Table[0,{6},{6}]},
inc= δ*{{-1,0,0},{0,-1,0},{0,0,-1},{1,0,0},{0,1,0},{0,0,1}};
For [i=1,i<=6,i++, du=inc[[i]];
dp=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ,u21+du,δ ];
dm=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ,u21-du,δ ];
S[[i]]=(dp-dm)/(2* δ );
]; Return[S]];

ClearAll[a,n,axyz,a,aX,aY,aZ,uX1,uY1,uZ1,uX2,uY2,uZ2,Φ, δ]; δ =1/100;

rep={aX->11,aY->10,aZ->2,uX1->3,uY1->-5,uZ1->-5,uX2->8,uY2->-4,uZ2->1,a0-
>15,a->21};
aXYZ={aX,aY,aZ}; u21={uX2-uX1,uY2-uY1,uZ2-uZ1};
axyz=aXYZ+u21; {ax,ay,az}=axyz;
Print["aXYZ=",aXYZ/.rep," u21=",u21/.rep," axyz=",axyz/.rep,
" a0=",Sqrt[aX^2+aY^2+aZ^2]/.rep," a=",Sqrt[ax^2+ay^2+az^2]/.rep];
u={uX1,uY1,uZ1,uX2,uY2,uZ2};
For [if=1,if<=6,if++, fa={a,a^2,1/a,1/a^2,Sqrt[a],Log[a]}[[if]];
{S,Q}=Simplify[LengthFunctionSecondDerivative[{fa,a},aXYZ,u21,{u,u}]];
Print["fa= ",fa,", S=",S//MatrixForm];
Print["fa= ",fa,", S=",N[Simplify[S/.rep]]//MatrixForm];
Print["eigs of S=", Chop[Eigenvalues[N[S/.rep]]]];
SFD=LengthFunctionSecondDerivativeByFD[{fa,a},aXYZ/.rep,u21/.rep,δ ];
Print["S by FD: ",Chop[N[SFD],.000001]//MatrixForm];
Print["eigs of SFD=", Chop[Eigenvalues[N[SFD]],.00001]];
];

Figure 14.5. Mathematica implementation of second-derivative computations in (14.11) for F(a)

and check by central finite differences.

0.000671548 − 0.00311033 − 0.00226206 − 0.000671548 0.00311033 0.00226206

− 0.00311033 0.00305731 − 0.00155516 0.00311033 − 0.00305731 0.00155516
− 0.00226206 − 0.00155516 0.00406464 0.00226206 0.00155516 − 0.00406464
fa= a, S=
− 0.000671548 0.00311033 0.00226206 0.000671548 − 0.00311033 − 0.00226206
0.00311033 − 0.00305731 0.00155516 − 0.00311033 0.00305731 − 0.00155516
0.00226206 0.00155516 − 0.00406464 − 0.00226206 − 0.00155516 0.00406464
eigs of S= 0.0103913, 0.0103913, − 0.00519566, 0, 0, 0
0.00067155 − 0.00311033 − 0.00226206 − 0.00067155 0.00311033 0.00226206
− 0.00311033 0.00305731 − 0.00155516 0.00311033 − 0.00305731 0.00155516
− 0.00226206 − 0.00155516 0.00406463 0.00226206 0.00155516 − 0.00406463
S by FD:
− 0.00067155 0.00311033 0.00226206 0.00067155 − 0.00311033 − 0.00226206
0.00311033 − 0.00305731 0.00155516 − 0.00311033 0.00305731 − 0.00155516
0.00226206 0.00155516 − 0.00406463 − 0.00226206 − 0.00155516 0.00406463
eigs of SFD= 0.0103913, 0.0103913, − 0.00519566, 0, 0, 0

√
Figure 14.6. Partial results from script of Figure 14.5 for F(a) = a and numerical values (14.16).
Symbolic results not shown to reduce clutter. Displacement increment δ = 0.01 used for verification
by central finite differences.

14–8
14–9 §14.3 THE CR BAR ELEMENT

Corotated configuration Deformed (current) x−

aligned with CD , with configuration C D
identical element midpoint −
y 2 (x 2 ,y 2 ,z 2 )

− Length L
∼
y z
1 (x1 ,y1 ,z 1 )
∼
z Length L 0
Y, y
E, A constant
10 (X1 ,Y1 ,Z 1 )
20 (X 2 ,Y2 ,Z 2 )
Base configuration C 0
X, x x∼
Z, z Global system

Figure 14.7. The 2-node bar element moving in 3D space.

§14.3. The CR Bar Element

The space bar element shown in Figure 14.7 will be presented here as an application of the foregoing
results on line segment moving through 3D space. The static EICR formulation is illustrated with a
linearly elastic, prismatic, 2-node bar element moving in 3D space, as depicted in that figure. The
element has six degrees of freedom collected in the 6-vector (14.3).
The area and length in the base configuration C0 are A0 and L 0 , respectively. These become A and
L in the deformed configuration C D . The element elongation is called d = L − L 0 . The elastic
modulus is E. In accordance with the usual assumptions of the CR description, bar deformations are
assumed to be small. This allows to carry all integrals over the initial volume A0 L 0 .
Application of the “best corotational fit criterion” shows that, as may be expected, the corotated
configuration C R is aligned with, and lies halfway from, the current end nodes. But this result will
not be required to develop the element equations.

§14.3.1. Internal Energy, Force and Stiffness

The axial strain and stress measures are denoted by e and s, respectively, with s being the energy
conjugate of e. Both are assumed to be constant over the element volume. For the moment the choice
of e and s will be left open. The strain and stress in C0 are 0 and s0 , respectively. In C D they become
e and s = s0 + Ee. The axial forces in C0 and C D are N0 = A0 s0 and N = A0 s = N0 + E A0 e,
respectively.
The energy density in C0 is taken as zero. It becomes U = s0 e + 12 E e2 in C D , which is constant over
the element. The total internal energy in C D is

U= U d V = U V0 = A0 L 0 (s0 e + 12 E e2 ) = L 0 (N0 e + 12 E A0 e2 ). (14.17)
V0

The strain measure is taken to be a function e = e(d, L 0 ) = u(L , L 0 ) to be chosen later. Since
d = L − L 0 and L 0 is fixed, e,d = ∂e/∂d = ∂e/∂ L = e,L and likewise e,dd = ∂ 2 e/(∂d∂d) =

14–9
Chapter 14: THE CR FORMULATION: SPACE BAR 14–10

Table 14.1. Strain Measures for Bar Element and their Derivatives

Strain Measure Symbol Definition for bar ∂e ∂ 2e

∂L ∂ L2
Engineering eE (L − L 0 )/L 0 1/L 0 0

Green-Lagrange eG (L 2 − L 20 )/(2L 20 ) L/L 20 1/L 20

Hencky eH log(L/L 0 ) 1/L −1/L 2

Midpoint eM 2(L − L 0 )/(L + L 0 ) 4L 0 /(L + L 0 )2 −8L 0 /(L + L 0 )3

∂ 2 e/(∂ L∂ L) = e,L L . The derivatives of e with respect to the nodal displacements DOF (14.3) are
obtained by the chain rule: ∂e/∂u = e,L (∂ L/∂u), etc. The necessary partials are already worked
out for the line segment in §14.2, with the substitutions a → L, a → L, â → L̂, etc. Recall that



∂L −L̂ ∂2 L H −H 1
=h= , =G= = (J − h hT ), (14.18)
∂u L̂ ∂u∂u −H H L
where L̂ is the â given in in (14.2), H is defined in (14.8) and J in (14.9), with the replacements
indicated above.
The internal force is the gradient of the internal energy with respect to the node displacements:
∂U ∂U ∂e ∂e ∂e ∂ L ∂e
p= = = L 0 (N0 + E A0 e) = L0 N = L0 N h. (14.19)
∂u ∂e ∂u ∂u ∂ L ∂u ∂L
The tangent stiffness is the Hessian of U :
 
∂ 2U ∂p ∂ N ∂e ∂ L ∂ 2e ∂ L ∂ L T ∂e ∂ 2 L
K= = = L0 + L0 N 2 + L0 N
∂u∂u ∂u ∂u ∂ L ∂u ∂ L ∂u ∂u ∂ L ∂u∂u
 T    
∂e ∂e ∂ L ∂ L ∂ 2e ∂ L ∂ L T
∂e ∂ 2 L 
= L 0 E A0 + L0 N +
∂ L ∂ L ∂u ∂u ∂ L 2 ∂u ∂u ∂ L ∂u∂u
(14.20)
 ∂e 2  ∂ 2e ∂e 
= E A0 L 0 h hT + N L 0 h h T
+ G
∂L ∂ L2 ∂L
 ∂e 2
 2  
1 ∂e ∂ e 1 ∂e
= E A0 L 0 h h + N L0
T
J+ − hh = K M + KG .
T
∂L L ∂L ∂ L2 L ∂L
Here K M and KG denote the material and geometric stiffness, a decomposition already encountered
in the TL description of Chapters 8–9.
§14.3.2. Matrices for Specific Strain Measures
Some specific strain measures and the values of their partials with respect to L are collected in Table
14.1. The appropriate choice should be replaced in (14.19) and (14.20) to get the final form of the
internal force vector and tangent stiffness matrix.

14–10
14–11 §14.3 THE CR BAR ELEMENT

Example 14.1. If the engineering strain e = e E = (L − L 0 )/L 0 is used,

E A0 T N
p = N h, KM = hh , KG = G, K = K M + KG . (14.21)
L0 L

Example 14.2. If the Green-Lagrange strain e = eG = (L 2 − L 20 )/(2L 20 ) is used,

 1 0 0 −1 0 0

 0 1 0 0 −1 0 
L E A0 L 2 T N N  0 0 1 0 0 −1 
,
p=N h, KM = hh , KG = J= K = K M + KG .
L0 L 30 L0 L0 
 −1 0 0 1 0 0

0 −1 0 0 1 0
0 0 −1 0 0 1
(14.22)
For this choice the geometric stiffness KG has the same form as that of the TL bar element derived in Chapter 8.
The material matrix, however, is different unless one sets L = L 0 .

Example 14.3. If the Hencky strain e = e H = log(L/L 0 ) is used,

L0 E A0 L 0 T N L0
p=N h, KM = hh , KG = (J − 2h hT ), K = K M + KG . (14.23)
L L2 L2

CRSpaceBar2InternalForce[ncoor0_,Em_,A0_,ue_,N0_,sm_,numer_]:=
Module[{X1,Y1,Z1,X2,Y2,Z2,LX,LY,LZ,LL0,L0,x1,y1,z1,x2,y2,z2,
Lx,Ly,Lz,LL,L,Lavg,d,EA,e,dedL,d2edL2,ND,h,p},
{{X1,Y1,Z1},{X2,Y2,Z2}}=ncoor0; {{uX1,uY1,uZ1},{uX2,uY2,uZ2}}=ue;
{LX,LY,LZ}={X2-X1,Y2-Y1,Z2-Z1}; LL0=LX^2+LY^2+LZ^2; L0=Sqrt[LL0];
{{x1,y1,z1},{x2,y2,z2}}=ncoor0+ue;
{Lx,Ly,Lz}={x2-x1,y2-y1,z2-z1}; LL=Lx^2+Ly^2+Lz^2; L=Sqrt[LL];
d=L-L0; Lavg=(L+L0)/2; e=Null;
If [sm=="eE", e=d/L0; dedL=1/L0; d2edL2=0];
If [sm=="eG", e=d*Lavg/L0^2; dedL=L/L0^2; d2edL2=1/L0^2];
If [sm=="eH", e=Log[L/L0]; dedL=1/L; d2edL2=-1/L^2];
If [sm=="eM", e=d/Lavg; dedL=L0/Lavg^2; d2edL2=-L0/Lavg^3];
If [e==Null,Print["CRSpaceBar2Force: Illegal sm arg"]; Return[Null]];
ND=N0+Em*A0*e; h={-Lx,-Ly,-Lz,Lx,Ly,Lz}/L;
p=ND*L0*dedL*h; If [numer,p=N[p]];
Return[p]];

ClearAll[ ]; ncoor0=N[{{0,0,0},{11,10,2}}];
Em=5000; A0=3; N0=10; numer=True;
For [iε=1, iε<=4, iε++, ={0,0.0001,0.01,0.1}[[iε]];
ue=N[{{2,3,-4},{-4,-5,8}}*(1+ε)^2];
For [ism=1,ism<=4,ism++, sm={"eE","eG","eH","eM"}[[ism]];
p=CRSpaceBar2InternalForce[ncoor0,Em,A0,ue,N0,sm,numer];
Print["sm: ",sm,", ε=", ε,", p=",p];
]];

Figure 14.8. Mathematica implementation of internal force vector calculation for 2-node space bar.

§14.3.3. Mathematica Implementation

The calculation of the internal force vector for the CR space bar is implemented via module
CRSpaceBar2InternalForce. This is listed in Figure 14.8, along with test statements described
below.

14–11
Chapter 14: THE CR FORMULATION: SPACE BAR 14–12

sm: eE, ε=0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eG, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eH, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eM, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eE, ε=0.0001, p= {− 3.87431, − 1.54886, − 10.8525, 3.87431, 1.54886, 10.8525}
sm: eG, ε= 0.0001, p= {− 3.87476, − 1.54904, − 10.8538, 3.87476, 1.54904, 10.8538}
sm: eH, ε= 0.0001, p= {− 3.87386, − 1.54868, − 10.8513, 3.87386, 1.54868, 10.8513}
sm: eM, ε= 0.0001, p= {− 3.87386, − 1.54868, − 10.8513, 3.87386, 1.54868, 10.8513}
sm: eE, ε= 0.01, p={− 56.577, − 21.3257, − 165.128, 56.577, 21.3257, 165.128}
sm: eG, ε= 0.01, p={− 57.5008, − 21.6739, − 167.824, 57.5008, 21.6739, 167.824}
sm: eH, ε= 0.01, p={− 55.6686, − 20.9833, − 162.477, 55.6686, 20.9833, 162.477}
sm: eM, ε=0.01, p={− 55.6664, − 20.9824, − 162.47, 55.6664, 20.9824, 162.47}
sm: eE, ε=0.1, p={− 430.732, − 36.854, − 1902.59, 430.732, 36.854, 1902.59}
sm: eG, ε= 0.1, p={− 517.786, − 44.3025, − 2287.12, 517.786, 44.3025, 2287.12}
sm: eH, ε= 0.1, p={− 358.761, − 30.6961, − 1584.69, 358.761, 30.6961, 1584.69}
sm: eM, ε=0.1, p={− 356.998, − 30.5453, − 1576.9, 356.998, 30.5453, 1576.9}

Figure 14.9. Results obtained by running the script of Figure 14.8.

The module is invoked as

p = CRSpaceBar2InternalForce[ncoor0, Em, A0, ue, N0, sm, numer ] (14.24)

The arguments are

ncoor0 Node coordinates of element base configuration, arranged as
{ { X1,Y1,Z1 },{ X2,Y2,Z2 } }.
Em Elastic modulus.
A0 Cross section area in base configuration.
ue Node displacements from base to deformed configuration, arranged as
{ { uX1,uY1,uZ1 },{ uX2,uY2,uZ2 } }.
N0 Axial force in base configuration.
sm A two-character string that specifies which strain measure to use for the bar constitutive
equations: "eE", "eG", "eH" and "eM" for the engineering, Green-Lagrange, Hencky
and midpoint measures, respectively. If sm is not one of these, an error message is
printed and the module returns Null.
numer A logical logical flag with the value True or False. If True the computations are
carried out in floating-point arithmetic. If False symbolic processing is assumed.
The module returns as function value
p The internal force vector arranged as a one-dimensional list: { pX1,pY1,pZ1,pX2,pY2,pZ2 }.
The module is numerically exercised by the test statements shown at the bottom of Figure 14.8.
The test case corresponds to the following data: E = 5000, A0 = 3, N0 = 10, base configuration
node coordinates (X 1 , Y1 , Z 1 ) = (0, 0, 0) and (X 2 , Y2 , Z 2 ) = (11, 12, 2) and node displacements
(u X 1 , u Y 1 , u Z 1 = (2, 3, −4) ∗ (1 + )2 and (u X 2 , u Y 2 , u Z 2 ) = (−4, −5, 8) ∗ (1 + epsilon)2 , in

14–12
14–13 §14.3 THE CR BAR ELEMENT

which  is an adjustable parameter. If  = 0 the bar moves in space but does not change length:
L 0 = L = 15. If  > 0 the bar stretches by approximately  L 0 .
The four strain measures e E , eG , e H and e M are exercised by cycling over sm = "eE", "eG", "eH",
and "eM", respectively, in the inner test loop. For values of : 0, 0.0001, 0.01 and 0.1 are tested in
the outer test loop.
As can be observed the internal force vectors are identical if  = 0, since for zero strain it does not
matter which measure is used. The forces are very close for the different measures if  = 0.0001,
which is very small strain (roughly 100 micros) but they differ substantially as  gets large.
The implementation and verification of the tangent stiffness matrix is deferred to Exercises 14.3 and
14.4.

The formulation of the mass and damping matrices for dynamic analysis is not provided in this
Chapter.

14–13
Chapter 14: THE CR FORMULATION: SPACE BAR 14–14

Homework Exercises for Chapter 14

The CR Formulation: Space Bar

EXERCISE 14.1 [A:15] Verify by hand the formulas given in §14..3.2.

EXERCISE 14.2 [A:15] Verify the expressions given in Examples 14.1–3, and append to these the p, K M and
KG matrices for the choice e = e M .

EXERCISE 14.3 [C:20] Implement the calculation of the tangent stiffness matrix for the space bar element
(any language is OK) to return K M and KG . If done in Mathematica, the code of Figure 14.8, which is posted
on the web site linked to the Chapter 14 Index, may be used as template. It is convenient to compute and return
K M and KG as two separate matrices as function value in Mathematica: Return[{ KM,KG }].
As numerical test, run the bar used in the test statements of that figure using  = 0 and any strain measure. Print
out K M , KG , K = K M + KG and the eigenvalues of the three matrices. Validation check: the rank of K M , KG
and K should be 1, 3 and 3, respectively (except in the case of engineering strain, in which case the rank should
be 1, 2 and 3, respectively.)

EXERCISE 14.4 [C:20] Using the module developed in the previous Exercise, compute and show K M , KG and
K for the bar data used in the example of Figure 14.8. For the inner loop cycle over the 4 strain measure choices
used in the internal force test. For the outer test loop, cycle over  equal to 0, 0.01 and −0.01. Compute and
show the eigenvalues of the three matrices. Validation check: the rank of K M , KG and K should be 1, 3 and 3,
respectively. For  = −0.01 you may see some negative eigenvalues in KG and K; do not be alarmed.

14–14
 13
.

CR Formulation
Overview II

13–1
Chapter 13: CR FORMULATION OVERVIEW II 13–2

TABLE OF CONTENTS

Page
§13.1. Internal Forces 13–3
§13.1.1. Force Transformations . . . . . . . . . . . . . . . 13–3
§13.1.2. Projector Properties . . . . . . . . . . . . . . . . 13–4
§13.2. Tangent Stiffness 13–5
§13.2.1. Definition . . . . . . . . . . . . . . . . . . . 13–5
§13.2.2. Material Stiffness . . . . . . . . . . . . . . . . 13–5
§13.2.3. Geometric Stiffness . . . . . . . . . . . . . . . . 13–6
§13.2.4. Consistency Verification . . . . . . . . . . . . . . 13–7
§13.3. Three Consistent CR Formulations 13–7
§13.3.1. Consistent CR formulation (C) . . . . . . . . . . . . 13–7
§13.3.2. Consistent Equilibrated CR Formulation (CE) . . . . . . . 13–8
§13.3.3. Consistent Symmetrizable Equilibrated CR Formulation (CSE) . . 13–8
§13.3.4. Formulation Requirements . . . . . . . . . . . . . 13–8
§13.3.5. Limitations of the EICR Formulation . . . . . . . . . . 13–9
§13.4. Conclusions 13–10

13–2
13–3 §13.1 INTERNAL FORCES

§13.1. Internal Forces

e
The element internal force vector p̄e and tangent stiffness matrix K̄ are computed in the CR configuration
based on small deformational displacements and rotations. Variations of the element DOF, collected
in ved as indicated in Table 13.2, must be linked to variations in the global frame to flesh out the EICR
interface of Figure 12.4. This section develops the necessary relations.

§13.1.1. Force Transformations

Consider an individual element e with N e nodes with six DOF (three translations and three rotations) at
each. Assume the element to be linearly elastic, undergoing only small deformations. Its internal energy
is assumed to be a function of the deformational displacements: U e = U e (v̄ed ), with array v̄ed organized
as shown in Table 13.2. U e is a frame independent scalar. The element internal force vector p̄e in the
CR frame is given by p̄e = ∂U e /∂ v̄ed . For each node a = 1, . . . N e :
 ∂U e 
 
∂U e p̄eua ∂ ūda 
p̄ae = e , or = (13.1)
∂ v̄da p̄eθa ∂U e
∂ θ̄da
where the second form separates the translational and rotational (moment) forces. To refer these to the
global frame we need to relate local-to-global kinematic variations:
 ∂ ūe ∂ ūedb 
  
Ne   db
δ ūeda δuae  ∂uae ∂ωae 
e = Jab , Jab =  e e . (13.2)
δ θ̄da b=1
δωae ∂ θ̄db ∂ θ̄db
∂uae ∂ωae
e
From virtual work invariance, (p̄eu )T δ ūed + (p̄eθ )T δ θ̄d = (peu )T δue + (peθ )T δωe , whence
  
Ne  
peua p̄eua
= T
Jab , a = 1, . . . N e . (13.3)
peθa b=1
p̄eθa

T T
It is convenient to split the Jacobian in (13.2) as Jab = H̄b P̄ab Ta and Jab
T
= TaT P̄ab H̄b . These matrices
are provided from three transformation stages, flowcharted in Figure 13.1:
      
δ ūedb I 0 δ ūeb
e
∂ θ̄db
e = , with H̄db = ,
δ θ̄db 0 H̄db δ ω̄eb ∂ ω̄edb
 ∂ ūe ∂ ūedb 
    db
 ∂ ūa ∂ ω̄ae
e
δ ūeb δ ūae 
= P̄ab with P̄ab =  , , (13.4)
δ ω̄eb δ ω̄ae
∂ ω̄edb ∂ ω̄edb
∂ ūe ∂ ω̄ae
 e  e    e a
δ ūa δua TR 0 δua
= Ta = ,
δ ω̄ae δωae 0 TR δωae

The 3 × 3 matrix L is the Jacobian derivative already encountered in Chapter 12. An explicit expression
in terms of θ is given in Appendix R. To express compactly the transformations for the entire element it
is convenient to assemble the 6N e × 6N e matrices

13–3
Chapter 13: CR FORMULATION OVERVIEW II 13–4

 
P̄11 P̄12 . . . P̄1N e _ _
CR deformational HPT global total
 P̄21 P̄22 . . . P̄2N e  _
displac_& rotations
P̄ =  , δud , δθd
displac & spins
δu, δω
... ... ... ... (13.5)
P̄ N e 1 P̄ N e 2 . . . P̄ N e N e _ rotation-to- global-to-CR T
T = diag [ T R T R . . . T R ] . H spin Jacobian frame rotator

and H̄ is defined in Appendix R. Then the element CR deformational Projector CR total

transformations can be written displac
_ & _spins _ _ _ displac
_ &_spins
δud , δωd P = Pu + Pω δu, δω
T T
δved = H̄ P̄ T δve , pe = TT P̄ H̄ p̄e . (13.6) Figure 13.1. Staged transformation sequence
from deformed to global DOFs.

The 6 × 6 matrix P̄ab in (13.4) extracts the deformational part of the displacement at node b in terms
of the total displacement at node a, both referred to the CR frame. At the element level, δ v̄ed = P̄ δ v̄e
extracts the deformational part by “projecting out” the rigid body modes. For this reason P̄ is called a
projector. As noted in Section 3.5, P̄ may be decomposed into a translational projector or T-projector P̄u
and a rotational projector or R-projector P̄ω , so that P̄ = P̄u + P̄ω . Each has a rank of 3. The T-projector
is a purely numeric matrix exemplified in Chapter 12. The R-projector can be expressed as P̄ω = S̄Ḡ,
where S̄ and Ḡ are defined in Chapter 12. Additional properties are studied below.
Remark 3. Rankin and coworkers [R.54,R.63–R.67] use an internal force transformation in which the incremental
−1
nodal rotations are used instead of the spins. This results in an extra matrix, H̄ appearing in the sequence (13.6).
The projector derived in those papers differs from the one constructed here in two ways: (1) only the R-projector is
considered, and (2) the origin of the CR frame is not placed at the element centroid but at an element node defined
by local node numbering. Omitting the T-projection is inconsequential if the element is “clean” with respect to
translational rigid body motions [R.30, Sec. 5].

§13.1.2. Projector Properties

In this section the bar over P, etc is omitted for brevity, since the properties described below are frame
independent. In Section 12.5 it was stated without proof that the orthogonal projector property P2 = P
was verified. Since P2 = (Pu − Pω )2 = P2u − 2Pu Pω + P2ω , satisfaction requires P2u = Pu , P2ω = Pω , and
Pu Pω = 0. Verification of P2u = Pu is trivial. Recalling that Pω = SG we get

P2ω = S (G S) G = S I G = S G = Pω . (13.7)

This assumes D = I; verification for non-identity D is immediate upon removal of zero rows and columns.
The orthogonality property Pu Pω = Pu S G = 0 follows by observing that Spin(xC ) = 0, where xC are
the coordinates of the element centroid in any frame with origin at C.
In the derivation of the consistent tangent stiffness, the variation of PT contracted with a force vector f,
where f is not varied, is required. The variation of the projector can be expressed as

δP = δPu − δPω = −δPω = −δS G − S δG. (13.8)

For the tangent stiffness one needs δPT f. This vector can be decomposed into a balanced (self-
equilibrated) force fb = Pf and an unbalanced (out of equilibrium) force fu = (I − P)f. Then

δPT f = −(GT δST + δGT ST ) (fb + fu ) = −GT δST PT f − (GT δST + δGT ST ) fu
(13.9)
= −GT δST PT f + δPT fu .

13–4
13–5 §13.2 TANGENT STIFFNESS

where ST fb = 0 was used. This comes from the fact that the columns of S are the three rotational rigid
body motions, which do not produce work on an self-equilibrated force vector.
The term δPT fu will be small if element configurations C R and C D are close because in this case fu will
approach zero. If G has the factorizable form shown below, however, we can show that δPT fu = 0
identically, regardless of how close C R and C D are, as long as f is in translational equilibrium. Assume
that G can be factored as
G = Ξ Γ, with δG = δΞ Γ. (13.10)
where Ξ is a coordinate dependent invertible 3 × 3 matrix, and Γ is a constant 3 × 6N e matrix. Since
GS = I, Ξ−1 = Γ S, and δG S + G δS = δΞ Γ S + G δS = 0, whence δΞ = −G δS Ξ. Then
δPT fu = −(GT δST + ΓT δΞT ST ) fu = −(GT δST − ΓT ΞT δST , GT ST ) fu
= −(GT δST − GT δST PωT ) fu = −(GT δST (I − PωT ) fu = −GT δST (I − PωT ) fu (13.11)
= −GT δST (I − PωT )(I − PT ) f = 0,

if f is in translational equilibrium: f = PωT f. This is always satisfied for any element that represents rigid
body translations correctly [R.30, Sec. 5].

§13.2. Tangent Stiffness

We consider here only the stiffness derived from the internal energy. The load stiffness due to noncon-
servative forces, such as aerodynamic pressures, has to be treated separately.

§13.2.1. Definition
The consistent tangent stiffness matrix Ke of element e is defined as the variation of the internal forces
with respect to element global freedoms:
def ∂pe
δpe = Ke δve , whence Ke = . (13.12)
∂ve
Taking the variation of pe in (13.6) gives rise to four terms:
T T T T T T T
δpe = δTT P̄ H̄ p̄e + TT δ P̄ H̄ p̄e + TT P̄ δHT p̄e + TT P̄ H̄ δ p̄e
(13.13)
= (KeG R + KeG P + KeG M + KeM ) δve .
The four terms identified in (13.13) receive the following names. K M is the material stiffness, KG M the
moment-correction geometric stiffness, KG P the equilibrium projection geometric stiffness, and KG R the
rotational geometric stiffness. If nodal eccentricities treated by rigid links are considered, one more term
appears, called the eccentricity geometric stiffness. This term is studied in great detail in [R.37].

§13.2.2. Material Stiffness

The material stiffness is generated by the variation of the element internal forces pe :
T T
KeM δve = TTR P̄ H̄ δ p̄e . (13.14)

The linear stiffness matrix in terms of the deformational freedoms in v̄ed is defined as the Hessian of the
internal energy:
e ∂ 2 Ū e ∂ p̄e
K̄ = e e = e (13.15)
∂ v̄d ∂ v̄d ∂ v̄d

13–5
Chapter 13: CR FORMULATION OVERVIEW II 13–6

Using the transformation of δ v̄e in (13.6) gives

T T e
KeM = TT P̄ H̄ K̄ H̄ P̄ T. (13.16)
e
Thus the material stiffness is given by a congruential transformation of the local stiffness K̄ to the global
frame. This is formally the same as in linear analysis but here the transformation terms depend on the
state. The expression (13.16) is valid only if Ke is independent of the deformational v̄ed freedoms.

§13.2.3. Geometric Stiffness

To express compactly the geometric stiffness components it is convenient to introduce the arrays
 n̄e     Spin(n̄e ) 
1 Spin(n̄e1 ) 1
 m̄1 e
 0   Spin(m̄e1 ) 
 .   ..   .. 
p̄eP = P̄ H̄ p̄e =   , F̄n =   , F̄mn =  .
T T
 . .  .   .  (13.17)
 e   Spin(n̄e e )   
n̄ N e N Spin(n̄ N e )
e

m̄eN e 0 Spin(m̄eN e )
These are filled with the projection node forces p̄eP . Only the final form of the geometric stiffness
components is given below, omitting the detail derivations of [R.37]. The rotational geometric stiffness
T T
is generated by the variation of T: KeRG δve = δTT P̄ H̄ p̄e and can be expressed as

KG R = −TT F̄nm Ḡ T, (13.18)

KeRG is the gradient of the internal force vector with respect to the rigid rotation of the element. This
interpretation is physically intuitive because a rigid rotation of a stressed element necessarily reorients
the stress vectors by that amount. Consequently the internal element forces must rigidly rotate to preserve
equilibrium.
The moment-correction geometric stiffness is generated by the variation of the jacobian H: KeRG δve =
T T
TTR P̄ δ H̄ p̄e . It is given by
T
KeG M = TT P̄ L̄ P̄T. (13.19)
where L is defined in Section 3.5.
The equilibrium projection geometric stiffness arises from the variation of the projector P̄ with respect
to the deformed element geometry: KeG P δve = TTR δPT HT p̄e . As in Section 4.2, decompose p̄e into a
T T
balanced (self-equilibrated) force p̄eb = P̄ p̄e and an unbalanced force p̄eu = p̄e − p̄eb . If δ P̄ p̄eu is either
identically zero or may be neglected as discussed in Section 4.2, KeG P is given by
T
KG P = −TT Ḡ F̄n P̄ T, (13.20)

in which the balanced force p̄eb is used in (13.17) to get F̄n .

T
If TT δ P̄ p̄eu cannot be neglected, as may happen in highly warped shell elements in a coarse mesh, the
following correction term may be added to KeRG :

T ∂G
KG P = −T T
Ḡ F̄nu P̄ + e
S̄ p̄u T, (13.21)
∂v
where Fnu is Fn of (13.17) when p̄eu is inserted instead of p̄e . In the computations reported in Part II
[R.38] this term was not included.

13–6
13–7 §13.3 THREE CONSISTENT CR FORMULATIONS

KeRG expresses the variation of the projection of the internal force vector p̄e as the element geometry
changes. This can be interpreted mathematically as follows: In the vector space of element force vectors
the subspace of self-equilibrium force vectors changes as the element geometry changes. The projected
force vector thus has a gradient with respect to the changing self-equilibrium subspace, even though the
element force f̄e does not change.
The complete form of the element tangent stiffness, excluding correction terms (13.21) for highly warped
elements, is
 T T e T T T  e
Ke = TT P̄ H̄ K̄ H̄ P̄ + P̄ L̄ P̄ − F̄nm Ḡ − Ḡ F̄n P̄ T = TT K̄ P T, (13.22)
e
in which K̄ R , which is the local tangent stiffness matrix (the tangent stiffness matrix in the local CR
frame of the element) is given by the parenthesized expression.

§13.2.4. Consistency Verification

e
The local tangent stiffness matrix K̄ R given in (13.22) has some properties that may be exploited to verify
the computer implementation [R.54,R.37]:
e T e
K̄ R S̄ = −F̄nm , S̄ K̄ R = −FnT , K̄ M S̄ = 0, K̄G R S̄ = −F̄nm ,
T T T
(13.23)
K̄G P S̄ = 0, S̄ K̄ M = 0, S̄ (K̄G R + K̄G P ) = −F̄n .
e
In addition, rigid-body-mode tests on the linear stiffness matrix K̄ using linearized projectors are
discussed in [R.30]. The set (13.23) tests the programming of the nonlinear projector P̄ since it checks
the null space of P. It also indicates whether the projector matrix is used correctly in the stiffness
formulation. However, satisfaction does not fully guarantee consistency between the internal force
and the tangent stiffness because H̄ and L̄ are left unchecked. Full verification of consistency can be
numerically done through finite difference techniques.

§13.3. Three Consistent CR Formulations

From the foregoing unified forms of the internal force and tangent stiffness, three CR consistent for-
mulations can be obtained by making simplifying assumptions at the internal force level. These satisfy
self-equilibrium and symmetry to varying degree. The following subsections describe the three versions
in order of increasing complexity. For all formulations one can take in account DOFs at eccentric nodes
as described in [R.37].

§13.3.1. Consistent CR formulation (C)

This variant is that developed by Bergan and coworkers in the 1980s at Trondheim and summarized in
the review article [R.56]. The internal force (13.6) is simplified by taking H̄ = I and P̄ = I, while
retaining δ v̄ed = H̄ P̄ δ v̄e for recovery of deformational DOFs. Since δ P̄ = δ H̄ = 0, the expression for
the tangent stiffness of (13.13) simplifies to the material and rotational geometric stiffness terms:
e e
pe = TT K̄ v̄ed , Ke = TT (K̄ H̄ P̄ − F̄nm Ḡ)T. (13.24)
e
Here F̄nm is computed according to (13.17) with p̄e = K̄ v̄d . The internal force is in equilibrium with
respect to the CR configuration C R . For a shell structure, the material stiffness approaches symmetry as
the element mesh is refined if the membrane strains are small. As the mesh is refined, the deformational
rotation axial vectors θ̄da become smaller and approach vector properties that in turn make H(θda

13–7
Chapter 13: CR FORMULATION OVERVIEW II 13–8
e
approach the identity matrix. With small membrane strains K̄ is indifferent with respect to post-
e e
multiplication with P̄ because the C R and C configurations will be close and K̄ P → K̄ . The consistent
geometric stiffness is always unsymmetric, even at equilibrium. Because of this fact one cannot expect
quadratic convergence for this formulation if a symmetric solver is used.
This formulation may be unsatisfactory for warped quadrilateral shell elements since the C R and C D
reference configurations may be far apart. Only in the limit of a highly refined element mesh will the
C R and C D references in general be close, and satisfactory equilibrium ensured.

§13.3.2. Consistent Equilibrated CR Formulation (CE)

The internal force (13.6) is simplified by taking H̄ = I so δH = 0, but the projector P̄ is retained. This
gives
T e T e T T
pe = TT P̄ K̄ v̄ed , Ke = TT (P̄ K̄ H̄ P̄ − F̄mn Ḡ − Ḡ F̄n P̄)T. (13.25)
T e
where F̄nm and F̄n are computed according to (13.17) with p̄e = P̄ K̄ v̄ed .
Due to the presence of P̄ on both sides, the material stiffness of the CE formulation approaches symmetry
as the mesh is refined regardless of strain magnitude. The geometric stiffness at the element level is non-
symmetric, but the assembled global geometric stiffness will become symmetric as global equilibrium is
approached, provided that there are no applied nodal moments and that displacement boundary conditions
are conserving. A symmetrized global tangent stiffness maintains quadratic convergence for refined
element meshes with this formulation.

§13.3.3. Consistent Symmetrizable Equilibrated CR Formulation (CSE)

All terms in (13.6) are retained, giving
T T e T T e T T T
pe = TT P̄ H̄ K̄ v̄ed , Ke = TT (P̄ H̄ K̄ H̄ P̄ + P̄ L̄ P̄ − F̄nm Ḡ − Ḡ F̄n P̄)T. (13.26)
T e
where F̄nm and F̄n are computed according to (13.17) with f̄ = P̄ H̄ K̄ v̄ed . The assembled global
geometric stiffness for this formulation becomes symmetric as global equilibrium is approached, as
in the CE case, as long as there are no applied nodal moments and the loads as well as boundary
conditions are conserving. Since the material stiffness is always symmetric, quadratic convergence with
a symmetrized tangent stiffness can be exppected without the refined-mesh-limit assumption of the CE
formulation.
Remark 4. The relative importance of including the H matrix, which is neglected by most authors, and the physical
significance of this Jacobian term are discussed in Part II [R.38].

§13.3.4. Formulation Requirements

It is convenient to set forward a set of requirements for geometrically nonlinear analysis with respect
to which different CR formulations can be evaluated. They are listed below in order of decreasing
importance.
Equilibrium. By this requirement is meant: to what extent the finite element internal force vector p is in
self-equilibrium with respect to the deformed configuration C D ? This is a fundamental requirement for
tracing the correct equilibrium path in an incremental-iterative solution procedure.
Consistency. A formulation is called consistent if the tangent stiffness is the gradient of the internal forces
with respect to the global DOF. This requirement determines the convergence rate of an incremental-
iterative solution algorithm. An inconsistent tangent stiffness may give poor convergence, but does not

13–8
13–9 §13.3 THREE CONSISTENT CR FORMULATIONS

Table 13.4. Attributes of Corotated Formulations C, CE and CSE.

Formulation Self-equil.(1) Consistent(2) Invariant(3) Symmetriz.(4) Elem.Indep.(5)

√ √ √
C √ √ √ √
CE √ √ √ √ √
CSE

checked if element is in self-equilibrium in deformed configuration C .

D
(1)
(2) checked if tangent stiffness is the v gradient of the element internal force.
(3) checked if the formulation is insensitive to choice of node numbering.
(4) checked if formulation maintains quadratic convergence of a true Newton solution
algorithm with a symmetrized tangent stiffness matrix.
(5) checked if the matrix and vector operations that account for geometrically nonlinear
effects are the same for all elements with the same node and DOF configuration.

alter the equilibrium path since this is entirely prescribed by the foregoing equilibrium requirement.
However, lack of consistency may affect the location of bifurcation (buckling) points and the branch
switching mechanism for post-buckling analysis. In other words, an inconsistent tangent stiffness matrix
may detect (“see”) a bifurcation where equilibrium is not satisfied from the residual equation. Subsequent
traversal by branch-switching will then be difficult because the corrector iterations need to jump to the
secondary path as seen by the residual equation.
Invariance. This requirement refers to whether the solution is insensitive to internal choices that may
depend on node numbering. For example, does a local element-node reordering give an altered equi-
librium path or change the convergence characteristics for the analysis for an otherwise identical mesh?
The main contributor to lack of invariance is the way the deformational displacement vector is extracted
from the total displacements, if the extraction is affected by the choice of the local CR frame. If lack of
invariance is observed, it may be usually traced to the matrix G, which links the variation of the rigid
body rotation to that of the nodal DOF degrees of freedom.
Symmetrizability. This means that a symmetrized K can be used without loss of quadratic convergence
rate in a true Newton solver even when the consistent tangent stiffness away from equilibrium is not
symmetric. In the examples studied in Part II [R.38] this requirement was met when the material stiffness
of the formulation was rendered symmetric.
Element Independence. This is used in the sense of the EICR discussed in Section 2.5. It means that
the matrix and vector operations that account for geometrically nonlinear effects are the same for all
elements that possess the same node and DOF configuration.
Attributes of the C, CE and CSE formulations in light of the foregoing requirements are summarized in
Table 13.4.

§13.3.5. Limitations of the EICR Formulation

The present CR framework, whether used in the C, CE or CSE formulation variants, is element inde-
pendent in the EICR sense discussed in Section 2.5 since it does not contain gradients of intrinsically
element dependent quantities such as the strain-displacement relationship. This treatment is appropriate
for elements where the restriction to small strains automatically implies that the CR and deformed ele-
ment configurations are close. This holds automatically for low order models such as two-node straight
bars and beams, and three-node facet shell elements.
The main reason for limiting element independence to low-order elements is the softening effect of the
nonlinear projector P. The use of P to restore the correct rigid body motions, and hence equilibrium with

13–9
Chapter 13: CR FORMULATION OVERVIEW II 13–10

respect to the deformed element geometry, effectively reduces the eigenvalues of the material stiffness
relative to the CR material stiffness K̄e before projection. This softening effect becomes significant if
the C R and C D geometries are far apart.
Such softening effects are noticeable in four-node initially-warped shell elements. Assume that the
element is initially warped with “positive” warping, and consider only the effect of P. The element
material stiffness of this initial positive warping is then K+ = P+ T
Ke P+ = Ke . Apply displacements
that switch this warping to the opposite of the initial one; that is, a “negative” warping. The new
element material stiffness then becomes K− = P− T
Ke P− = Ke One will intuitively want the two element
configurations to have the same rigidity in the sense of the dominant nonzero eigenvalues of the tangent
stiffness matrix. But it can be shown that the eigenvalues of the projected material stiffness matrix K−
can be significantly lower than those of the initial stiffness matrix K+ . If the element stiffness Ke is
referred to the flat element projection, one can restore symmetry of K+ and K− with respect to dominant
nonzero eigenvalues, but it is not possible to remove the softening effect.
This argument also carries over to higher order bar, arch and shell elements that are curved in the initial
reference configuration. It follows that the EICR is primarily useful for low-order elements of simple
geometry.

§13.4. Conclusions

This article presents a unified formulation for geometrically nonlinear analysis using the CR kinematic
description, assuming small deformations. Although linear elastic material behavior has been assumed
for brevity, extension to materially nonlinear behavior such as elastoplasticity and fracture within the
confines of small deformations, is feasible as further discussed below. All terms in the internal force and
tangent stiffness expressions are accounted for. It is shown how dropping selected terms in the former
produces simpler CR versions used by previous investigators.
These versions have been tested on thin shell and flexible-mechanism structures, as reported in Part II
[R.38]. Shells are modeled by triangle and quadrilateral elements. The linear stiffness of these elements
is obtained with the ANDES (Assumed Natural DEviatoric Strains) formulation of high performance
elements [R.27–R.32,R.53]. Test problems include benchmarks in buckling, nonlinear bifurcation and
collapse.
Does the unified formulation close the book on CR? Hardly. Several topics either deserve further
development or have been barely addressed:
(A) Relaxing the small-strain assumption to allow moderate deformations.
(B) Robust handling of extremely large rotations involving multiple revolutions.
(C) Integrating CR elements with rigid links and joint elements for flexible multibody dynamics.
(D) Using substructuring concepts for CR modeling of structural members with continuum elements.
(E) Achieving a unified form for CR dynamics, including nonconservative effects and multiphysics.
Topic (A) means the use of CR for problems where strains may locally reach moderate levels, say 1–10%,
as in elastoplasticity and fracture, using appropriate strain and stress measures in the local frame. The
challenge is that change of metric of the CR configuration should be accounted for, even if it means
dropping the EICR property. Can CR compete against the more established TL and UL descriptions?
It seems unreasonable to expect that CR can be of use in overall large strain problems such as metal
forming, in which UL reigns supreme. But it may be competitive in localized failure problems, where
most of the structure remain elastic although undergoing finite rotations.

13–10
13–11 §13.4 CONCLUSIONS

Some data points are available: previous large-deformation work presented in [R.46,R.47,R.50,R.51,R.57].
More recently Skallerud et al. reported [R.75] that a submerged-pipeline failure shell code using the AN-
DES CR quadrilateral of [R.37] plus elastoplasticity [R.74] and fracture mechanics [R.21] was able to
beat a well known commercial TL-based code by a factor of 600 in CPU time. This speedup is of obvious
interest in influencing both design cycle and deployment planning.
Topic (B) is important in applications where a floating (free-free) structure undergoes several revolutions,
as in combat airplane maneuvers, payload separation or orbital structure deployment. The technical
difficulty is that expressions presented in the Appendix cannot handle finite rotations beyond ±2π , and
thus require occasional resetting of the base configuration. While this can be handled via restarts for
structures such as full airplanes, it can be more difficult when the relative rotation between components
exceeds ±2π , as in separation, fragmentation or deployment problems.
Topics (C) and (D) have been addressed in the FEDEM program developed by SINTEF at Trondheim,
Norway. This program combines CR shell and beam elements of [R.37], grouped into substructures,
with kinematic objects typical of rigid-body dynamics: eccentric links and joints. Basic tools used in
FEDEM for combining joint models with flexible continuum elements are covered in a recent book
[R.73].
Finally, topic (E) is fertile ground for research. The handling of model components such as mass,
damping and nonconservative effects in fluid-structure interaction and aeroelasticity is an active ongoing
research topic. For example a recent paper [R.26] describes flight maneuver simulations of a complete
F-16 fighter using CR elements to model the aircraft. As in statics, a key motivation for CR in dynamics
is reuse of linear FEM force-stiffness libraries. Can that reuse extend to mass and damping libraries?
And how do standard time integration methods perform when confronted with unsymmetric matrices?
These topics have barely been addressed.

13–11
 12
.

CR Formulation
Overview I

12–1
Chapter 12: CR FORMULATION OVERVIEW I 12–2

TABLE OF CONTENTS

Page
§12.1. Introduction 12–3
§12.2. The Emergence of CR 12–3
§12.2.1. Continuum Mechanics Sources . . . . . . . . . . . . 12–3
§12.2.2. FEM Sources . . . . . . . . . . . . . . . . . . 12–4
§12.2.3. Shadows of the Past . . . . . . . . . . . . . . . . 12–5
§12.2.4. Linking FEM and CR . . . . . . . . . . . . . . . 12–6
§12.2.5. Element Independent CR . . . . . . . . . . . . . . 12–7
§12.3. Corotational Kinematics 12–9
§12.3.1. Configurations . . . . . . . . . . . . . . . . . 12–9
§12.3.2. Coordinate Systems . . . . . . . . . . . . . . . . 12–9
§12.3.3. Coordinate Transformations . . . . . . . . . . . . . 12–11
§12.3.4. Rigid Displacements . . . . . . . . . . . . . . . . 12–11
§12.3.5. Rotator Formulas . . . . . . . . . . . . . . . . 12–12
§12.3.6. Degrees of Freedom . . . . . . . . . . . . . . . . 12–12
§12.3.7. EICR Matrices . . . . . . . . . . . . . . . . . 12–13
§12.3.8. Deformational Translations . . . . . . . . . . . . . . 12–16
§12.3.9. Deformational Rotations . . . . . . . . . . . . . . 12–17

12–2
12–3 §12.2 THE EMERGENCE OF CR

Note: This and the following Chapters on the CR formulation of geometrically nonlinear FEM are new.
References and Appendices for this material have been placed in Appendix R.

§12.1. Introduction

Three Lagrangian kinematic descriptions are in present use for finite element analysis of geometrically
nonlinear structures: (1) Total Lagrangian (TL), (2) Updated Lagrangian (UL), and (3) Corotational
(CR). The CR description is the most recent of the three and the least developed one. Unlike the others,
its domain of application is limited by a priori kinematic assumptions:

Displacements and rotations may be arbitrarily large, but deformations must be small. (12.1)

Because of this restriction, CR has not penetrated the major general-purpose FEM codes that cater to
nonlinear analysis. A historical sketch of its development is provided in Section 2.
As typical of Lagrangian kinematics, all descriptions: TL, UL and CR, follow the body (or element) as
it moves. The deformed configuration is any one taken during the analysis process and need not be in
equilibrium during a solution process. It is also known as the current, strained or spatial configuration
in the literature, and is denoted here by C D . The new ingredient in the CR description is the “splitting”
or decomposition of the motion tracking into two components, as illustrated in Figure 12.1.
1. The base configuration C0 serves as the origin
of displacements. If this happens to be one
actually taken by the body at the start of the Corotated C R
analysis, it is also called initial or undeformed. Deformed (current,
spatial) C
D
The name material configuration is used pri-
marily in the continuum mechanics literature.
Motion splits into Deformational
2. The corotated configuration C R varies from deformational and rigid motion
element to element (and also from node to
node in some CR variants). For each individ- Base (initial, undeformed,
ual element, its CR configuration is obtained material) configuration C 0
through a rigid body motion of the element
Rigid body motion
base configuration. The associated coordinate
system is Cartesian and follows the element Global
frame
like a “shadow” or “ghost,” prompting names
such as shadow and phantom in the Scandi-
Figure 12.1. The CR kinematic description.
navian literature. Element deformations are
Deformation from corotated to deformed (current)
measured with respect to the corotated con- configuration grossly exaggerated for visibility.
figuration.

In static problems the base configuration usually remains fixed throughout the analysis. In dynamic
analysis the base and corotated configurations are sometimes called the inertial and dynamic reference
configurations, respectively. In this case the base configuration may move at uniform velocity (a Galilean
inertial system) following the mean trajectory of an airplane or satellite.
From a mathematical standpoint the explicit presence of a corotated configuration as intermediary be-
tween base and current is unnecessary. The motion split may be exhibited in principle as a multiplicative
decomposition of the displacement field. The device is nonetheless useful to teach not only the physical
meaning but to visualize the strengths and limitations of the CR description.

12–3
Chapter 12: CR FORMULATION OVERVIEW I 12–4

§12.2. The Emergence of CR

The CR formulation represents a confluence of developments in continuum mechanics, treatment of

finite rotations, nonlinear finite element analysis and body-shadowing methods.

§12.2.1. Continuum Mechanics Sources

In continuum mechanics the term “corotational” (often spelled “co-rotational”) appears to be first men-
tioned in Truesdell and Toupin’s influential exposition of field theories [R.81, Sec. 148]. It is used
there to identify Jaumann’s stress flux rate, introduced in 1903 by Zaremba. By 1955 this rate had been
incorporated in hypoelasticity [R.82] along with other invariant flux measures. Analogous differential
forms have been used to model endochronic plasticity [R.85]. Models labeled “co-rotational” have been
used in rheology of non-Newtonian fluids; cf. [R.17,R.78]. These continuum models place no major
restrictions on strain magnitude. Constraints of that form, however, have been essential to make the idea
practical in nonlinear structural FEA, as discussed below.
The problem of handling three-dimensional finite rotations in continuum mechanics is important in all
Lagrangian kinematic descriptions. The challenge has spawned numerous publications, for example
[R.1,R.4,R.5,R.34,R.42,R.43,R.62,R.71,R.72]. For use of finite rotations in mathematical models, par-
ticularly shells, see [R.60,R.70,R.77]. There has been an Euromech Colloquium devoted entirely to that
topic [R.61].
The term “corotational” in a FEM paper title was apparently first used by Belytschko and Glaum [R.8].
The survey article bt Belystchko [R.9] discusses the concept from the standpoint of continuum mechanics.

§12.2.2. FEM Sources

In the Introduction of a key contribution, Nour-Omid and Rankin [R.54] attribute the original concept
of corotational procedures in FEM to Wempner [R.86] and Belytschko and Hsieh [R.7].
The idea of a CR frame attached to individual elements was introduced by Horrigmoe and Bergan
[R.39,R.40]. This activity continued briskly under Bergan at NTH-Trondheim with contributions by
Kråkeland [R.46], Nygård [R.15,R.56,R.57], Mathisen [R.48,R.49], Levold [R.47] and Bjærum [R.18].
It was summarized in a 1989 review article [R.56]. Throughout this work the CR configuration is labeled
as either “shadow element” or “ghost-reference.” As previously noted the device is not mathematically
necessary but provides a convenient visualization tool to explain CR. The shadow element functions
as intermediary that separates rigid and deformational motions, the latter being used to determine the
element energy and internal force. However the variation of the forces in a rotating frame was not directly
used in the formation of the tangent stiffness, leading to a loss of consistency. Crisfield [R.22–R.24]
developed the concept of “consistent CR formulation” where the stiffness matrix appears as the true
variation of the internal force. An approach blending the TL, UL and CR descriptions was investigated
in the mid-1980s at Chalmers [R.50–R.52].
In 1986 Rankin and Brogan at Lockheed introduced [R.63] the concept of “element independent CR
formulation” or EICR, which is further discussed below. The formulation relies heavily on the use of
projection operators, without any explicit use of “shadow” configurations. It was further refined by
Rankin, Nour-Omid and coworkers [R.54,R.64–R.67], and became essential part of the nonlinear shell
analysis program STAGS [R.68].
The thesis of Haugen on nonlinear thin shell analysis [R.37] resulted in the development of the formulation
discussed in this article. This framework is able to generate a set of hierarchical CR formulations. The
work combines tools from the EICR (projectors and spins) with the shadow element concept and assumed

12–4
12–5 §12.2 THE EMERGENCE OF CR

Corotated (a.k.a.
dynamic reference) frame

PR CR

P C
Corotated (shadow)
configuration C R
Deformed configuration C D
(Shown separate from C R for
visualization convenience)
P0
Base (a.k.a. inertial) frame

C0 Base (initial) configuration C 0

Figure 12.2. The concept of separation of base (a.k.a. inertial) and CR (a.k.a. dynamic)
configurations in aircraft dynamics. Deformed configuration (with deformations grossly exaggerated)
shown separate from CR configuration for visibility. In reality points C and C R coincide.

strain element formulations. Spins (instead of rotations) are used as incremental nodal freedoms. This
simplifies the EICR “front end” and facilitates attaining consistency.
Battini and Pacoste at KTH-Stockholm [R.2,R.3,R.58] have recently used the CR approach, focusing
on stability applications. The work by Teigen [R.79] should be cited for the careful use of offset nodes
linked to element nodes by eccentricity vectors in the CR modeling of prestressed reinforced-concrete
members.

§12.2.3. Shadows of the Past

The CR approach has also roots on an old idea that preceeds FEM by over a century: the separation of rigid
body and purely deformational motions in continuum mechanics. The topic arose in theories of small
strains superposed on large rigid motions. Truesdell [R.80, Sec. 55] traces the subject back to Cauchy in
1827. In the late 1930s Biot advocated the use of incremental deformations on an initially stressed body
by using a truncated polar decomposition. However this work, collected in a 1965 monograph [R.16],
was largely ignored as it was written in an episodic manner, using full notation by then out of fashion.
A rigurous outline of the subject is given in [R.83, Sec. 68] but without application examples.
Technological applications of this idea surged after WWII from a totally different quarter: the aerospace
industry. The rigid-plus-deformational decomposition idea for an entire structure was originally used by
aerospace designers in the 1950s and 1960s in the context of dynamics and control of orbiting spacecraft
as well as aircraft structures. The primary motivation was to trace the mean motion.
The approach was systematized by Fraeijs de Veubeke [R.25], in a paper that essentially closed the subject
as regards handling of a complete structure. The motivation was clearly stated in the Introduction of that
article, which appeared shortly before the author’s untimely death:
“The formulation of the motion of a flexible body as a continuum through inertial space is unsatisfactory from
several viewpoints. One is usually not interested in the details of this motion but in its main characteristics
such as the motion of the center of mass and, under the assumptions that the deformations remain small, the
history of the average orientation of the body. The last information is of course essential to pilots, real and
artificial, in order to implement guidance corrections. We therefore try to define a set of Cartesian mean
axes accompanying the body, or dynamic reference frame, with respect to which the relative displacements,

12–5
Chapter 12: CR FORMULATION OVERVIEW I 12–6

Deformed (membrane) Deformed (shell)

Deformed (bar) Deformed (beam)

;;;
(a)
Base
(b)
;;;; Base

Figure 12.3. Geometric tracking of CR frame: (a) Bar or beam element

in 2D; (b) Membrane or shell element in 3D.

velocities or accelerations of material points due to the deformations are minimum in some global sense.
If the body does not deform, any set of axes fixed into the body is of course a natural dynamic reference
frame.”

Clearly the focus of this article was on a whole structure, as illustrated in Figure 12.2 for an airplane.
This will be called the shadowing problem. A body moves to another position in space: find its mean
rigid body motion and use this information to locate and orient a corotated Cartesian frame.
Posing the shadowing problem in three dimensions requires fairly advanced mathematics. Using two
“best fit” criteria Fraeijs de Veubeke showed that the origin of the dynamic frame must remain at the
center of mass of the displaced structure: C R in Figure 12.2. However, the orientation of this frame
leads to an eigenvalue problem that may exhibit multiple solutions due to symmetries, leading to non
uniqueness. (This is obvious by thinking of the polar and singular-value decompositions, which were not
used in that article.) That this is not a rare occurrence is demonstrated by considering rockets, satellites
or antennas, which often have axisymmetric shape.

Remark 12.1. Only C D (shown in darker shade in Figure 12.2) is an actual configuration taken by the pictured
aircraft structure. Both reference configurations C and C are virtual in the sense that they are not generally
0 R

occupied by the body at any instance. This is in contrast to the FEM version of this idea.

§12.2.4. Linking FEM and CR

The practical extension of Fraeijs de Veubeke’s idea to geometrically nonlinear structural analysis by
FEM relies on two modifications:
1. Multiple Frames. Instead of one CR frame for the whole structure, there is one per element. This
is renamed the CR element frame.
2. Geometric-Based RBM Separation. The rigid body motion is separated directly from the total
element motion using elementary geometric methods. For example in a 2-node bar or beam one
axis is defined by the displaced nodes, while for a 3-node triangle two axes are defined by the plane
passing through the points. See Figure 12.3.
The first modification is essential to success. It helps to fulfill assumption (12.1): the element defor-
mational displacements and rotations remain small with respect to the CR frame. If this assumption
is violated for a coarse discretization, break it into more elements. Small deformations are the key to
element reuse in the EICR discussed below. If intrinsically large strains occur, however, the breakdown
prescription fails. In that case CR offers no advantages over TL or UL.

12–6
12–7 §12.2 THE EMERGENCE OF CR

Deformational
Global element element
equations Incorporate equations Form element
Assembler rigid body mass, stiffness
motions & forces

System equations
of motion

Extract Evaluate
Solver deformational element
Total motions Deformational stresses
displacements displacements
Finite Element
CR "Filters" Library

Figure 12.4. The EICR as a modular interface to a linear FEM library. The
flowchart
is mainly conceptual. For computational efficiency the interface logic
may be embedded with each element through inlining techniques.

The second modification is inessential. Its purpose is to speed up the implementation of geometrically
simple elements. The CR frame determination may be refined later, using more advanced tools such as
polar decomposition and best-fit criteria, if warranted.

Remark 12.2. CR is ocassionally confused with the convected-coordinate description of motion, which is used in
branches of fluid mechanics and rheology. Both may be subsumed within the class of moving coordinate kinematic
descriptions. The CR description, however, maintains orthogonality of the moving frame(s) thus achieving an
exact decomposition of rigid-body and deformational motions. This property enhances computational efficiency as
transformation inverses become transposes. On the other hand, convected coordinates form a curvilinear system
that “fits” the change of metric as the body deforms. The difference tends to disappear as the discretization becomes
progressively finer, but the fact remains that the convected metric must encompass deformations. Such deformations
are more important in solid than in fluid mechanics (because classical fluid models “forget” displacements). The
idea finds more use in UL descriptions, in which the individual element metric is updated as the motion progresses.

§12.2.5. Element Independent CR

As previously noted, one of the sources of the present work is the element-independent corotational
(EICR) description developed by Rankin and coworkers [R.54,R.63–R.67]. Here is a summary descrip-
tion taken from the Introduction to [R.54]:
“In the co-rotation approach, the deformational part of the displacement is extracted by purging the rigid
body components before any element computation is performed. This pre-processing of the displacements
may be performed outside the standard element routines and thus is independent of element type (except
for slight distinctions between beams, triangular and quadrilateral elements).”

Why is the EICR worth study? The question fits in a wider topic: why CR? That is, what can CR do
that TL or UL cannot? The topic is elaborated in the Conclusions section, but we advance a practical
reason: reuse of small–strain elements, including possibly materially nonlinear elements.
The qualifier element independent does not imply that the CR equations are independent of the FEM
discretization. Rather it emphasizes that the key operations of adding and removing rigid body motions
can be visualized as a front end filter that lies between the assembler/solver and the element library, as
sketched in Figure 12.4. The filter is purely geometric. For example, suppose that a program has four
different triangular shell elements with the same node and degree-of-freedom configuration. Then the

12–7
Chapter 12: CR FORMULATION OVERVIEW I 12–8

Table 12.1 Configurations in Nonlinear Static Analysis

by Incremental-Iterative Methods

Name Alias Explanation Equilibrium Identification

Required?
Generic Admissible A kinematically admissible configuration No C
Perturbed Kinematically admissible variation No C + δC
of a generic configuration.

Deformed Current Actual configuration taken during No CD

Spatial the analysis process. Contains
others as special cases.

Base∗ Initial The configuration defined as the Yes C0

Undeformed origin of displacements.
Material
TL: C , UL: C ,
0 n−1
Reference Configuration to which TL,UL: Yes.
CR: C no, C yes CR: C and C
R 0 R 0
computations are referred
Iterated† Configuration taken at the k th No Cnk
iteration of the n th increment step
Target† Equilibrium configuration accepted Yes Cn
at the n th increment step

Corotated‡ Shadow Body or element-attached configuration No CR

obtained from C through a rigid
0
Ghost
body motion (CR description only)

Globally- Connector Corotated configuration forced to align No CG

aligned with the global axes. Used as “connector”
in explaining the CR description.
∗
C0 is often the same as the natural state in which body (or element) is undeformed and stress-free.
† Used only in Part II [R.38] in the description of solution procedures.
‡ In dynamic analysis C and C are called the inertial and dynamic-reference configurations,
0 R

respectively, when they apply to the entire structure.

front end operations are identical for all four. Adding a fifth small-strain element of this type incurs
relatively little extra work to “make it geometrically nonlinear.”
This modular organization is of interest because it implies that the element library of an existing FEM
program being converted to the CR description need not be drastically modified, as long as the analysis
is confined to small deformations. Since that library is typically the most voluminous and expensive part
of a production FEM code, element reuse is a key advantage because it protects a significant investment.
For a large-scale commercial code, the investment may be thousands of man-years.
Of course modularity and computational efficiency can be conflicting attributes. Thus in practice the
front end logic may be embedded with each element through techniques such as code inlining. If so the
flowchart of Figure 12.4 should be interpreted as conceptual.

12–8
12–9 §12.3 COROTATIONAL KINEMATICS

§12.3. Corotational Kinematics

This section outlines CR kinematics of finite elements, collecting the most important relations. Mathe-
matical derivations pertaining to finite rotations are consigned to Appendix A. The presentation assumes
static analysis, with deviations for dynamics briefly noted where appropriate.

§12.3.1. Configurations
To describe Lagrangian kinematics it is convenient to introduce a rich nomenclature for configurations.
For the reader’s convenience those used in geometrically nonlinear static analysis using the TL, UL
or CR descriptions are collected in Table 12.1. Three: base, corotated and deformed, have already
been introduced. Two more: iterated and target, are connected to the incremental-iterative solution
process covered in Part II [R.38]. The generic configuration is used as placeholder for any kinematically
admissible one. The perturbed configuration is used in variational derivations of FEM equations.
Two remain: reference and globally-aligned. The reference configuration is that to which element
computations are referred. This depends on the description chosen. For Total Lagrangian (TL) the
reference is base configuration. For Updated Lagrangian (UL) it is the converged or accepted solution
of the previous increment. For corotational (CR) the reference splits into CR and base configurations.
The globally-aligned configuration is a special corotated configuration: a rigid motion of the base that
makes the body or element align with the global axes introduced below. This is used as a “connector”
device to teach the CR description, and does not imply the body ever occupies that configuration.
The separation of rigid and deformational components of motion is done at the element level. As noted
previously, techniques for doing this have varied according to the taste and background of the investigators
that developed those formulations. The approach covered here uses shadowing and projectors.

§12.3.2. Coordinate Systems

A typical finite element, undergoing 2D motion to help visualization, is shown in Figure 12.5. This
diagram as well as that of Figure 12.6 introduces kinematic quantities. For the most part the notation
follows that used by Haugen [R.37], with subscripting changes.
Configurations taken by the element during the response analysis are linked by a Cartesian global frame,
to which all computations are ultimately referred. There are actually two such frames: the material
global frame with axes {X i } and position vector X, and the spatial global frame with axes {xi } and
position vector x. The material frame tracks the base configuration whereas the spatial frame tracks
the CR and deformed (current) configurations. The distinction agrees with the usual conventions of
dual-tensor continuum mechanics [R.81, Sec. 13]. Here both frames are taken to be identical, since for
small strains nothing is gained by separating them (as is the case, for example, in the TL description).
Thus only one set of global axes, with dual labels, is drawn in Figures 12.5 and 12.6.
Lower case coordinate symbols such as x are used throughout most of the paper. Occasionally it is
convenient for clarity to use upper case coordinates for the base configuration, as in Appendix C.
The global frame is the same for all elements. By contrast, each element e is assigned two local Cartesian
frames, one fixed and one moving:
{x̃i } The element base frame (blue in Figure 12.5). It is oriented by three unit base vectors ii0 , which are
rows of a 3 × 3 orthogonal rotation matrix (rotator) T0 , or equivalently columns of T0T .
{x̄i } The element corotated or CR frame (red in Figure 12.5). It is oriented by three unit base vectors iiR ,
which are rows of a 3 × 3 orthogonal rotation matrix (rotator) T R , or equivalently columns of T0T .

12–9
Chapter 12: CR FORMULATION OVERVIEW I 12–10

Base C is element centroid in statics, Deformed (current)

(initial, but center of mass in dynamics. udP
undeformed) P P R xPCR

uP
CR C
0
xPC P 0 uPR _
xP x2
C0 c
0 X 2,x2
xPR
xP b
a Corotated
x~1 O //x~1 ~
//x 2
X ,x _
element x~2 global frame 1 1 x 1 element
base frame (with material & rigid body
spatial coalesced) rotation CR frame

Figure 12.5. CR element kinematics, focusing on the motion of generic point P.

Two-dimensional kinematics pictured for visualization convenience.
deformational
R rotation Rd
P (a "drilling
(depends on P) rotation" in 2D)
P0 C R _C
c x2
C0
T X2,x2
T0 b
x~ TRT element
1 a
element x~ O global X1,x1 _ CR frame
base frame 2
frame
x1
R0

Figure 12.6. CR element kinematics, focusing on rotational transformation between

frames.

Note that the element index e has been suppressed to reduced clutter. That convention will be followed
throughout unless identification with elements is important. In that case e is placed as supercript.
The base frame {x̃i } is chosen according to usual FEM practices. For example, in a 2-node spatial beam
element, x̃1 is defined by the two end nodes whereas x̃2 and x̃3 lie along principal inertia directions. An
important convention, however, is that the origin is always placed at the element centroid C 0 . For each
deformed (current) element configuration, a fitting of the base element defines its CR configuration, also
known as the element “shadow.” Centroids C R and C ≡ C D coincide. The CR frame {x̄i } originates
at C R . Its orientation results from matching a rigid motion of the base frame, as discussed later. When
the current element configuration reduces to the base at the start of the analysis, the base and CR frames
coalesce: {x̃i ≡ x̄i }. At that moment there are only two different frames: global and local, which agrees
with linear FEM analysis.
Notational conventions: use of G, 0, R and D as superscripts or subscripts indicate pertinence to the
globally-aligned, base, corotated and deformed configurations, respectively. Symbols with a overtilde
or overbar are measured to the base frame {x̃i } or the CR frame {x̄i }, respectively. Vectors without a
superposed symbol are referred to global coordinates {xi ≡ xi }. Examples: x R denote global coordinates
of a point in C R whereas x̃G denote base coordinates of a point in CG . Symbols a, b and c = b − a are
abbreviations for the centroidal translations depicted in Figure 12.5, and more clearly in Figure 12.7(b).
A generic, coordinate-free vector is denoted by a superposed arrow, for example u  , but such entities
rarely appear in this work.
The rotators T0 and T R are the well known local-to-global displacement transformations of FEM analysis.

12–10
12–11 §12.3 COROTATIONAL KINEMATICS

(a) (b) (c)

x~ 1
CR CD
x~ 2 (moving) C0 ϕ0 CR
C0 _ C0
(fixed) x2 c
a ϕR CR
X2, x2 b
_
C G
x1 CG CG
(fixed) X1, x1

Figure 12.7. Further distillation to essentials of Figure 12.5. A bar moving in 2D is shown:
(a) Rigid motion from globally-aligned to base and corotated configurations; (b) key geometric
quantities that define rigid motions in 2D; (c) as in (a) but followed by a stretch from corotated
to deformed. The globally-aligned configuration is fictitious: only a convenient link up device.

Given a global displacement u, ũ = T0 u and ū = T R u.

§12.3.3. Coordinate Transformations

Figures 12.5 and 12.6, although purposedly restricted to 2D, are still too busy. Figure 12.7, which
pictures the 2D motion of a bar in 3 frames, displays essentials better. The (fictitious) globally-aligned
configuration CG is explicitly shown. This helps to follow the ensuing sequence of geometric relations.
Begin with a generic point xG in CG . This point is mapped to global coordinates x0 and x R in the base
and corotated configurations C0 and C R , respectively, through
x0 = T0T xG + a, x R = TTR xG + b, (12.2)
in which rotators T0 and T R were introduced in the previous subsection. To facilitate code checking, for
the 2D motion pictured in Figure 12.7(b) the global rotators are
   
c0 s0 0 cR sR 0
T0 = −s0 c0 0 , T R = −s R c R 0 , c0 = cos ϕ0 , s0 = sin ϕ0 , etc. (12.3)
0 0 1 0 0 1
When (12.2) are transformed to the base and corotated frames, the position vector xG must repeat:
x̃0 = xG and x̄ R = xG , because the motion pictured in Figure 12.7(a) is rigid. This condition requires
x̃0 = T0 (x0 − a), x̄ R = T R (x R − b). (12.4)
These may be checked by inserting x0 and x R from (12.2) and noting that xG repeats.

§12.3.4. Rigid Displacements

The rigid displacement is a vector joining corresponding points in C0 and C R . This may be referred to the
global, base or corotated frames. For convenience call the C0 →C R rotator R0 = TTR T0 . Also introduce
c̃ = T0T c and c̄ = TTR c. Some useful expressions are
ur = x R − x0 = (TTR − T0T )xG + c = (R0 − I)T0T xG + c = (R0 − I)T0T x̃0 + c
= (R0 − I)T0T x̄ R + c = (R0 − I)(x0 − a) + c = (I − R0T )(x R − b) + c,
(12.5)
ũr = T0 ur = T0 (R0 − I)T0T x̃0 + c̃ = (R̃0 − I)x̃0 + c̃,
T
ūr = T R ur = T R (I − R0T )TTR x̄ R + c̄ = (I − R̄0 )x̄ R + c̄.

12–11
Chapter 12: CR FORMULATION OVERVIEW I 12–12

Here I is the 3 × 3 identity matrix, whereas R̃0 = T0 R0 T0T and R̄0 = T R R0 TTR denote the C0 →C R
rotator referred to the base and corotated frames, respectively.

§12.3.5. Rotator Formulas

Traversing the links pictured in Figure 12.8 shows that any rotator
can be expressed in terms of the other two:

T0 = T R R0 , T R = T0 R0T , R0 = TTR T0 , R0T = T0T T R . (12.6) R0T

element element
base frame CR frame
In the CR frame: R̄0 = T R R0 TTR , whence X
~ R0 _
x

T
R̄0 = T0 TTR , R̄0 = T R T0T . (12.7) TT0 TR
T0 global TRT
Notice that T0 is fixed since C and C are fixed throughout the anal-
G 0 frame
ysis, whereas T R and R0 change. Their variations of these rotators X, x
are subjected to the following constraints:
Figure 12.8. Rotator frame links.

δT0 = δT0T = 0, δT R = T0 δR0T , δTTR = δR0 T0T , δR0 = δTTR T0 ,

(12.8)
δR0T = T0T δT R , TTR δT R + δTTR T R = 0, R0T δR0 + δR0T R0 = 0.

The last two come com the orthogonality conditions TTR T R = I and R0T R0 = I, respectively, and provide
δR0 = −R0 δR0T R0 , δR0T = −R0T δR0 R0T , etc.
We denote by ω and ω̄ the axial vectors of R0 and R̄0 , respectively, using the exponential map form of
the rotator described in Section A.10. The variations δω and δ ω̄ are used to form the skew-symmetric
T
spin matrices Spin(δω) = δR0 R0T = −Spin(δω)T and Spin(δ ω̄) = δ R̄0 R̄0 = −Spin(δ ω̄)T . These
matrices are connected by congruential transformations:

Spin(δω) = T0T Spin(δ ω̄) T0 , Spin(δ ω̄) = T0 Spin(δω) T0T . (12.9)

Using these relations the following catalog of rotator variation formulas can be assembled:

δT R = T0 δR0T = −T R δR0 R0T = −T R Spin(δω) = −R0T Spin(δ ω̄) T0 ,

δTTR = δR0 T0T = −R0 δR0T TTR = Spin(δω) TTR = T0T Spin(δ ω̄) R0 .
δR0 = δTTR T0 = −R0 δR0T R0 = Spin(δω) R0 = T0T Spin(δ ω̄) R̄0 T0 ,
T (12.10)
δR0T = T0T δT R = −R0T δR0 R0T = −R0T Spin(δω) = −T0T R̄0 Spin(δ ω̄) T0 ,
δ R̄0 = T0 δR0 T0T = −T0 R0 δR0T TTR = T0 Spin(δω) TTR = Spin(δ ω̄) R̄0 ,
T T
δ R̄0 = T0 δR0T T0T = −T R δR0 R0T T0T = −T R Spin(δω) T0T = −R̄0 Spin(δ ω̄).

12–12
12–13 §12.3 COROTATIONAL KINEMATICS

Table 12.2. Degree of Freedom and Conjugate Force Notation

Notation Frame Level Description

v̂ = [ v̂1 . . . v̂ N ]T Global Structure Total displacements and rotations at structure nodes.

 
ua Translations: ua , rotations: Ra , for a = 1, . . . N .
with v̂a =
Ra
δv = [ δv1 . . . δv N ]T Global Structure Incremental displacements and spins at structure
 
δua nodes used in incremental-iterative solution procedure.
with δva =
δωa Translations: δua , spins: δωa ; conjugate forces: na and
ma , respectively, for a = 1, . . . N .

δ v̄e = [ δ v̄e1 . . . δ v̄eN e ]T Local CR Element Localization of above to element e in CR frame. Trans-
 e
δ ūa lations: δ ūae , spins: δ ω̄ae ; conjugate forces: n̄ae and m̄ae ,
with δ v̄ae =
δ ω̄ae respectively, for a = 1, . . . N e .

v̄ed = [ v̄ed1 . . . v̄ed N e ]T Local CR Element Deformational displacements and rotations at element
 ūe  e
with v̄eda = da
e
nodes. Translations: ūeda , rotations: θ̄da ; conjugate
θ̄da forces: n̄a and m̄a , respectively, for a = 1, . . . N e .
N = number of nodes in structure; N e = number of nodes in element e; a, b : node indices.

§12.3.6. Degrees of Freedom

For simplicity it will be assumed that an N e -node CR element has six degrees of freedom (DOF) per node:
three translations and three rotations. This assumption covers the shell and beam elements evaluated in
Part II [R.38]. The geometry of the element is defined by the N e coordinates xa0 , a = 1, . . . N e in the
base (initial) configuration, where a is a node index.
The notation used for DOFs at the structure and element level is collected in Table 12.2. If the structure
has N nodes, the set {ua , Ra } for a = 1, . . . N collectively defines the structure node displacement vector
v. Note, however, that v is not a vector in the usual sense because the rotators Ra do not transform as
vectors when finite rotations are considered. The interpretation as an array of numbers that defines the
deformed configuration of elements is more appropriate.
The element total node displacements ve are taken from v in the usual manner. Given ve , the key CR
operation is to extract the deformational components of the translations and rotations for each node. That
sequence of operations is collected in Table 12.3. Note that the computation of the centroid is done by
simply averaging the coordinates of the element nodes. For 2-node beams and 3-node triangles this is
appropriate. For 4-node quadrilaterals this average does not generally coincides with the centroid, but
this has made little difference in actual computations.

§12.3.7. EICR Matrices

Before studying element deformations, it is convenient to introduce several auxiliary matrices: P =
Pu − Pω , S, G, H and L that appear in expressions of the EICR front-end. As noted, elements treated
here possess N e nodes and six degrees of freedom (DOF) per node. The notation and arrangement used
for DOFs at different levels is defined in Table 12.2. Subscripts a and b denote node indices that run from
1 to N e . All EICR matrices are built node-by-node from node-level blocks. Figure 12.9(a) illustrates
the concept of perturbed configuration C D + δC, whereas Figure 12.9(b) is used for examples. The CR
and deformed configuration are “frozen”; the latter being varied in the sense of variational calculus.

12–13
Chapter 12: CR FORMULATION OVERVIEW I 12–14

Table 12.3. Forming the Deformational Displacement Vector.

Step Operation for each element e and node a = 1, . . . a

1. From the initial global nodal coordinates xae compute centroid position ae =
N e
e
xC0 = (1/N e ) a=1 xae . Form rotator Te0 as per element type convention.
Compute node coordinates in the element base frame: x̃ae = Te0 (xae − ae ).

2. Compute node coordinates in deformed (current) configuration: xae = xae + uae

N e
and the centroid position vector be = xCe = (1/N e ) a=1 xae . Establish the
deformed local CR system Te by a best-fit procedure, and Re0 = Te (Te0 )T . Form
local-CR node coordinates of CR configuration: x̄eRa = Te (xae − be ).

3. Compute the deformational translations ūda = x̄ae − x̄eRa . R̃d = Tn Ra T0T Com-
e
pute the deformational rotator R̄da = Te Rae (Te0 )T . Extract the deformational
e
angles θ̄da
e
from the axial vector of R̄da .

The translational projector matrix Pu or simply T-projector is dimensioned 6N e × 6N e . It is built from

3 × 3 numerical submatrices Uab = (δab − 1/N e ) I, in which I is the 3 × 3 identity matrix and δab the
Kronecker delta. Collecting blocks for all N e nodes and completing with 3 × 3 zero and identity blocks
as placeholders for the spins and rotations gives a 6N e × 6N e matrix Pu . Its configuration is illustrated
below for N e = 2 (e.g., bar, beam, spar and shaft elements) and N e = 3 (e.g., triangular shell elements):
 2 
3
I 0 − 13 I 0 − 13 I 0
 1   0
2
I 0 − 12 I 0  I 0 0 0 0
 0   − 1 2
− 1 
Ne = 2: Pu = 
I 0 0  , Ne = 3: Pu =   3
I 0 3
I 0 3
I 0 
 . (12.11)
 −1I 0 1
I 0   0 0 0 I 0 0 
2 2  1 
0 0 0 I − I 0 − I 0 1 2
I 0
3 3 3
0 0 0 0 0 I

For any N e ≥ 1 it is easy to verify that P2u = Pu , with 5N e unit eigenvalues and N e zero eigenvalues. Thus
Pu is an orthogonal projector. Physically, it extracts the deformational part from the total translational
displacements.
Matrix S is called the spin-lever or moment-arm or matrix. It is dimensioned 3N e × 3 and has the
configuration (written in transposed form to save space):

S = [ −S1T I −S2T I . . . −STN e I ]T , (12.12)

in which I is the 3 × 3 identity matrix and Sa are node spin-lever 3 × 3 submatrices. Let xa =
[ x1a x2a x3a ]T generically denote the 3-vector of coordinates of node a referred to the element centroid.
Then Sa = Spin(xa ). The coordinates, however, may be those of three different configurations: C0 , C R
and C D , referred to two frame types: global or local. Accordingly superscripts and overbars (or tildes)
are used to identify one of six combinations. For example
   R   D 
0 −x3a 0
−a3 x2a
0
−a2 0 −x̄3a R
x̄2a 0 −x̄3aD
x̄2a
Sa0 =  x3a −a1  , S̄a =  x̄3a R 
, S̄a =  x̄3a D 
R D
0
−a3 0 −x1a0 R
0 −x̄1a D
0 −x̄1a ,
−x2a −a2 x1a −a1
0 0
0 −x̄2a x̄1a
R R
0 −x̄2a x̄1a
D D
0
(12.13)

12–14
12–15 §12.3 COROTATIONAL KINEMATICS

(a) Instantaneous (b)

rotation axis _
Corotated CR
x3 Perturbed C D + δC _
– CR C _ δu2j
δω x2 _
C + δC jD _ δu1j
Deformed CD
x–Da _ jR δu3j(+up)
δc– x2 _
perturbed a _ x1
x1 a _
C + δC –
δω3 CR C _
δθ x3 (+up)
x– aD x– δω
– x
–D
a δv– a (includes iR Corotated CR
displacements _
Perturbed C + δC
D
and spins) δu2i iD _ Deformed CD
_ δu1i
δu3i (+up) L

Figure 12.9. Concept of perturbed configuration to illustrate derivation of EICR matrices:

(a) facet triangular shell element moving in 3D space; (b) 2-node bar element also in 3D
but depicted in the {x̄1 , x̄2 } plane of its CR frame. Deformations grossly exaggerated for
visualization convenience; strains and local rotations are in fact infinitesimal.

are node spin-lever matrices for base-in-global-frame, CR-in-local-frame and deformed-in-local-frame,

R D
respectively. The element matrix (12.12) inherits the notation; in this case S0 , S̄ and S̄ , respectively.
For instance, S matrices for the 2-node space i − j bar element pictured in Figure 12.9(b) are 12 × 3. If
the length of the bar in C D is L, the deformed bar spin-lever matrix referred to the local CR frame is
 T
0 0 0 0 0 0 0 0 0 0 0 0
D
S̄ = 12 L 0 0 1 0 0 0 0 0 −1 0 0 0 . (12.14)
0 −1 0 0 0 0 0 1 0 0 0 0
The first row is identically zero because the torque about the bar axis x̄1 vanishes in straight bar models.
Matrix G, introduced by Haugen [R.37], is dimensioned 3 × 6N e , and will be called the spin-fitter
matrix. It links variations in the element spin (instantaneous rotations) at the centroid of the deformed
configuration in response to variations in the nodal DOFs. See Figure 12.9(a). G comes in two flavors,
global and local:

def
def
N e
δω = G δve = G δvae ,
a a
δ ω̄ = Ḡ δ v̄e = Ḡ δ v̄ae ,
a a
with a
≡ a=1
. (12.15)

Here the spin axial vector variation δωe denotes the instantaneous rotation at the centroid, measured
in the global frame, when the deformed configuration is varied by the 6N e components of δve . When
referred to the local CR frame, these become δ ω̄e and δ v̄e , respectively. For construction, both G and
Ḡ may be split into node-by-node contributions using the 3 × 6 submatrices Ga and Ḡa shown above.
As an example, G matrices for the space bar element shown in Figure 12.9(b) is 3 × 12. The spin-lever
matrix in C D referred to the local CR frame is
 
D 1 0 0 0 0 0 0 0 0 0 0 0 0
Ḡ = 0 0 1 0 0 0 0 0 −1 0 0 0 . (12.16)
L 0 −1 0 0 0 0 0 1 0 0 0 0

The first row is conventionally set to zero as the spin about the bar axis x̄1 is not defined by the nodal
freedoms. This “torsion spin” is defined, however, in 3D beam models by the end torsional rotations.

12–15
Chapter 12: CR FORMULATION OVERVIEW I 12–16

Unlike S, the entries of G depend not only on the element geometry, but on a developer’s decision: how
the CR configuration C R is fitted to C D . For the triangular shell element this matrix is given in Appendix
B. For quadrilateral shells and space beam elements it is given in Part II [R.38].
D D
Matrices S̄ and Ḡ satisfy the biorthogonality property

G S = D. (12.17)

where D is a 3 × 3 diagonal matrix of zeros and ones. A diagonal entry of D is zero if a spin component
D D
is undefined by the element freedoms. For instance in the case of the space bar, the product Ḡ S̄ of
(12.14) and (12.16) is diag(0, 1, 1). Aside from these special elements (e.g., bar, spars, shaft elements),
D = I. This property results from the fact that the three columns of S are simply the displacement
vectors associated with the rigid body rotations δ ω̄i = 1. When premultiplied by G one merely recovers
the amplitudes of those three modes.
The rotational projector or simply R-projector is generically defined as Pω = S G. Unlike the T-projector
Pu such as those in (12.11), the R-projector depends on configuration and frame of reference. Those are
R R R
identified in the usual manner; e.g., P̄ω = S̄ω Ḡω . This 6N e × 6N e matrix is an orthogonal projector of
rank equal to that of D = G S. If G S = I, Pr has rank 3. The complete projector matrix of the element
is defined as
P = Pu − Pω . (12.18)
This is shown to be a projector, that is P2 = P, in Section 4.2.
Two additional 6N e × 6N e matrices, denoted by H and L, appear in the EICR. H is a block diagonal
matrix built of 2N e 3 × 3 blocks:

H = diag [ I H1 I H2 . . . I H N e ] , Ha = H(θa ), H(θ) = ∂θ/∂ω. (12.19)

Here Ha denotes the Jacobian derivative of the rotational axial vector with respect to the spin axial vector
evaluated at node a. An explicit expression of H(θ) is given in (R.48) of Appendix A. The local version
in the CR frame is

H̄ = diag [ I H̄d1 I H̄d2 . . . I H̄d N e ] , H̄da = H̄(θ̄da ), H̄(θ̄d ) = ∂ θ̄d /∂ ω̄d . (12.20)

L is a block diagonal matrix built of 2N e 3 × 3 blocks:

L = diag [ 0 L1 0 L2 . . . 0 L N e ] , La = L(θa , ma ). (12.21)

where ma is the 3-vector of moments (conjugate to δωa ) at node a. The expression of L(θ, m) is
provided in (R.49) of Appendix A. The local form L̄ has the same block organization with La replaced
by L̄a = L(θ̄da , m̄a ).

§12.3.8. Deformational Translations

Consider a generic point P 0 of the base element of Figure 12.5, with global position vector x0P . P 0 rigidly
moves to P R in C R with position vector x RP = x0P + u RP = x0P + c + x RPC . Next the element deforms to
occupy C D . P R displaces to P, with global position vector x P = x0P + u P = x0P + c + x RPC + ud P .
The global vector from C 0 to P 0 is x0P − a, which in the base frame becomes x̃0P = T0 (x0P − a). The
global vector from C R ≡ C to PR is x RP −b, which in the element CR frame becomes x̄ RP = T R (x RPC −b).

12–16
12–17 §12.3 COROTATIONAL KINEMATICS

But x̃0P = x̄ RP since the C0 →C R motion is rigid. The global vector from PR to P is ud P = x P − x RP ,
which represents a deformational displacement. In the CR frame this becomes ūd P = T R (x P − x RP ).
The total displacement vector is the sum of rigid and deformational parts: u P = ur P + ud P . The rigid
displacement is given by expressions collected in (12.5), of which ur P = (R0 − I) (x0P − a) + c is the
most useful. The deformational part is extracted as ud P = u P − ur P = u P − c + (I − R0 ) (x0P − a).
Dropping P to reduce clutter this becomes

ud = u − c + (I − R0 ) (x0 − a). (12.22)

The element centroid position is calculated by averaging its node coordinates. Consequently
N e
c = (1/N e ) u ,
b b
ua − c = U u ,
b ab b
with b
≡ b=1
(12.23)

in which Uab = (δab − 1/N e ) I is a building block of the T-projector introduced in the foregoing
subsection. Evaluate (12.22) at node a, insert (12.23), take variations using (12.10) to handle δR0 , use
(12.2) to map R0 (x0 − a) = x R − b, and employ the cross-product skew-symmetric property (R.3) to
extract δω:

δuda = δ(ua −c) − δR0 (xa0 −a) = U δub − Spin(δω) R0 (xa0 −a)
b ab

= Uab δub − Spin(δω) (x R
a − b) = U δub + Spin(xaR − b) δω
b ab
(12.24)
b

= U δub +
b ab
S R Gb δvb .
b a

Here matrices S and G have been introduced in (12.12)–(12.15). The deformational displacement in the
T
element CR frame is ūd = T R ud . From the last of (12.5) we get ūd = ū − c̄ − (I − R̄0 )x̄ R , where
R̄0 = T R R0 TTR . Proceeding as above one gets
D
δ ūda = U δ ūb +
b ab b
S̄a Ḡb δ v̄b . (12.25)

D
The node lever matrix SaR of (12.24) changes in (12.25) to S̄a , which uses the node coordinates of the
deformed element configuration.

§12.3.9. Deformational Rotations

Denote by R P the rotator associated with the motion of the material particle originally at P 0 ; see
Figure 12.6. Proceeding as in the translational analysis this is decomposed into the rigid rotation R0
and a deformational rotation: R P = Rd P R0 . The sequence matters because Rd P R0 = R0 Rd P . The
order Rd P R0 : rigid rotation follows by deformation, is consistent with those used by Bergan, Rankin
and coworkers; e.g. [R.54,R.56]. (From the standpoint of continuum mechanics based on the polar
decomposition theorem [R.81, Sec. 37] the left stretch measure is used.) Thus Rd P = R P R0T , which
can be mapped to the local CR system as R̄d = T R Rd TTR . Dropping the label P for brevity we get

Rd = R R0T = R T0T T R , R̄d = T R Rd TTR = T R R T0T . (12.26)

The deformational rotation (12.26) is taken to be small but finite. Thus a procedure
 to extract a rotation
axial vector θd from a given rotator is needed. Formally this is θ̄d = axial Loge (R̄d )], but this can be
prone to numerical instabilities. A robust procedure is presented in Section A.11. The axial vector is
evaluated at the nodes and identified with the rotational DOF.

12–17
Chapter 12: CR FORMULATION OVERVIEW I 12–18

Evaluating (12.26) at a node a, taking variations and going through an analysis similar to that carried
out in the foregoing section yields

∂θda ∂ωda  
δθda = δω b = H a δab [ 0 I ] − Gb δvb .
∂ωda b ∂ω
b
b
(12.27)
∂ θ̄da ∂ ω̄da  
δ θ̄da = δ ω̄b = H̄a δab [ 0 I ] − Ḡb δ v̄b .
∂ ω̄da b ∂ ω̄
b
b

where Gb is defined in (12.15) and Ha in (12.19).

Overview continues in next Chapter.

12–18
 11
.

The Core Congruential

Formulation:
Stiffness Equations

11–1
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–2

TABLE OF CONTENTS

Page
§11.1. DCCF Transformation to Physical Freedoms 11–3
§11.2. DCCF Transformation Examples 11–4
§11.2.1. The Bar Element . . . . . . . . . . . . . . . . 11–4
§11.2.2. Iso-P Plane Stress Element . . . . . . . . . . . . 11–4
§11.3. The Generalized CCF 11–5
§11.3.1. Generalized Coordinates as Generic Target . . . . . . . 11–6
§11.3.2. Algebraic Transformation . . . . . . . . . . . . . 11–6
§11.3.3. Differential Transformation . . . . . . . . . . . . . 11–7
§11.3.4. Multistage Transformation . . . . . . . . . . . . 11–7
§11.4. A 2-Node 2D Timoshenko Beam Element 11–8
§11.4.1. Generalized Coordinates and Stress Resultants . . . . . . 11–8
§11.4.2. Transformation Matrices . . . . . . . . . . . . . 11–9
§11.4.3. Internal Force Vector . . . . . . . . . . . . . . . 11–10
§11.4.4. Tangent Stiffness Matrix . . . . . . . . . . . . . 11–10
§11.4.5. Can a Secant Stiffness be Constructed? . . . . . . . . . 11–13
§11.5. A 2-Node 3D Timoshenko Beam Element 11–13
§11.5.1. Transformation to Generalized Gradients . . . . . . . . 11–13
§11.5.2. Transformation to the Rotational Vector . . . . . . . . 11–18
§11.5.3. Transformation to Finite Element Freedoms . . . . . . . 11–19
§11.6. Equivalence of DCCF and Standard TL Formulation 11–20
§11.7. References 11–22
§11. Exercises . . . . . . . . . . . . . . . . . . . . . . 11–24

11–2
11–3 §11.1 DCCF TRANSFORMATION TO PHYSICAL FREEDOMS

The present Chapter completes the development of the Core Congruential Formulation (CCF) of
Total Lagrangian (TL) elements. It present procedures to transform core equations to finite element
stiffness equations.

§11.1. DCCF Transformation to Physical Freedoms

Core expressions for the internal-force vector and stiffness matrices of an individual TL element are
given in (10.20)and (10.26)-(10.29), respectively. These expressions pertain to material particles
of the structure. The behavior of each particle is expressed in terms of its displacement gradients
collected in vector g. To create a discrete model the structure is subdivided into finite elements.
Finite elements equations in terms of the physical DOFs collected in vector v are constructed
through a combination of core-to-physical transformations and integration over element domains.
In this section we stay within the scope of the Direct CCF by assuming that the transformations
between g and v are linear. Because all subsequent developments pertain to an individual element,
no element identifiers are used to reduce indexing clutter.
Over an individual element the displacement field uT = (u 1 , u 2 , u 3 ) is interpolated as

u = N v, (11.1)

where v now collects the element node-displacement degrees of freedom (DOFs) and N =
N(X 1 , X 2 , X 3 ) is a matrix of shape functions independent of v. Differentiating (11.1) with re-
spect to the X i and taking the first two v variations yields

g = Gv, δg = G δv, δ 2 g = 0, (11.2)

(for the last one see Remark 10.2). Invariance of the strain energy variations δU and δ 2 U obtained
by integrating (10.14)-(10.15) over the element reference volume yields
  
K =
U T U
G S G d V, K =
r T r
G S G d V, K= GT SG d V, (11.3)
V0 V0 V0
  
f= GT Φ d V, p= GT Ψ d V, p0 = GT Ψ0 d V, (11.4)
V0 V0 V0

Although the dependency of Slevel and Ψ on g is not made implicit in these equations, it must be
remembered that the transformation g = Gv also appears there. Because of the ensuing algebraic
complexity, numerical integration is generally required unless the gradients are constant over the
element.
Often G is expressed as a chain of transformations, some of which are position dependent and
remain inside the element integral whereas others are not and may be taken outside. For example,
in the bar element treated below, G = T Ḡ, where Ḡ transforms g to local node displacements
while T transforms local to global node displacements.

11–3
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–4

§11.2. DCCF Transformation Examples

§11.2.1. The Bar Element

The core equations for a geometrically nonlinear TL bar were derived in Section 5.1. These equa-
tions are now applied to the formulation of a two-node, linear-displacement, prismatic TL bar
element. The element has constant reference area A0 and initial length L 0 . The two end nodes are
located at (X 1 , Y1 , Z 1 ) and (X 2 , Y2 , Z 2 ), respectively. The node displacements are (v X 1 , vY 1 , v Z 1 )
and (v X 2 , vY 2 , v Z 2 ). The element displacement field in local coordinates { X̄ , Ȳ , Z̄ } may be inter-
polated as
 v̄ 
X1
   
ū X N1 0 0 N2 0 0  v̄Y 1 
 v̄ 
ū = ū Y = 0 N1 0 0 N2 0  Z 1  = N v̄, (11.5)
 v̄ X 2 
ū Z 0 0 N1 0 0 N2 v̄Y 2
v̄ Z 2

where N1 = 1 − X̄ /L 0 and N2 = X̄ /L are linear shape functions. Differentiating with respect to

the reference coordinate we get
 
1 −1 0 0 1 0 0 1
G= 0 −1 0 0 1 0 v̄ = Ḡv̄, (11.6)
L0 0 0 −1 0 0 1 L0

This transformation may be applied to the core matrices and vectors derived in Chapter 10. For
example, application to the core tangent stiffness (10.38) yields

1 T A0 T
K̄ = 2 Ḡ SḠ d V = Ḡ (EbbT + sH) Ḡ, (11.7)
L0 V0 L0

Finally, transformation to node displacements (v Xi , vY i , v Zi ), i = 1, 2 is handled in the usual

manner by writing the local-to-global transformation equation
    
ū X TX X TX Y TX Z uX
ū = ū Y = TY X TY Y TY Z uY = T u, (11.8)
ū Z TZ X TZ Y TZ Z uZ

which is valid for both end nodes giving v̄i = Tv, i = 1, 2. Consequently the element tangent
stiffness matrix in local coordinates is given by
   
A0 TT 0 T T 0
K= Ḡ (Ebb + sH) Ḡ
T
. (11.9)
L0 0 TT 0 T

For this simple element all entries may be obtained in closed form and no numerical integration is
necessary.

11–4
11–5 §11.3 THE GENERALIZED CCF

§11.2.2. Iso-P Plane Stress Element

For the case of plane stress considered in §10.5.2, we shall assume that the associated finite elements
are isoparametric displacement models with n nodes, and that (as usual for such models) the nodal
freedoms are of translational type. The transformation to physical DOFs can then be handled within
the purview of the DCCF.
As in §10.5.2 the reference system, current system and in-plane displacement components are
denoted by {X, Y }, {x, y} and {u X , u Y }, respectively. The element nodes are located at {X i , Yi },
(i = 1, . . . n) in the reference configuration C0 and move to {xi = X i + u Xi , yi = Yi + u Y i },
(i = 1, . . . n) in the current configuration C. The element displacement field may be expressed as
 
vX 1
     vY 1 
uX N1 0 N2 . . . 0  
 v X 2  = N v,
= (11.10)
uY 0 N1 0 . . . Nn  ..  

.
vY n
in which Ni are appropriate isoparametric shape functions written in terms of natural coordinates
such as ξ and η for quadrilaterals. The G matrix follows upon differentiation with respect to X
and Y , and all core equations transformed as per (11.3)–(11.4). For example, the physical tangent
stiffness is 
K= GT (S M + SG )G d V, (11.11)
V0
where S M and SG are given by (10.45) and (10.46), respectively. As in the case of linear elements,
(11.11) is most conveniently evaluated by numerical integration. Because several of the integrand
matrices are sparse, in the interest of efficiency in the computer implementation the integrand
may be symbolically evaluated through a computer algebra system such as Macsyma, Maple
or Mathematica, and automatically converted to Fortran or C program statements before being
encapsulated in the Gauss quadrature loop.
§11.3. The Generalized CCF
As discussed in §10.2, the Generalized Core Congruential Formulation or GCCF is required when
the relation between displacement gradients g and finite element degrees of freedom v is nonlinear.
This complication occurs in elements with rotational freedoms, such as beams, plates and shells, if
finite rotations are exactly treated.
Recall the expression (10.15) of the second variation δ 2 U of the internal energy density. This
expression has the core tangent stiffness S as kernel of the quadratic form in δg. The core internal
T
force Φ also appears in the inner product δ 2 g Φ. This second term may either survive or drop
out depending on the relation of g with the target physical or generalized coordinates (the latter
term is explained below) chosen in the CCF transformation phase. In the case of the DCCF, this
term drops out and
S = S M + SG (11.12)
is the tangent core stiffness, which forward transforms as per (11.3). This is the situation considered
so far. But if that term survives two things happen. First, (11.12) is relabeled as
S = S M + SG P , (11.13)

11–5
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–6

in which S and SG P are called the principal core tangent stiffness and principal geometric stiffness,
T
respectively. Second, transforming the term δ 2 g Φ to freedoms v produces a extra term in
accordance with the schematics
T
K = K M + KG P + KGC , S M → K M , SG P → KG P , δ 2 g Φ → δvT KGC δv, (11.14)

where → and → symbolize DCCF-transformation and GCCF-transformation-styles, respectively.

As can be seen the transformation phase produces a new term KGC called the complementary
geometric stiffness. That term cannot be expressed in terms of the variation δg of the displacement
gradients. Consequently there is no “core complementary core geometric stiffness” SGC that can be
added to (11.13). Instead it appears as a “carry forward term” that materializes as a quadratic-form
kernel upon transforming.

§11.3.1. Generalized Coordinates as Generic Target

For elements that require the GCCF treatment a one-shot transformation between g and v is often
replaced by a multistage transformation. The degree of freedom sets used as intermediate targets
of this process will be collectively referred to as “generalized coordinates” and identified as q.
Of course the final target: element node displacements v, is a particular instance of such array of
choices.
In §10.2 it was noted that two variants of the GGCF, qualified as algebraic and differential, should be
distinguished in terms of consequences on the existence of physical stiffness equations at various
variational levels. These variants are examined below. The ensuing development examines the
transformation from displacement gradients g to a “generic target” set of generalized coordinates
qi collected in vector q. These coordinates are assumed to be independent, a restriction removed
later. Symbols K and f are used to denote tangent stiffness matrices and internal force vectors,
respectively, in terms of q.

§11.3.2. Algebraic Transformation

The Algebraic GCCF, or AGCCF, applies if the relation between g (source) and q (target) is nonlinear
but algebraic. We have g = g(q) or in index notation, gi = gi (q j ). Differentiating with respect to
the qi variables yields
∂gi
δgi = δq j = G i j δq j , or δg = G δq,
∂q j
(11.15)
∂ 2 gi ∂gi 2 0
δ gi =
2
∂q j ∂qk
δq j δqk +
∂q j

δ q j = Fi jk δq j δqk , or δ g = (F δq) δq,
2

Here (Fδq) is the matrix Fi jk δqk = Fki j δqk ; F being a cubic array. The array G receives the name
tangent transformation matrix. The second term in the expansion of δ 2 gi vanishes because the qi
are assumed to be independent target variables.
Enforcing invariance of δ 2 U yields the tangent stiffness transformation

 T 
K= G (S M + SG P )G + Q d V = K M + KG P + KGC = K M + KG , (11.16)
V0

11–6
11–7 §11.3 THE GENERALIZED CCF

where the entries of Q are (cf. Remark 4.1) Q i j = Q ji = Fki j k with summation on k = 1, . . . n g .
Note that Q is symmetric because Fki j = Fk ji . Integration of Q over V0 yields the complementary
portion KGC of the geometric stiffness KG .
The internal, applied and prestress force vectors transform according to the formulas in (11.4) with
the G defined in (11.15):
  
f= T
G Φ d V, p= T
G Ψ d V, p =
0
GT Ψ0 d V. (11.17)
V0 V0 V0

What happens to KU and Kr ? They can be obtained, somewhat artificially, by constructing the
matrix equation
g = Wq, (11.18)
where W is called a secant transformation matrix. Generally this matrix is far from unique because
its n g × n q entries must satisfy only n g conditions. (Care has often to be given to the q j → 0 if
0/0 limits appear in W.) Using (11.18) we can proceed to form
 
K =
U T U
W S W d V, K =
r
GT Sr W d V. (11.19)
V0 V0

Because in general W = G, symmetry in the secant stiffness Kr cannot be expected even if Sr is

symmetric.

Remark 11.1. The AGGCF is applicable to finite elements with degrees of freedoms that include fixed-axis
rotations, because such rotations are integrable. Examples are provided by two-dimensional beams as well as
plane stress (membrane) elements with drilling freedoms if only in-plane motions are allowed.

Remark 11.2. Why is KGC called a geometric stiffness? Because it vanishes if the current configuration is
stress free, in which case the core internal force Φ vanishes and so does Q.

§11.3.3. Differential Transformation

The Differential GCCF, or DGCCF, is required if the relation between g (source) and q (target) is
only available as a non-integrable differential form between their variations:

δgi = G i j δq j , or δg = G δq,
∂G i j (11.20)
δ 2 gi = δq j δqk = Fi jk δq j δqk , or δ 2 g = (Fδq) δq.
∂qk

The transformation equation (11.16) still applies for K whereas (11.17) holds for the force vectors.
But no integral g = g(q) as in the AGGCF exists. Consequently KU and Kr , which require a
secant matrix relation of the form (11.18), cannot be constructed. Furthermore Q is not necessarily
symmetric; a condition for that being Fki j = Fk ji or equivalently ∂G ki /∂q j = ∂G k j /∂qi .

Remark 11.3. For mechanical finite elements the DGCCF naturally arises when three-dimensional finite
rotations are present as nodal degrees of freedom, because such rotations are non-integrable.

11–7
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–8

Remark 11.4. The relations (11.20) have points of resemblance with the case of non-holonomic constraints
in analytical dynamics.

§11.3.4. Multistage Transformation

Up to this point the q have been assumed to be independent variables. But as previously noted,
for complicated elements the GCCF transformations are more conveniently applied in stages. The
target variables in one stage become the source variables for the next one.
What happens if the q are intermediate variables in a transformation chain? If the q are linear in the
final independent degrees of freedom v, all previous formulas hold because the DCCF applies for
the remaining transformations, which are strictly congruential. But if the q are nonlinear in v, or
only a non-integrable differential relation exists, term (∂gi /∂q j ) δ 2 q j = G i j δ 2 q j in the second of
(11.15) survives. The net effect is that the geometric stiffness acquires a higher order component,
implicitly defined as the kernel of 
i G i j δ 2 q j d V, (11.21)
V0
This term cannot be resolved (“resolution” meaning explicit extraction of its stiffness kernel in the
form of a complementary geometric stiffness) until the transformation chain reaches downstream
variables that either are the final degrees of freedom (and thus independent), or depend linearly
on such. It is difficult to state detailed rules that encompass all possible situations. Instead the
treatment of the 2D and 3D beam element transformations in §11.4 and §11.11 illustrates the basic
techniques for “carrying forward” terms such as (11.21).

§11.4. A 2-Node 2D Timoshenko Beam Element

We continue here with the derivation of a 2D, isotropic Timoshenko beam element started in §10.5.4.
This example serves to illustrate the Algebraic GCCF. The specific element constructed here has
two end nodes, six degrees
 of freedom, and reference length L 0 . The cross section area A ≡ A0 and
 I = A Y d A are constant along the element. Axis X is made to pass through the
2
moment of inertia
centroid so that A Y d A = 0. Furthermore it is assumed that the cross section is doubly symmetric
so that A Y 3 d A = 0.
The element displacement field, defined by u 0X (X ), u 0Y (X ) and θ(X ), is interpolated with linear
shape functions:
 
vx1
   v 
u 0X N1 0 0 N2 0 0  y1 
θ 
u 0Y = 0 N1 0 0 N2 0  1  = Nv, (11.22)
 vx2 
θ 0 0 N1 0 0 N2  
v y2
θ2
where N1 = 1 − (X/L 0 ) and N2 = 1 − N1 = X/L 0 . Consequently
∂u 0X vX 2 − vX 1 ∂u 0Y vX 2 − vX 1 ∂θ θ2 − θ1
= = , γ = = , κ= = , (11.23)
∂X L0 ∂X L0 ∂X L0
are constant over the element.

11–8
11–9 §11.4 A 2-NODE 2D TIMOSHENKO BEAM ELEMENT

§11.4.1. Generalized Coordinates and Stress Resultants

As intermediate set of generalized coordinates we take qT = [  γ κ θ ]. These four quantities
are constant over each cross section and may be viewed as cross-section orientation coordinates.
Consequently when obtaining stiffness matrices and internal forces in terms of q it is convenient to
integrate over the beam cross section. The resulting quantities appear naturally in terms of cross
section stress-resultants as shown below. In terms of these generalized coordinates the auxiliary
vectors bi listed in (10.61) become

       
1 + g1 1 +  − Y κ cos θ g3 − sin θ
 g2   γ − Y κ sin θ   1 + g4   cos θ 
b1 =  = , b2 =  = ,
0 0 1 + g1 1 +  − Y κ cos θ
0 0 g2 γ − Y κ sin θ
(11.24)
The well known stress resultants of beam theory are the axial force N , transverse shear force V
and bending moment M. They are obtained by integrating the PK2 stresses over the beam cross
section: 

N= s1 d A = E A  + 12 ( 2 + γ 2 ) + 12 E I κ 2 + N 0 ,
A
 0
V = s2 d A = G A s ω γ + V 0 , (11.25)
A0

M= s1 Y d A = −E I κω + M 0 ,
A0

where ω = (1 + ) cos θ + γ sin θ and ωγ = γ cos θ − (1 + ) sin θ can be viewed as generalized

skew strains. In (11.25) N 0 , V 0 and M 0 denote initial-stress resultants (stress resultants in C0 , also
called prestress forces), A ≡ A0 , I = A0 Y 2 d A, and As = µA, in which µ is the usual shear
correction factor of Timoshenko beam theory. Because of  the doubly-symmetric cross-section
assumption, a term containing the third-section-moment A0 Y 3 d A has been omitted from the
expression for M.
In addition to N , V and M, the following higher order moment, which is absent from the linear
theory, appears in the residual force and tangent stiffness:

 
C= s1 Y 2 d A = E I ( + 12 ( 2 + γ 2 ) + 12 EHκ 2 + C 0 , (11.26)
A0

in which H = A0 Y 4 d A. If terms in κ 2 are neglected,

C − C 0 = (N − N 0 )(I /A) = (N − N 0 )r 2 , (11.27)

√
where r = I /A is the radius of gyration of the cross section. If such terms are retained this
relation is only exact if r 2 = H/I and approximate otherwise.

Remark 11.5. One may verify that sY
A 2
d A vanishes identically. This serves as a check of the strain
distribution equations.

11–9
 11–10
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

§11.4.2. Transformation Matrices

The differential relations required to establish the tangent transformation are obtained from (10.57)
as   
1 0 −Y cos θ Y κ sin θ δ
∂g  0 1 −Y sin θ −Y κ cos θ   δγ
δg = δq =     = G1 δq, (11.28)
∂q 0 0 0 − cos θ δκ
0 0 0 − sin θ δθ
   δv X 1 
−1 0 0 1 0 0
∂q 1  0 −1 0 0 1 0  δvY 1 
δq = δv =    ..  = G2 δv, (11.29)
∂v L0 0 0 −1 0 0 1 . 
0 0 L0 − X 0 0 X δθ2
The transformation relating δg = G δv may be obtained as the product

 
−1 0 Y (cos θ + (L 0 − X )κ sin θ) 1 0 Y (− cos θ + X κ sin θ )
1  0 −1 Y (sin θ − (L 0 − X )κ cos θ) 0 1 Y (− sin θ − X κ cos θ ) 
G = G1 G2 =  
L0 0 0 −(L 0 − X ) cos θ 0 0 −X cos θ
0 0 −(L 0 − X ) sin θ 0 0 −X sin θ
(11.30)
but it is more instructive (as well as conducive to higher efficiency in the computer implementation)
to perform the transformation phase in two stages.
Observe that the first transformation (from g to q) is nonlinear and algebraic whereas the second
one (from q to v) is linear. Consequently we have to use the AGCCF for the first transformation
but the second one can be done simply through the DCCF.

§11.4.3. Internal Force Vector

The internal force vector in terms of q, denoted by fq , is obtained from the core expression (10.64)
for Φ and the matrix G1 given in (11.28):
 q  
 f N (1 + ) − Mκ cos θ − V sin θ
q
 f   N γ − Mκ sin θ + V cos θ 
fq = G1T Φ d A0 =  γq  =  . (11.31)
A0 fκ −Mω + Cκ
q
fθ −Mκωγ − V ω
Finally, application of (11.29) and integration over the element length yields
 q 
− f
 − fγ
q 
 L0  
− fq + 1L fq 
 κ 0 θ 
f= G2T fq d X =  2
. (11.32)
 f
q

0
 
 fγ
q 
q q
f κ + 12 L 0 f θ
This vector satisfies translational equilibrium.

11–10
11–11 §11.4 A 2-NODE 2D TIMOSHENKO BEAM ELEMENT

§11.4.4. Tangent Stiffness Matrix

Transforming to generalized coordinates q produces three components of the tangent stiffness
matrix: 
T  q q q
K =
q
G1 (S M + SG )G1 + Q d A = K M + KG P + KGC . (11.33)
A0
q
The entries of KM , obtained through symbolic manipulation, are
q
K M (1, 1) = E A(1 + )2 + G As sin2 θ + E I κ 2 cos2 θ,
q
K M (1, 2) = E A(1 + )γ − G As sin θ cos θ + E I κ 2 sin θ cos θ,
q

K M (1, 3) = E I κ (1 + )(1 + cos2 θ ) + γ sin θ cos θ ,
q
K M (1, 4) = E I κ 2 ωγ cos θ + G As ω sin θ,
q
K M (2, 2) = E Aγ 2 + G As cos2 θ + E I κ 2 sin2 θ,
q
 (11.34)
K M (2, 3) = E I κ (1 + ) sin θ cos θ + γ (1 + sin2 θ ) ,
q
K M (2, 4) = E I κ 2 ωγ sin θ − G As ω cos θ,
q
K M (3, 3) = E I ω 2 + EHκ 2 ,
q
  
K M (3, 4) = E I κ (1 + )γ (cos2 θ − sin2 θ + γ 2 − (1 + )2 sin θ cos θ ,
q
K M (4, 4) = E I κ 2 φg2 + G As ω 2 .

The principal geometric stiffness, which is readily worked out by hand, is

 
N 0 −M cos θ Mκ sin θ − V cos θ
q  N −M sin θ −Mκ cos θ − V sin θ 
KG P =   (11.35)
C 0
symm Cκ 2
q
The new term contributed by the AGCCF to Kq is the complementary geometric stiffness KGC . Its
source is the matrix Q introduced in Section 8.2. The entries of Q are Q i j = (∂ 2 gk /∂qi ∂q j )k ,
where the components of g and Φ = s1 b1 + s2 b2 may be obtained from (10.57) and (11.24),
respectively.
The entries of Q were symbolically generated by the following Mathematica module:

QmatrixOf2DTimoBeamElement[eps_,gamma_,kappa_,theta_,Em_,Gm_,Y_]:=
Module[{g,h1,h2,H1,H2,e1,e2,s1,s2,b1,b2,phi,i,j,k},
q={eps,gamma,kappa,theta}; phi={1,1,1,1};
g={eps-Y*kappa*Cos[theta],gamma-Y*kappa*Sin[theta],
-Sin[theta],Cos[theta]-1};
gg={{g[[1]]},{g[[2]]},{g[[3]]},{g[[4]]}};
h1={{1},{0},{0},{0}}; h2={{0},{1},{1},{0}};
H1={{1,0,0,0},{0,1,0,0},{0,0,0,0},{0,0,0,0}};
H2={{0,0,1,0},{0,0,0,1},{1,0,0,0},{0,1,0,0}};
e1=(Transpose[h1].gg+(1/2)*Transpose[gg].H1.gg)[[1,1]];
e2=(Transpose[h2].gg+(1/2)*Transpose[gg].H2.gg)[[1,1]];
s1=Simplify[Em*e1]; s2=Simplify[Gm*e2];

11–11
 11–12
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

b1={1+eps-Y*kappa*Cos[theta],gamma-Y*kappa*Sin[theta],0,0};
b2={-Sin[theta],Cos[theta],1+eps-Y*kappa*Cos[theta],
gamma-Y*kappa*Sin[theta]};
phi=Simplify[s1*b1+s2*b2];
Q=Table[0,{4},{4}];
For[i=1,i<=4,i++, For[j=1,j<=4,j++, For[k=1,k<=4,k++,
Q[[i,j]]=Q[[i,j]]+(D[D[g[[k]],q[[i]]],q[[j]]])*phi[[k]] ]]];
Return[Q]
];

The output of this module was integrated over the cross section and pattern matched with the
expression of the stress resultants (11.25)-(11.26) to produce
 
0 0 0 0
q  0 0 0 
KGC =  , (11.36)
0 −Mωγ
symm −V ωγ + Mκω − Cκ 2
which added to (11.35) yields the geometric stiffness
 
N 0 −M cos θ Mκ sin θ − V cos θ
q  N −M sin θ −Mκ cos θ − V sin θ 
KG =  . (11.37)
C −Mωγ
symm −V ωγ + Mκω
q
Finally, the tangent stiffness in terms of q is Kq = K M + KG . Denoting the entries of Kq by K i j ,
i, j = 1, . . . 4 the tangent stiffness matrix K in terms of node displacements v is formed through
the DCCF transformation
 q q q q
K 11 K 12 K 13 − 12 L 0 K 14
 q q q
K 23 − 12 L 0 K 24
 L0  K 22

 q q q
K 33 − L 0 K 34 + 13 L 20 K 44
K= G2T Kq G2 d X = 

0


symm
q q q q  (11.38)
−K 11 −K 12 −K 13 − 12 L 0 K 14
−K 12
q
−K 22
q q
−K 23 − 12 L 0 K 24
q 


q q
−K 13 + 12 L 0 K 14
q
−K 23 + 12 L 0 K 24
q q
−K 33 + 16 L 20 K 44
q

.
q q q q
K 13 + 12 L 0 K 14 
K 11 K 12 

q
K 22
q q
K 23 + 12 L 0 K 24 
q q q
K 33 + L 0 K 34 + 13 L 20 K 44
The above rule can be applied to K M and KG should separate formation be desirable, as when
setting up a stability eigenproblem.
If the reference configuration is not aligned with X , the preceding expressions apply to the local
system { X̄ , Ȳ }. A final local-to-global transformation step, similar to that discussed for the 3D bar

11–12
11–13 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT

in §11.3, is then necessary. This step can be handled by a simple DCCF transformation, because
the finite rotation θ remains the same in global coordinates.
Remark 11.6. The foregoing exact expressions contain curvature-squared terms typically in the combination
I κ 2 . This can be shown to be of order (r/R)2 compared to other terms, where r is the radius of gyration
of the cross section and R = 1/κ the radius of curvature of the current configuration. For typical beams
(r/R)2 is 10−6 or less; consequently all such tiny terms may be dropped without visible loss of accuracy. For
highly-bent extremely-thin beams, however, that ratio may go up to 0.01 in which case the κ 2 terms might
have a noticeable though small effect if retained.

§11.4.5. Can a Secant Stiffness be Constructed?

To attempt the construction of a secant stiffness Krq in terms of generalized coordinates q one
should obtain a secant matrix form of the relationship g = g(q). As noted previously such form is
far from unique. One possible choice is
    
g1 1 0 −Y cos θ 0 
 g   0 1 −Y sin θ 0 γ 
g= 2=    = W1 q, (11.39)
g2 0 0 0 − sin θ/θ κ
g4 0 0 0 (cos θ − 1)/θ θ
which has the merit of not beingtoo dissimilar from G1 . Note that some care must be taken as regards
some 0/0 limits. Then Krq = A G1T Sr W1 d A, which may be easily worked out in closed form but
L
is unsymmetric. Because q is linear in v, the next transformation is simply Kr = 0 0 G2T Krq G2 d X
which can be handled through a scheme similar to (11.38) but with an unsymmetric kernel matrix.
§11.5. A 2-Node 3D Timoshenko Beam Element
We continue here the development of a two-node 3D Timoshenko beam element started in §10.5.5.
As can be surmised, the development is more complex and demanding than for its 2D counterpart.
Only a summary taken from Crivelli’s thesis [1] and Crivelli and Felippa [3] is presented here. The
transformation phase to pass from the core equations to the element nodal degrees of freedom is
carried out in three stages:
1. From particle displacement gradients g to generalized gradients w at each cross section. An
integration over the cross section area is involved.
2. From generalized gradients w to cross-section orientation coordinates q. The rotational
parametrization is introduced at this stage.
3. From cross-section orientation to finite-element nodal degrees of freedom v. An integration
over the element length, as defined by the shape functions, is involved.
These transformation stages are summarized in Tables 11.1 and 11.2, which together also serve to
define notation

§11.5.1. Transformation to Generalized Gradients

The first set of target variables are the generalized gradients w(X ) at each reference cross section
defined by the longitudinal coordinate X . The components of w are indirectly given through their
first variation:  T
δw = d δu0 dδΘ δΘ , (11.40)
dX dX
11–13
 11–14
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

Table 11.1 Internal energy and its variations for 3D Timoshenko beam element

Core Section Gradients Section Orientation Physical DOF

Particle Cross-Section Cross-Section Whole Element
g w z v

U = 12 gT SU g + gT Ψ0 — — —

δU = δgT (Sr g + Ψ0 ) δUG = δwT R δUz = δzT fz δU = δvT f

δ 2 U = δgT S δg + δ 2 gT Φ δ 2 UG = δwT S δw + F δ 2 Uz = δzT Kz δz δ 2 U = δvT K δv

Table 11.2. Core-to-physical-DOFs transformations for 3D beam element

Core Level Section Gradients Section Orientation Physical DOF

Particle Cross-Section Cross-Section Whole Element
g w z v
  L0
Φ R= T
W ΦdA fz = Z R
T
f= GzT fz d X
A0 0
  L0
S S= WT SW d A Kz = ZT S Z + SGC z K= GzT Kz Gz d X
A0 0

δg = W δw δw = Z δz δz = Gz δv

where δΘ, defined in (10.68), measures the variation of angular orientation. Because this quantity
is not generally integrable for three-dimensional motions, it is not possible to express Θ as a unique
function of the displacements. The variation of g1 is
dδu0 T dδΘ T
 ζ̃T κ,
δg1 = + RT ζ̃ + RT ζ̃ κ̃δΘ + RT δΘ (11.41)
dX dX
T
where we used the relation [1] δκ = dδΘ/d X + κ̃δΘ. On using the commutative law ãb = b̃ a
 = ãb̃ − b̃ã we may rewrite (11.41) as
and Jacobi’s identity ãb
∂ δu0 T ∂ δΘ
δg1 = + RT ζ̃ + RT κ̃T ζ̃ δΘ (11.42)
∂X ∂X

 2 = RT h̃T δΘ and δg3 = RT h̃T δΘ,

For the other gradient vectors we have δg2 = δRT h2 = RT δΘh 2 3

11–14
11–15 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT

which can be collected in matrix form as

   I RT ζ̃T   d δu0   
δg1 RT κ̃T ζ̃ dX W1
 
 
 =  W2 
  δΘ
T
δg =  δg2  =  0 0 RT h̃2 d
  δw = W δw, (11.43)
dX
δg3 0 0
T
RT h̃3 δΘ W3

where I is the 3-by-3 identity matrix and Wi are 3-by-9 matrices. The second variation of g, which
is required for the complementary geometric stiffness, is

 ζ̃ T d δΘ  ζ̃T d δΘ
 ζ̃T κ̃ δΘ + RT δΘ
δ 2 g1 = RT δΘ + RT δΘ
dX dX
 ζ̃T κ̃δΘ + δ 2 RT ζ̃T κ + RT ζ̃T δ 2 κ,
+ RT δΘ (11.44)

δ 2 g2 = δ 2 RT i2 , δ 2 g3 = δ 2 RT i3

At this point it is appropriate to introduce the following section resultants:

P = Aσb + P0 , sb = Eeb ,
Q = µs A + Q0 , τ = τ2 + τ3 , τ2 = Gγ2 h2 , τ3 = Gγ3 h3 ,

Mσ = E I S Ke + Mσ ,
0
IS = ζζT d A, Ke = φ̃ κ, (11.45)
A0

T
Mτ = µt G I P κ + Mτ ,
0
IP = ζ̃ ζ̃ d A.
A0

Here P, Q, Mσ and Mτ are axial forces, shear forces, bending moments and torsional moments,
respectively, at the current configuration C; P0 , Q0 , M0σ and M0τ are similar quantities at the
reference configuration C0 ; µs and µt are transverse-shear and torsion coefficients that account for
the actual shear stress distributions, respectively; and I S and I P are the cartesian and polar inertia
tensors, respectively, of the cross section. Should the axes Y and Z be aligned with the principal
inertia axes the latter simplified to
   
0 0 0 I22 + I33 0 0
IS = 0 I22 0 , IP = 0 I33 0 . (11.46)
0 0 I33 0 0 I22

Because the relation between g and w is of differential type the applicable transformation rules are
those the DGCCF, and no energy or secant stiffness survives. Thus only the internal force vector
R and tangent stiffness S associated with w are derived below.

Internal Force Vector. The generalized internal force vector is

 
R= W ΦdA =
T
si WT bi d A = Rσ + Rτ , (11.47)
A0 i A0

11–15
 11–16
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

where Rσ and Rτ are the contributions of the normal and shear stresses respectively. Detailed
calculations result [1] in the following exact expressions:
 RT (P φ + κ̃M )   
σ RT Q
Rσ =  T
φ̃ Mσ , Rτ =  Mτ . (11.48)
T
 T
Ke Mσ φ̃ Q + κ̃ Mτ
T

For small deformations in which the squared curvature may be neglected, R ≈ I, φ̃ ≈ 

h1 , K e ≈ κ
and κ̃Mσ ≈ 0. If these approximations are made,
   
P h1 Q
 T 
Rσ =   h1 Mσ  , Rτ =  Mτ  . (11.49)
0 T
h1 Q

These resemble the classic linearized theory equations. Furthermore observe that the term P RT φ
corresponds to the internal force of the TL 3D bar.
Tangent Stiffness. For the tangent stiffness we have the decomposition

S = S M + SG P + SGC . (11.50)

Furthermore, since w is nonlinear in downstream variables, the complementary geometric stiffness

splits into two components:
SGC = SGCw + SGCq , (11.51)
where SGCw and SGCq contains terms that depend on the first and second variations, respectively,
of R and κ. The notation is suggested by the fact that SGCw can be merged into SG P to yield the
geometric stiffness SGw = SG P + SGCw , which is associated with the generalized gradients w and
independent of the rotational parametrization selected in the next set of target variables q. On the
other hand, the kernel SGCq cannot be extracted at the w level and must be carried forward to the
q level because it is parametrization dependent. Each of the components in (11.50)-(11.51) may
be expressed as the sum of two contributions, one from the normal stresses and one from the shear
stresses:

S M = S Mσ + S Mτ , SG P = SG Pσ + SG Pτ , SGC x = SGC xσ + SGC xτ , x = w, q. (11.52)

Material Stiffness. The generalized core material stiffness is given by the congruential transforma-
tion  
SM = W SM W d A =
T
E i WT bi biT W d A = S Mσ + S Mτ . (11.53)
A0 i A0

Carrying out the algebraic manipulations one obtains

 RT (φφT + κ̃T I κ̃)R RT κ̃I S φ̃ RT κ̃I S 
Ke

S
S Mσ =E  T
φ̃ I S φ̃
T
Ke  ,
φ̃ I S  (11.54)
symm Ke I S 
T
Ke

11–16
11–17 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT
   
ART I⊥ R 0 ART I⊥ φ̃ 0 0 0
S Mτ = µG  IP I P κ̃ , in which I⊥ = 0 1 0 . (11.55)
T
symm Aφ̃ I⊥ φ̃ + κ̃ I P κ̃
T 0 0 1

The contribution RT φφT R is the core material stiffness of a TL 3D bar.

Geometric Stiffness due to Normal Stresses. It is convenient to work out together all geometric
stiffness terms produced by the normal stresses, i.e.

SGσ = SG Pσ + SGCwσ + SGCqσ = SGwσ + SGCqσ . (11.56)

The appropriate definitions are


SG Pσ = s11 W1T HW1 d A,
A
 0 (11.57)
SGCσ = s11 b1 δ 2 g d A = δwT SGCwσ δw + F(δ 2 R, δ 2 κ),
A0

where F contains SGCq as q level kernel. Carrying out the algebraic manipulations one obtains
  T
σ 
PI RT M σ RT κ̃T M
SGwσ = SG Pσ + SGCwσ =  0  σ φ̃
M  (11.58)
symm  σ κ̃ + κ̃T M
φ̃M  σ φ̃

The term P I corresponds to the core geometric stiffness of the 3D Tl bar.

The higher order term in (11.57) may be expressed as
 T

Fσ (δ 2 R, δ 2 κ) = MσT φ̃δ 2 κ + φT Rδ 2 RT κ̃Mσ δqT V(φ̃ Mσ ) + U(κ̃Mσ ; φ) δq, (11.59)

Consequently
T
SGCqσ = V(φ̃ Mσ ) + U(κ̃Mσ ; φ). (11.60)
Because the next-level target variables q include the finite rotation parametrization, matrices V and
U depend on that choice. They are the source of unsymmetries in the stiffness matrices when certain
rotational parametrizations are adopted, such as the incremental rotation vector. If the rotational
vector is chosen these matrices are symmetric.
Geometric Stiffness due to Shear Stresses. The contribution of the shear stresses to the geometric
stiffness is
SGτ = SG Pτ + SGCwτ + SGCqτ = SGwτ + SGCqτ . (11.61)
The appropriate definitions are

SG Pτ = s12 (W1T HW2 + W2 HW1 ) + s13 (W1T HW3 + W3 HW1 ) d A
A
 0 (11.62)
SGCτ = (s12 b2 + s13 b3 ) δ 2 g d A = δwT SGCwτ δw + Fτ (δ 2 R, δ 2 κ).
A0

11–17
 11–18
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

Carrying out manipulations one obtains the surprisingly simple form for SGwτ
 
0 0 RT Q T
SGwτ = SG Pτ + SGCwτ =  0 0 . (11.63)
symm 0
The terms due to the second variation of g become
Fτ = QT δ 2 RΦ + MτT δ 2 κ. (11.64)
The kernel carried forward to the q level is
SGCqτ = V(Mτ ) + U(Q ; Φ). (11.65)

§11.5.2. Transformation to the Rotational Vector

The second transformation stage passes from w to z, which is a vector of generalized displacements,
also associated with a beam section, which embodies the parametrization of the cross section
rotation:  T  T
du0 dα d δu0 d δα
z= α , δz = δα . (11.66)
dX dX dX dX
Here α denotes the rotational vector parametrization defined by the standard formulas
α = axial (α̃), R = exp(α̃T ), (11.67)
and which may be extracted from R by
arcsin(τ )
α̃ = log R = axial (RT − R), τ = 12 || axial (RT − R)||. (11.68)
2τ
Because only the variations of w are known the relation between w and z is also of differential type:
   d δu0 
I 0 0 dX 
δw = Z δz, or δw =  0 Y(z) dY(z)  δα  ,
d (11.69)
dX dX
0 0 Y(z) δα
in which  
sin |α| sin |α| ααT 1 − cos |α|
Y(α) = I+ 1− − α̃. (11.70)
|α| |α| |α|2 |α|2
On applying the transformations (11.69) we find for the internal force and the material and principal-
geometric components of the tangent stiffness matrix:
fq = ZT (Rσ + Rτ ), K Mq = ZT (S M )Z, KG Pq = ZT (SGw Z. (11.71)
The materialization of the geometric stiffness terms SGCqσ and SGCqτ for the rotational vector
needs additional work. We state here only the final result:
   
0 0 0 0 0 0
U(τ ; Φ) =  0 0 , T(Mτ ) =  0 Tτ1  . (11.72)
τ τ
symm U symm T2

11–18
11–19 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT

where
  
τ
U = c1 τ ΦI + c2 τΦ + Φ τ + c3 τ Φαα + τ α̃ΦI + τ̃ Φα + αΦ τ̃
T T T TT T T T T

  
+ c5 αT Φ ταT + ατT + αT τ αΦT + ΦαT + τT ααT ΦI
+ c4 τT α̃ΦααT + c6 τT ααT ΦααT ,

 T + c3 αMT + c5 ααT M
Vτ1 = c2 M  τ + c7 Mτ αT + αT Mτ I + c8 αT Mτ ααT ,
τ τ
 
dα T
dα T 
dα
T
dα 
dα
Vτ2 = −c3 Mτ I − c4 Mτ ααT + c5 Mτ αT + α MτT + αT Mτ I
dX dX dX dX dX
  
T dα dα dα T
+ c6 α Mτ αα + c7
T
Mτ + Mτ
T
+
dX dX dX
dα T  dα T
+ c8 α αMτT + Mτ αT + αT Mτ I + c9 ααT Mτ ααT ,
dX dX
(11.73)
in which
sin α 1 − cos α sin α − α cos α
c1 = − , c2 = , c3 = ,
α α2 α3
c1 + 3c3 c1 + 2c2 c3 + 4c5
c4 = − , c5 = − , c6 = − , (11.74)
α2 α2 α2
1 + c1 3c3 − 2c2 c5 − 5c8
c7 = , c8 = , c9 = .
α2 α2 α2
A similar approach can be taken with (11.69), which defines Fσ . The tangent stiffness matrix can
be obtained by superposing all contributions.

§11.5.3. Transformation to Finite Element Freedoms

The final stage introduces a finite element representation for the degrees of freedom. The beam
or beam assembly is divided into a set of two-node finite elements. Each of these nodes has three
displacement degrees of freedom and three rotational degrees of freedom corresponding to the three
{α X , αY , α Z } components of the rotational vector α. Each element in turn has twelve freedoms
which are collected in the array vT = {un αn }T where dn collects the six translational freedoms
while αn collects the six rotations. The cross-section state vector z is approximated inside each
element by
 
N 0    
dn dn
z=0 dX  d N = Gz = Gz v. (11.75)
αn αn
0 N
where N is a matrix of linear shape functions. Since δq = Gz δv the final internal force vector f and
tangent stiffness matrix K of each element are obtained through the DCCF transformations
 L0  L0
f= GzT fz d X, K= GzT Kz Gz d X. (11.76)
0 0

11–19
 11–20
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

The choice of shape functions for the rotational vector poses some subtle questions. In small-
deflection analysis it is common practice to select all Timoshenko beam shape functions to be linear
in X . This choice obviously enforces nodal compatibility while preserving constant curvature states.
But for finite deflections a linear interpolation for the rotational vector components cannot exactly
represent a constant curvature state unless the rotations are about a single axis (plane rotations).
The same is true if the rotation matrix R(X ) is interpolated linearly. On the other hand, linear
interpolation of Euler parameters does preserve the constant curvature state. This motivated the
development of an interpolation scheme that starts from the 4 Euler parameters i (X ), i = 0, 1, 2, 3,
i i = 1 that orient the normal of a cross section at X . These are collected in the 4-vector
2

 = [ 0 1 2 3 ]T . Given the eight end values (0) and (L) the interpolation that can copy
a constant curvature vector κ is found to be [1]
 
tan(ζ ) sin(ζ )
(ζ ) = cos(ζ ) 1 − (0) + (L), (11.77)
tan(ζ L ) sin(ζ L )
√
where ζ = 12 κ X , ζ L = 12 κ L, κ = κT κ. The constant curvature vector can be extracted from the
end values through the formula
1     
 − 20 (0)I (0) ,
κ= (L) − 20 (L)I (0) − (0) (11.78)
β2 L
This interpolation is then transformed to the variations in terms of the rotational vector. Details are
provided in Reference [1].
§11.6. Equivalence of DCCF and Standard TL Formulation
The correspondence between the Direct Core Congruential Formulation (DCCF) and the Standard Formulation
(SF) of the Total Lagrangian (TL) kinematic description is established below for 3D continuum finite elements.
This connection was worked out in a course term project [4]. Such elements fit within the DCCF framework
because their physical DOFs (node displacements) are of translational type.
The Standard Formulation is based on the same scheme used for linear finite elements: first interpolate, then
vary. As in the linear case, the departure point is extremization of the Total Potential Energy functional (TPE)
over the element domain:
   
=U −W = e s0 d V +
T 1
2
e Ee d V −
T
u b dV −
T
uT t d S, (11.79)
V0 V0 V0 St0

where as usual conservative dead loading is assumed. In (11.79), b is the prescribed body force field, t are
surface tractions prescribed over portion St0 of the boundary in C0 , and other quantities are as defined in Section
4. The weak equilibrium equations are obtained on making (11.79) stationary:
   
δ = δU − δW = δe s0 d V +
T
δe Ee d V −
T
δu b d V −
T
δuT t d S = 0. (11.80)
V0 V0 V0 St0

The displacement and strain fields are interpolated in terms of the element degrees of freedom v:

u = N v, δu = N δv, δe = B δv, (11.81)

where B = B(v) depends in v but N does not. Substituting these interpolations into (11.80) yields the residual
equilibrium equations
δ = δvT r = δvT (f − p) = 0. (11.82)

11–20
11–21 §11.6 EQUIVALENCE OF DCCF AND STANDARD TL FORMULATION

where    
f= B (s0 + Ee) d V =
T T
B s d V, p= N b dV +
T
Ns d S, (11.83)
V0 V0 V0 St0

where f and p are the internal and external force vectors, respectively, and s = s0 + Ee are the PK2 stresses in
C. Because the variations δv are arbitrary, the residual-force nonlinear equilibrium equation is r = f − p = 0
or f = p. The tangent stiffness matrix is given by
∂r ∂f
K= = , (11.84)
∂v ∂v
because p (for conservative dead loading) does not depend on v. Splitting B = Bc + Bv (v), where Bc is
constant but Bv depends on v, gives the well known decomposition

K = K0 + K D + KG , (11.85)

where K0 , K D and KG denote the linear, initial-displacement and geometric stiffness matrices, respectively.
These are given by 
K0 = BcT EBc d V,

V0

KD = (BcT EBv + BvT EBc + BvT EBv ) d V, (11.86)


V0

KG δv = δBT s d V.
V0

To correlate these standard forms with those produced by the DCCF, we note that the GL strains can be also
split as e = ec + ev , where ec and ev are linear and nonlinear in v, respectively. The latter may be expressed
in terms of the displacement gradients as
ev = 12 Ag, (11.87)
where A is the 6 × 9 matrix
 gT 0 0
 g g2 g3 0 0 0 0 0 0

1 1
0 g2T 0  0 0 0 g4 g5 g6 0 0 0
0 g3T  0 g9 
A= = ,
0 0 0 0 0 0 g7 g8
0  0 (11.88)
g3T g2T 0 0 g7 g8 g9 g4 g5 g6 
 T   
g3 0 g1T g7 g8 g9 0 0 0 g1 g2 g3
g2T g1T 0 g4 g5 g6 g1 g2 g3 0 0 0
in which the displacement gradients are vector-arranged as
 ∂u 1 ∂u 2 ∂u 2 ∂u 3  .
gT = [ g1 g2 ··· g8 g9 ] = ∂ X1 ∂ X1 ··· ∂ X3 ∂ X3 (11.89)

Comparing
δev = 1
2
δA g + 12 A δg = A δg, (11.90)
to the DCCF transformation relation δg = G δv, in which G is independent of v, we see that

Bv = AG. (11.91)

The other expression we require is δAT s, which appears in the geometric stiffness matrix contracted with δv:
 
KG δv = δB s d V =
T
GT δAT s d V. (11.92)
V0 V0

11–21
 11–22
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

It is well known — see for instance Chapter 19 of Zienkiewicz [5] — that

 
s1 I s4 I s5 I
δAT s = M δg = MG δv, with M= s4 I s2 I s6 I , (11.93)
s5 I s6 I s3 I

where I is the 3 × 3 identity matrix and si , i = 1, . . . 6 are components of the PK2 stress tensor ordered
s1 = s11 , s2 = s22 , . . . s6 = s23 . Using this relation, KG can be placed in the standard form

KG = GT MG d V, (11.94)
V0

which by inspection is seen to be the DCCF-transformation of the core geometric stiffness M ≡ SG = si Hi ,

with the Hi matrices defined in (10.10).
To correlate other terms, write the linear part of the GL strains in terms of gradients as
1 0 0 0 0 0 0 0 0

0 0 0 0 1 0 0 0 0
0 0 1
with D =  .
0 0 0 0 0 0
ec = Dg = DG δv, 0 1 (11.95)
0 1 0 0 0 0 0
 
0 0 1 0 0 0 1 0 0
0 0 0 0 0 1 0 1 0

The numerical D matrix can be easily related to the hi vectors introduced in (10.10). Because both D
and G are independent of v it follows that δec = DG δv and consequently Bc = DG. Partitioning A as
[ a1T a2T . . . a6T ] one easily finds that ai = Hi g. Now the following identities can be verified through
simple algebra:

DT ED = E i j hi hiT = S0 ,
DT EA = E i j hi aTj = E i j hi gT H j = S1 , AT ED = S1T ,
(11.96)
AT EA = E i j ai a j = E i j Hi gi gTj H j = S2 = S2T ,
M = si0 Hi + E i j hi gH j + 12 (gT Hi g)H j = si0 Hi + S∗1 + 12 S∗2 = si Hi .

Comparing these to the expressions of §10.4.3 we conclude that

 
K0 = G D EDG d V =
T T
GT S0 G d V,
 
V0 V0

KD = G (D EA + A ED + A EA)G d V =
T T T T
GT S D G d V, (11.97)
 
V0 V0

KG = G MG d V =
T
GT SG G d V,
V0 V0

which displays the equivalence of both formulations when no approximations are made. This proof may be
extended without difficulty to the AGCCF in which case G is a function of v, although as noted in the text that
situation is sometimes mishandled in the Standard Formulation through the introduction of a priori kinematic
approximations. The equivalence between DGCCF and SF is more difficult to prove because there is no TPE
functional from which the latter can be derived, and such connection should be regarded as an open problem.

11–22
11–23 §11.7 REFERENCES

§11.7. References

[1] L. A. Crivelli, A Total-Lagrangian beam element for analysis of nonlinear space structures, Ph. D.
Dissertation, Dept. of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, 1990.
[2] C. A. Felippa and L. A. Crivelli, A congruential formulation of nonlinear finite elements, in Nonlinear
Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner, Springer-Verlag,
Berlin, pp. 283–302, 1991.
[3] L. A. Crivelli and C. A. Felippa, A three-dimensional non-linear Timoshenko beam element based on
the core-congruential formulation, Int. J. Numer. Meth. Engrg., 36, pp. 3647–3673, 1993.
[4] F. Abedzadeh Anaraki, A. Barzegar Mehrabi and H. R. Lofti, Correspondence between CC-TL and
C-TL formulations, in Term Projects in Nonlinear Finite Element Methods, ed. by C. A. Felippa, Report
CU-CSSC-91-12, Center for Space Structures and Controls, University of Colorado, Boulder, CO, May
1991.
[5] O. C. Zienkiewicz, The Finite Element Method, 3r d ed., McGraw-Hill, London, 1976.

11–23
 11–24
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS

Homework Exercises for Chapter 11

Not Assigned

EXERCISE 11.1
Consider a two-node geometrically TL nonlinear bar in three dimensions built with the CCF. Write subroutines
to compute the tangent stiffness matrix K and the PK2 stress s in the current configuration given the following
input data:

(1) The coordinates of both element nodes in the reference configuration global Cartesian system X, Y, Z .
(2) The X, Y, Z displacements v Xi , vY i , v Zi of the two element end nodes i = 1, 2.
(3) The elastic modulus E and the reference bar cross section A0 .
(4) The bar axial stress s 0 in the reference configuration.
The subroutines you have to write (in Fortran, C or C++) have the following names and calling sequence
interface:
BAR3F (X0, v, E, A0, s0, F, status)
BAR3K (X0, v, E, A0, s0, K, status)
BAR3S (X0, v, E, s0, s, status)

BAR3F and BAR3K compute the internal force vector f and tangent stiffness matrix K, respectively, in the current
configuration. Subroutine BAR3S computes the PK2 stress s in the current configuration.
Input variables (Fortran assumed in description):

Variable Declaration Description

X0 double precision x0(3,2) Global coordinates of end nodes.
The coordinates of node i go in the
i th column of X0
v double precision v(3,2) v X , vY , v Z displacements of end nodes.
The displacements of node i go in the
i th column of v
E double precision e Elastic modulus
A0 double precision a0 Reference cross section area
s0 double precision s0 PK2 stress in reference configuration

The output variables are:

Variable Declaration Description

F double precision f(6) internal force vector
K double precision k(6,6) tangent stiffness matrix
S double precision s PK2 axial stress s in current configuration
STAT character*(*) stat Blank if no error; else error message

11–24
11–25 Exercises

For intermediate manipulations, construct the local bar axis X joining nodes 1 to 2 in C0 . Then construct Y ,
and Z normal to X forming a right-handed system. As noted §11.2.1, there is some arbitrariness because Y , Z
may be “gyrated” around X without changing the final answer; select whatever orientation rule seem to be
more computationally “robust” in the sense that it should not fail for arbitrary bar orientations.
The answer for this exercise should be a listing of BAR3F, BAR3K and BAR3S.
EXERCISE 11.2
Explain the rule you chose to orient Y and Z . (Appropriate comments in the BARK source code should be
sufficient to answer this one.)
EXERCISE 11.3
Test the subroutines on the following input data:

Argument Input value(s)

X0 [1.23, 2.34, 3.45, 5.43, 4.32, 3.21 ]
v [0.76, −2.12, 1.67, −2.45, 3.01, −3.28]
E 1.82
A0 0.765
S0 3.21

To feed these values write a short test driver that calls the two subroutines in turn; a suggested driver is listed
under “Programming Recommendations.”
Compare the output of BAR3K with the following tangent stiffness matrix:
 0.935471 0.097502 −0.071172 −0.935471 −0.097502 0.071172

 0.097502 1.622137 −0.511148 −0.097502 −1.622137 0.511148 
 −0.071172 −0.511148 0.511148 −1.295011 
K= 
1.295011 0.071172
 −0.935471 −0.097502 (E11.1)
0.071172 0.935471 0.097502 −0.071172 
 
−0.097502 −1.622137 0.511148 0.097502 1.622137 −0.511148
0.071172 0.511148 −1.295011 −0.071172 −0.511148 1.295011
Compare the output of BAR3F to the following internal force vector:
 −0.912675 
 −6.554668 
 4.784631 
f=
 0.912675 
 (E11.2)
 
6.554668
−4.784631

The computed PK2 stress s in the current configuration returned by BAR3S should be 5.603088.

11–25
 10
.

The Core-Congruential
Formulation:
Core Equations

10–1
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–2

TABLE OF CONTENTS

Page
§10.1. Introduction 10–3
§10.2. Overview 10–3
§10.2.1. Basic Concepts . . . . . . . . . . . . . . . . . 10–4
§10.2.2. Direct and Generalized CCF . . . . . . . . . . . . 10–4
§10.2.3. The CCF Philosophy: Divide and Conquer . . . . . . . 10–5
§10.3. Historical Background 10–6
§10.4. Core Stiffness Equations 10–8
§10.4.1. TL Description of Particle Motion . . . . . . . . . . 10–8
§10.4.2. Energy Variations . . . . . . . . . . . . . . . . 10–9
§10.4.3. Parametrized Forms . . . . . . . . . . . . . . . 10–9
§10.4.4. Spectral Forms . . . . . . . . . . . . . . . . . 10–11
§10.4.5. Generalization to H(g) . . . . . . . . . . . . . . 10–11
§10.5. Core Stiffness Derivation Examples 10–12
§10.5.1. Bar in 3D Space . . . . . . . . . . . . . . . . 10–12
§10.5.2. Plate in Plane Stress . . . . . . . . . . . . . . . 10–14
§10.5.3. Plate Bending . . . . . . . . . . . . . . . . . 10–16
§10.5.4. 2D Timoshenko Beam . . . . . . . . . . . . . . 10–18
§10.5.5. 3D Timoshenko Beam: Kinematics . . . . . . . . . . 10–20
§10.5.6. 3D Timoshenko Beam: Core equations . . . . . . . . 10–23
§10.6. References 10–24
§10. Exercises . . . . . . . . . . . . . . . . . . . . . . 10–25

10–2
10–3 §10.2 OVERVIEW

In Chapter 8 it was noted that two methods for developing geometrically nonlinear elements based
on the Total Lagrangian description exist: the Standard Formulation (SF) and the Core-Congruential
Formulation or CCF. Chapters 10 and 11 cover the second method. The present exposition is taken
largely from a recent survey article [1].

§10.1. Introduction
There is an elegant Total Lagrangian (TL) formulation of geometrically nonlinear mechanical finite
elements that has received little attention in the literature. This will be referred to as the Core-
Congruential Formulation, or CCF, in the sequel. The key concepts, presented by Rajasekaran and
Murray [2] in 1973, evolved from the analysis and reinterpretation of the pioneer work of Mallet
and Marcal [3] as well as Murray’s previous work in geometrically nonlinear finite element analysis
[4]. The discussion of Reference [2] by Felippa [5] provided parametric expressions for the stiffness
matrices that appear at various levels of the discrete governing equations. This work originated
what is called here the Direct Core Congruential Formulation, or DCCF.
In 1987 this course presented the derivation of several elements using the DCCF. Preparation
of homework assignments and feedback from students in this and follow-up offerings helped to
streamline the material. Subsequently Crivelli’s doctoral thesis [6] used the CCF in the systematic
development of a three-dimensional nonlinear Timoshenko beam element capable of undergoing
arbitrarily large rotations. Challenges posed by this application pushed this formulation beyond
frontiers hitherto deemed impassable by a TL element with rotational degrees of freedom. This
development was summarily reported in a survey article by Felippa and Crivelli [7] and explained
in more detail in a subsequent paper by Crivelli and Felippa [8].
A lesson gained from this research is that, when dealing with 3D finite rotations, the CCF should
be applied in a staged fashion that allows the systematic examination of additional terms arising in
the transformations to physical degrees of freedom. That transformation methodology gave rise to
what is here called the Generalized CCF, or GCCF.
Both DCCF and GCCF share the same “divide and conquer” philosophy. However, the core
equations as well as subsequent steps that transform those equations to physical freedoms vary
in complexity. To simplify the exposition while focusing on the essential aspects, Sections 10.3
through 10.7 focus on the DCCF. Examples of application to elements amenable to the direct
treatment are presented. The GCCF is discussed in Chapter 11, and illustrated with applications to
2D and 3D beam elements.

Remark 10.1. Several authors have expressed the belief that the approximation performance of TL-based
elements degrades beyond moderate rotations, and an updated Lagrangian or corotational description is nec-
essary for handling truly large motions. For example, in 1986 Mathiasson, Bengtsson and Samuelsson [9]
concluded that “The TL formulation can only be used in problems with small or moderate displacements.”
More recently Bergan and Mathisen [10] voice a similar opinion: “it is commonly known that in a step by step
TL formulation artificial strains easily arise in beam elements due to nonhomogeneities in the displacement
expansions in transverse and longitudinal directions.” Our experience shows that such limitations are not in-
herent in the TL description but instead emerge when a priori kinematic approximations are made to simplify
element derivations. The 3D beam element just cited exhibits computational and approximation performance
for very large rotations comparable to those based on the co-rotational and Updated Lagrangian descriptions
while retaining certain advantages listed in the Conclusions.

10–3
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–4

§10.2. Overview

§10.2.1. Basic Concepts

The original development of the CCF was concerned with the construction of TL stiffness matrices
for geometrically nonlinear analysis through the congruential-transformation pattern

K level
= GT Slevel G d V, (10.1)
V0

where S is the core stiffness matrix, K the physical stiffness in terms of the nodal degrees of
freedom v, G a core-to-physical-freedom transformation matrix assumed to be independent of
v, V0 the appropriate reference integration volume, and in which “level” identifies the governing
equation level at which the stiffness matrix is used.
The three variational levels of interest in practice are: energy (level 0), force equilibrium (level 1),
and first-order incremental equilibrium (level 2). Qualifiers “residual-force” and “secant-stiffness”
are also used for level 1, and “tangent-stiffness” used for level 2.
The core stiffness matrix is expressed in terms of the displacement gradients at each material point.
Displacement gradients g make a better choice of core variables than finite strains because for
elements with translational degrees of freedom (DOFs) they can be expressed linearly in terms of
node displacements v as g = Gv, a property that validates (10.1) for all levels. As discussed below,
such elements fall under the purview of the Direct CCF.
The qualifier “core” emphasizes the goal of independence of Slevel with respect to discretization
decisions such as element geometry, shape functions, and choice of nodal degrees of freedom.
Such a dependence is introduced by the congruential transformation indicated in (10.1) and the
integration over the element volume.

§10.2.2. Direct and Generalized CCF

The basic schematics of the CCF, mathematically expressed through (10.1), may be diagrammed
as

Core Congruential Physical-DOF

Stiffness ⇒ Transformation ⇒ Stiffness
Equations Equations Equations

But this panoramic view needs to be rendered more precise. If the relation between core DOFs
(the displacement gradients g) and the physical DOFs (the node displacements v of a finite element
model) is linear, these transformations do not depend on level:

(0) Core Energy Stiffness Physical Energy Stiffness

Congruential
(1) Core Secant Stiffness ⇒ Transformation ⇒ Physical Secant Stiffness
(1) Core Internal Force Equations Physical Internal Force
(2) Core Tangent Stiffness Physical Tangent Stiffness

10–4
10–5 §10.2 OVERVIEW

In this diagram, numbers annotated within the “core box” denote the variational level of the gov-
erning equation in use. Internal force and secant stiffness are two alternative governing-equation
expressions at level 1. The energy level (level 0) may also be expressed in several ways, but this is
not shown in the diagrams to reduce clutter. Under the aforementioned assumption we obtain the
Direct Core Congruential Formulation, or DCCF.
If the relation between displacement gradients g and node displacements v is nonlinear, the trans-
formations sketched above are not only more complex but depend on variational level and possibly
the expression form used within a level. This complication arises when elements with rotational
degrees of freedom such as beams, plates and shells are considered. It gives rise to the Generalized
Core Congruential Formulation, or GCCF.
Two variants of the GCCF may be distinguished. If the relation between g and v is nonlinear
but algebraic, the transformation equations do vary with level but in principle are still possible as
illustrated in the following diagram.

S-Congruential
(0) Core Energy Stiffness ⇒ Transformation ⇒ Physical Energy Stiffness
(1) Core Secant Stiffness Equations Physical Secant Stiffness
T-Congruential
(1) Core Internal Force ⇒ Physical Internal Force
Transformation ⇒
(2) Core Tangent Stiffness Equations Physical Tangent Stiffness

Here “T-Congruential” and “S-Congruential” are abbreviations for “Tangent Congruential” and
“Secant-Congruential,” respectively. Such a distinction is elaborated upon in Chapter 11.
If the relation between g and v is nonlinear and can be expressed only in non-integrable differential
form, the “Secant Transformation Equations” of the preceding diagram do not generally exist, and
the diagram must be truncated:

(0) Core Energy Stiffness

(1) Core Secant Stiffness

T-Congruential
(1) Core Internal Force ⇒ Physical Internal Force
Transformation ⇒
(2) Core Tangent Stiffness Equations Physical Tangent Stiffness

These two variants of the GCCF are called Algebraic GCCF and Differential GCCF and denoted by
acronyms AGCCF and DGCCF, respectively, in the sequel. The main distinction between AGCCF
and DGCCF is that it makes no sense to talk about missing quantities, such as the physical secant
stiffness, with the latter.
The original development of the CCF outlined in the Introduction focused on elements with
translational-degree-of-freedom configurations. For such elements the Direct form of the CCF,
or DCCF, is sufficient. Sections §10.3 through §10.7 focus on that form, leaving the development
and application of the GCCF to the next Chapter.

10–5
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–6

§10.2.3. The CCF Philosophy: Divide and Conquer

The CCF derivation of the finite element equations naturally reflects the outlined framework. It
proceeds through two phases: a core phase followed by a transformation phase. In the initial phase
core energy, secant and tangent stiffness matrices as well as internal force vectors are obtained.
These matrices and vectors pertain to individual particles. For the stiffness matrices they are
collectively represented by the term Slevel in (10.1).
The key goal is to try to make such core equations as independent as possible with respect to finite-
element discretization decisions such as element geometry, shape functions, selection of nodal
degrees of freedom and (in the case of rotational DOFs) rotational parametrizations. To emphasize
this independence, the term core was coined. Complete independence is in fact achievable if the
relation between displacement gradients g and v is linear, which characterizes the DCCF. The goal
has to be tempered if the relation is nonlinear because dependencies may arise at the tangent stiffness
level. Such dependencies create the so-called complementary geometric stiffness terms, which are
characteristic of elements that fall under purview of the GCCF.
In the transformation phase, these core forms are transformed to physical DOFs, i.e. element node
displacements. The transformation may be done directly for simple elements and in multistage
fashion for complex ones. In particular, multistage transformations are recommended for elements
that require the Differential GCCF such as 3D beam and shell elements. In this case the trans-
formation phase is decomposed into transformation stages that progressively “bind” particles into
lines, areas or volumes through kinematic constraints, and eventually link the element domain to
the nodal degrees of freedom. Decisions such as the choice of specific parametrizations for finite
rotations may be deferred to final stages.
What are the differences between the CCF and the more conventional Total Lagrangian formulation
of nonlinear finite elements? If kinematic exactness is maintained throughout, the final discrete
equations are identical. This is shown in Appendix 1 for the DCCF applied to continuum elements.
But in geometrically nonlinear analysis approximations of various kinds are common, especially
in structural elements with rotational degrees of freedom such as beams, plates and shells. In the
conventional formulation it is quite difficult to assess a priori the effect of seemingly innocuous
approximations “thrown into the pot,” and a posteriori exhaustive testing of complex situations
becomes virtually impossible. Sample: how does the neglect of higher order terms in the axial
deformation of a spinning 3D beam affects torsional buckling?
The staged approach recommended for the GCCF permits a better control over such assumptions.
The core equations are physically transparent, clearly displaying the effect of material behavior,
displacement gradients and prestresses. In the ensuing transformation sequence the origin of each
term can be accurately traced, and on that basis informed decisions on retention or dropping made.
This process can be aided by computer by testing subproblems that isolate the physics modeled by
specific terms.
From this discussion it follows that, from the standpoint of element development, evaluation and
testing, the most significant advantage that can be claimed for the CCF is the clean separation of
physical effects. The importance of this factor should not be underestimated, because physical
transparency is the key to success in nonlinear analysis.

§10.3. Historical Background

10–6
10–7 §10.3 HISTORICAL BACKGROUND

In 1968 Mallet and Marcal [3] attempted to establish a standard nomenclature for geometrically
nonlinear finite element structural analysis based on the Total Lagrangian (TL) kinematic descrip-
tion. Consider a discrete, finite element model of a static structural system under dead loading
with nodal displacement degrees of freedom collected in array v. Displacements are measured
from a fixed reference configuration C0 to a current configuration C. The virtual-work conjugate
forces, independent of v, are collected in array p. The system has a total potential energy function
 = U − P that is the difference between the strain energy U and the loads potential P = pT v.
The residual node forces are r = ∂/∂v, and the symbol
denotes increment associated with
the variation of the current configuration. (In keeping up with the spirit of Reference [3] actual
variations are used below rather than virtual ones; the latter are identified by the usual δ prefix.)
Mallet and Marcal expressed the total potential energy, the residual (force-balance) equilibrium
equations, and the incremental equilibrium equations as follows:
 
 = U − P = 12 vT K0 + 13 N1 + 16 N2 v − pT v, (10.2)
∂  
r= = K0 + 12 N1 + 13 N2 v − p = 0, (10.3)
∂v

r = [K0 + N1 + N2 ]
v −
p = 0. (10.4)

Here K0 is the linear stiffness matrix evaluated at the reference configuration, whereas N1 and N2
are nonlinear stiffness matrices, also evaluated at the reference configuration, that depend linearly
and quadratically, respectively, on the node displacements v. The N matrices were said “to repeat”
in the foregoing expressions. (This old notation has not survived; presently symbol N is most
commonly used to identify matrices of element shape functions.)
Five years later Rajasekaran and Murray [2] examined more critically the structure of the matrices
that appear in the above equations. In that investigation they chose to start from the “core” stiffness
matrices corresponding to K, N1 and N2 expressed in terms of displacement gradients, and in
doing so laid down the main idea of the CCF. Working with specific elements they showed that
the nonlinear stiffness matrices N1 and N2 are not uniquely determined. Indeed (10.2)-(10.4) as
written are unique only for a single degree of freedom. They did not present, however, a general
expression valid for arbitrary elements. This was partly done by Felippa [5], who in the discussion
of Reference [2] considered again those equations, rewritten here in a more general and compact
form:
 T
 = 12 vT KU v + p0 − p v, (10.5)
r = Kr v + p0 − p = f − p = 0, (10.6)

r = K
v −
p = 0, (10.7)

in which the notation of this paper — rather than that of Reference [5] — is used. Here KU , Kr
and K denote the energy, secant and tangent stiffness matrices, respectively. (Energy and secant
stiffnesses are not denoted by Ke and Ks because such symbols are used for other purposes in
the finite element course noted in the Introduction.) In addition, p0 is the prestress force vector,
which vanishes if the reference configuration is stress free and was omitted in that discussion, [5]
and f = Kr v + p0 is the internal force vector. The tangent stiffness is of course fundamental in
incremental-iterative solution methods and stability analysis, while the secant stiffness (by itself or

10–7
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–8

in the internal-force form Kr v + p0 ) is important in pseudo-force methods. The energy stiffness

enjoys limited application per se but has theoretical importance as source for the other two.
In linear problems KU = Kr = K = K0 and the three stiffness matrices coalesce. But in nonlinear
problems not only do the matrices differ but, as shown in the next section, KU and Kr may involve
arbitrary scalar coefficients. Such parametrized expressions were given by Felippa [5] under the
following restrictions:
(R1) Kr is symmetric.
(R2) The reference configuration is stress free.
(R3) The finite strain measure is quadratic in the displacement gradients.
(R4) The transformation between core and physical freedoms is linear.
The following treatment eliminates restrictions (R1) and (R2) altogether, and the other two se-
lectively. It should be noted that restriction (R4) is the condition that, with present terminology,
characterizes the DCCF.

§10.4. Core Stiffness Equations

§10.4.1. TL Description of Particle Motion

A conservative, geometrically nonlinear structure under dead loading is viewed as a continuum
undergoing finite displacements u. These displacements are measured from a fixed reference
configuration C0 to a variable current configuration C. No discretization into finite elements is
implied at this stage. We confine our attention to the case in which the material behavior stays within
the linear elastic range, thus implying small deformational strains but arbitrarily large rotations.
Corresponding points or particles in the reference and current configuration are referred to a fixed
Cartesian coordinate system and have the coordinates X i and xi (i = 1, . . . n d ), respectively, where
n d is the number of space dimensions. The displacement field components are u i = xi − X i .
Let the state of strain at a particle in the current configuration be characterized by n s strains
ei (i = 1, 2, . . . n s ) collected in an array e, and let the corresponding conjugate stresses be si
(i = 1, 2, . . . n s ), collected in an array s. Using the summation convention the elastic stress-strain
relations are written

si = si0 + E i j e j , with E i j = E ji , or s = s0 + Ee, (10.8)

where si0 are stresses in the reference configuration (stresses that remain if ei = 0, also called
prestresses) and E i j are elastic moduli arranged as a n s × n s square array in the usual manner.
Let π, U, P, Ψ, Φ and Υ denote the analogues of , U , P, p, f and r, respectively, at the particle
level. (The first three acquire the meaning of energy densities, whereas Ψ is a dead-loading body
force density independent of u.) The strain energy density can be expressed as

U = ei si0 + 12 ei E i j e j = eT s0 + 12 eT Ee. (10.9)


The total strain energy U is obtained by integrating (10.9) over the structure volume: U = V0 U d V ;
the integration taking place — as can be expected in a TL description — over the reference config-
uration geometry.

10–8
10–9 §10.4 CORE STIFFNESS EQUATIONS

Next, introduce the n g displacement gradients gmn = ∂u m /∂ X n . These are subsequently identified
as gi (i = 1, 2, . . . n g ) so they can be conveniently arranged in a one-dimensional array g. Following
Rajasekaran and Murray [2] and Felippa [5] assume that the strains ei are linked to the displacement
gradients through matrix relations of the form

ei = hiT g + 12 gT Hi g , i = 1, 2, . . . n s (10.10)
where hi and Hi are arrays of dimension n g × 1 and n g × n g , respectively, with Hi symmetric.
In the original References [2,5] it was assumed that Hi is independent of g, which is the case for
the Green-Lagrange strain measure. This restriction, labeled (R3) in §10.3, will be enforced below
except in §10.4.5.

§10.4.2. Energy Variations

As noted previously, for deriving core equations we regard the displacement gradients g as degrees
of freedom. On substituting (10.8) and (10.10) into (10.9) we obtain the “core counterparts” of
(10.5)–(10.7), in which v has become g:

π = U − P = 12 gT SU g + (Ψ 0 − Ψ)T g, (10.11)
∂H
Υ= = Sr g + Ψ 0 − Ψ = Φ − Ψ = 0, (10.12)
∂g

Υ = S
g −
Ψ = 0. (10.13)

Here SU , Sr and S denote the energy, secant and tangent core stiffness matrices, and Ψ0 , which is
independent of g, is the core counterpart of p0 .
With this notation the first and second variations of the strain energy density can be expressed as
 
δU = δgT (SU g + Ψ0 ) + 12 gT δSU g = δgT Sr g + Ψ0 = δgT Φ, (10.14)
δ 2 U = δgT Sr δg + δgT δSr g + (δ 2 g)T Φ = δgT S δg + (δ 2 g)T Φ. (10.15)
These variational equations implicitly determine Sr , Φ and S from SU and Ψ0 . If the linearity
restriction (R4) holds, the term in δ 2 g drops out as explained in the Remark below, and

δ 2 U = δgT S δg. (10.16)

Remark 10.2. If g = Gv with G independent of v, δ 2 g = G δ 2 v = 0 because v are independent variables.

On the other hand, if displacement gradients are nonlinear functions of node displacements expressable as
gi = gi (v j ), then

∂gi ∂ 2 gi ∂gi 2 0
δgi =
∂v j
δv j = G i j δv j , δ 2 gi =
∂v j ∂vk
δv j δvk +
∂v j
δ v j = Fi jk δv j δvk . (10.17)

Thus δg is still G δv but δ 2 g = (F δv) δv, where F is a cubic array. The presence of the term δ 2 g is taken into
account in the GCCF discussed in Chapter 11.

10–9
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–10

§10.4.3. Parametrized Forms

For convenience introduce the following n g × n g matrices (with summation convention on i, j =
1, . . . n g implied):
S0 = E i j hi h j , S1 = E i j hi gT H j , S∗1 = E i j (hiT g) H j ,
(10.18)
S2 = E i j Hi ggT H j , S∗2 = E i j (gT Hi g) H j ,
in which parentheses are used to emphasize the grouping of scalar quantities such as gT Hi g. It
may be then verified that, if assumptions (R3)-(R4) of §10.3 hold, the core stiffnesses and prestress
vector in (10.13)–(10.15) possess the general form:
SU (α, β) = S0 + 12 α(S1 + S1T ) + (1 − α)S∗1 + 14 βS2 + 14 (1 − β)S∗2 + si0 Hi
= S0 + 12 α(S1 + S1T ) + ( 12 − α)S∗1 + 14 β(S2 − S∗2 ) + 12 (si0 + si )Hi ,
= S0 + 12 α(S1 + S1T ) − αS∗1 + 14 βS2 − 14 (1 + β)S∗2 + si Hi ,
Sr (φ, ψ) = S0 + 12 S1 + φS1T + (1 − φ)S∗1 + 14 (2 − ψ)S2 + 14 ψS∗2 + si0 Hi
= S0 + 12 S1 + φS1T + ( 12 − φ)S∗1 + 14 (2 − ψ)S2 + 14 (ψ − 1)S∗2 + 12 (si0 + si )Hi ,
= S0 + 12 S1 + φS1T − φS∗1 + 14 (2 − ψ)(S2 − S∗ ) + si Hi ,
S = S0 + S1 + S1T + S∗1 + S2 + 12 S∗2 + si0 Hi = S0 + S1 + S1T + S2 + si Hi ,
Ψ 0 = si0 hi .
(10.19)
Here α, β, φ and ψ are arbitrary scalar coefficients in the sense that gT SU g and Sr g are independent
of them. In fact,
Φ = Sr g + Ψ0 = si bi , (10.20)
where bi is defined in (10.25) below. The expressions (10.19) are more general than those originally
given by Felippa [5] because restrictions (R1)-(R2) noted in §10.3 are no longer enforced. Note
that the secant core stiffness Sr becomes symmetric if φ = 1/2.
The “repeatable forms” (10.2)–(10.4) of Mallet and Marcal are obtained if α = β = ψ = 2/3 and
φ = 1/2, in which case the combinations S1 + S1T + S∗1 and S2 + 12 S∗2 become the core counterparts
of N1 and N2 , respectively. But this observation has largely historical interest. More physically
relevant are the following combinations:

S D = S1 + S1T + S2 , S M = S0 + S D ,
(10.21)
SG = S∗1 + 12 S∗2 + si0 Hi = si Hi .
These are the core versions of the initial-displacement, material and geometric stiffness, respec-
tively. The core tangent stiffness is S = S0 + S D + SG = S M + SG .
If the Generalized CCF is required for downstream element development as explained in Chapter
11, SG = si Hi is called the principal core geometric stiffness and is denoted by SG P . In this case
the combination
S = S M + SG P , (10.22)
receives the name principal core tangent stiffness.

10–10
10–11 §10.4 CORE STIFFNESS EQUATIONS

Remark 10.3. Finite element practicioners may be surprised at the nonuniqueness of SU and Sr . It appears to
contradict the fact that, given two square matrices A1 and A2 and an arbitrary nonzero test vector x, A1 x = A2 x
for all x implies A1 = A2 . But this is not necessarily true if A1 and A2 are functions of x. More precisely, the
energy core stiffness is not unique because
gT (S1 − S∗1 )g = 0, gT (S1T − S∗1 )g = 0, gT (S2 − S∗2 )g = 0, (10.23)
and the secant core stiffness is not unique because
(S1T − S∗1 )g = 0, (S2 − S∗2 )g = 0. (10.24)
Adding “gage terms” such as those of (10.24) multiplied by arbitrary coefficients does not change δU and
consequently the secant stiffness acquires two free parameters. Uniqueness holds for the tangent stiffness
because the test vectors are the virtual displacement gradient variations, and S is not a function of δg.

Remark 10.4. Because of (10.23), an additional free parameter appears in SU if unsymmetry is allowed. If
symmetry is enforced the first two gage expressions must be combined to read gT (S1 + S1T − 2S∗1 )g = 0.

§10.4.4. Spectral Forms

There is a more compact alternative expression of the core stiffnesses that offers theoretical as well
as implementational advantages at the cost of some generality. Define vectors bi and ci as
∂ei
ei = ciT g, ci = hi + 12 Hi g, bi = = hi + Hi g. (10.25)
∂g
Then the spectral forms (so called because of the formal similarity of equations (10.26)–(10.28)
with the spectral decomposition of a matrix as the sum of rank-one matrices) are

SU (1, 1) = SU α=β=1 = E i j ci cTj + si0 Hi , (10.26)

Sr (0, 0) = Sr φ=ψ=0 = E i j bi cTj + si0 Hi , (10.27)

Sr ( 12 , 1) = Sr  1 = E i j ci cTj + 12 (si + si0 )Hi , (10.28)
φ= ,ψ=1
2
S= E i j bi bTj + si Hi = S M + SG . (10.29)
Note that Sr ( 12 , 1) is symmetric but Sr (0, 0) is not. It is seen that for energy and secant stiffnesses,
compactness is paid in terms of settling for specific coefficients.
Remark 10.5. The foregoing relations may be easily verified by noting that

E i j ci cTj = S0 + 12 (S1 + S1T ) + 14 S2 ,

E i j bi cTj = S0 + 12 S1 + S1T + 12 S2 ,
E i j bi bTj = S0 + S1 + S1T + S2 ,

T
 
∂(ci cTj ) ∂c j ∂ci (10.30)
Ei j = Ei j ci + cTj = S∗1 + 12 S∗2 = E i j e j Hi = (si − si0 )Hi ,
∂g ∂g ∂g

T
∂ 2 (ci cTj ) ∂ci ∂c j
Ei j = 2E i j = 12 S∗2 ,
∂g2 ∂g ∂g
and seeking these patterns in the general parametrized expressions (10.20).

10–11
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–12

§10.4.5. Generalization to H(g)

If the Hi depend on g, as it generally happens if strain measures other than Green-Lagrange’s

are used, the secant and tangent stiffness core equations become more complex because of the
presence of first and second g-derivatives of Hi . The changes in the core variational equations
(10.14)–(10.15) can be succintly expressed as

δU = δgT (Sr + 
r
 0 = δgT (Φ + Φ),
S )g + Ψ0 + Ψ  (10.31)

δ 2 U = δgT (S +  
S) δg + (δ 2 g)T (Φ + Φ). (10.32)

S ,
where 
r
 are additional core terms that arise on account of the dependence of the Hi on g.
S and Φ
The parametrization and efficient characterization of such terms for several strain measures of
interest in practice, notably logarithmic and midpoint strains, are presently open problems. Such
topics would in fact be good candidates for term projects in advanced nonlinear finite element
courses.

§10.5. Core Stiffness Derivation Examples

Because the core equations reflect the motion of an individual particle, their form is primarily
determined by the choice of components of s, e and g that are retained in the strain energy density.
This choice is in turn a byproduct of the mathematical idealization of the actual structure or structural
component.
Several cases are worked out below to illustrate the basic steps. The core expressions developed in
these examples do not force commitment to specific elements, only to a mathematical model. For
example the bar core equations may be subsequently used to develop 2-node straight elements or
3-node curved ones. Some specific elements based on these equations are derived in Chapter 11.

§10.5.1. Bar in 3D Space

The particle belongs to a bar moving in 3D space. The only energy contribution is due to the axial
(longitudinal) stress. We have n d = 3, n s = 1 and n g = 3. To simplify node subscripting, Cartesian
systems and displacement components will be denoted by {X, Y, Z }, {x, y, z} and {u X , u Y , u Z }
rather than {X 1 , X 2 , X 3 }, {x1 , x2 , x3 } and {u 1 , u 2 , u 3 }, respectively. In the reference configuration
C0 the bar is referred to a local Cartesian system { X̄ , Ȳ , Z̄ }, with X̄ located along the bar axis. See
Figure 10.1.
With reference to this local system, the motion of a particle initially at X̄ is defined by the displace-
ment components ū X = ū X ( X̄ ), ū Y = ū Y ( X̄ ) and ū Z = ū Z ( X̄ ). The three displacement gradients
that intervene in the definition of nonlinear strains are
 
g1 ∂ ū X /∂ X̄
g= g2 = ∂ ū Y /∂ X̄ . (10.33)
g3 ∂ ū Z /∂ X̄

10–12
10–13 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

Z, z

2'
Y, y X, x C

Z̄ v̄Y 1 1'
v̄ Z 2
Ȳ
v̄ Z 1
1
v̄Y 2
C0
E, A0

L0
2 v̄ X 1
X̄
v̄ X 2

Figure 10.1. A 2-node bar element in 3D space.

As uniaxial strain measure we adopt the Green-Lagrange (GL) axial strain, defined as

2
2
2 
∂ ū X ∂ ū X ∂ ū Y ∂ ū Z
e ≡ e1 = + 12 + + = g1 + 12 (g12 + g22 + g32 )
∂ X̄ ∂ X̄ ∂ X̄ ∂ X̄
T  T   (10.34)
1 g1 g1 1 0 0 g1
= 0 g2 + 2 g2
1
0 1 0 g2 = hT g + 12 gT Hg.
0 g3 g3 0 0 1 g3

Thus for this choice of strain, h1T ≡ hT = [ 1 0 0 ] and H1 ≡ H is the 3 × 3 identity matrix.
The conjugate stress measure s1 ≡ s is the second Piola-Kirchhoff (PK2) axial stress. The stress-
strain relation is s = s 0 + Ee, where s 0 and s are PK2 axial stresses in the reference and current
configurations, respectively, and E is Young’s modulus.
Because H is independent of g, to form the core stiffnesses in local coordinates we can directly use
the spectral expressions (10.26)–(10.29). First construct the vectors
1 + 1g  
2 1 1 + g1
c ≡ c1 =  12 g2  , b ≡ b1 = g2 , (10.35)
1
g3 g3
2

which inserted into the spectral forms yield

10–13
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–14

3'
C

4'
C0 3
4 2'

1'
v̄Y 1
v̄ X 1
Y, y
1 2

X, x

Figure 10.2. A 4-node plane stress element in 2D space.

  
(1 + 12 g1 )2 1
g (1 + 12 g1 )
2 2
1
g (1 + 12 g1 )
2 3 1 0 0
SU (1, 1) = E ccT + s 0 H = E  1 2
g
4 2
1
g g
4 2 3
 + s0 0 1 0 , (10.36)
symm 1 2
g 0 0 1
4 3

  
(1 + 12 g1 )2 1
g (1 + 12 g1 )
2 2
1
g (1 + 12 g1 )
2 3 1 0 0
Sr ( 12 , 1) = E ccT + s m H = E  1 2
g
4 2
1
g g
4 2 3
 + sm 0 1 0 , (10.37)
symm 1 2
g 0 0 1
4 3

 

(1 + g1 )2 g2 (1 + g1 ) g3 (1 + g1 ) 1 0 0
S = E bbT + sH = E  g22 g2 g3  + s 0 1 0 . (10.38)
symm g32 0 0 1

In equation (10.37), s m = 12 (s 0 + s) = s 0 + 12 Ee is the average or “half-way” stress. The clean

separation into material and geometric (initial-stress) stiffnesses should be noted.

§10.5.2. Plate in Plane Stress

As second example we consider a particle that pertains to a plate in plane stress (membrane),
constrained to move in its plane. See Figure 10.2. As usual we consider only the motion of
the midplane. The Cartesian reference system and displacement components will be denoted by
{X, Y }, {x, y} and {u X , u Y } rather than {X 1 , X 2 }, {x1 , x2 } and {u 1 , u 2 }, respectively. The element
displacement field of a generic particle originally at (X, Y ) is defined by the two components
u X = u X (X, Y ) and u Y = u Y (X, Y ). Three in-plane PK2 stresses contribute to the strain energy

10–14
10–15 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

and four displacement gradients appear in the corresponding GL strain. Consequently n d = 2,

n s = 3 and n g = 4. The four displacement gradients are arranged as
   
g1 ∂u X /∂ X
 g   ∂u /∂ X 
g= 2= Y  (10.39)
g3 ∂u X /∂Y
g4 ∂u Y /∂Y
The strain measures chosen are the three components ei (i = 1, 2, 3) of the GL strains defined in
the usual manner:
 T  
1 1 0 0 0
0 0 1 0 0
e1 = e X X = g1 + 12 (g12 + g22 ) =   g + 12 gT   g, (10.40)
0 0 0 0 0
0 0 0 0 0
 T  
0 0 0 0 0
0 0 0 0 0
e2 = eY Y = g4 + 12 (g32 + g42 ) =   g + 12 gT   g, (10.41)
0 0 0 1 0
1 0 0 0 1
 T  
0 0 0 1 0
1 0 0 0 1
e3 = e X Y + eY X = g2 + g3 + g1 g3 + g2 g4 =   g + 12 gT   g, (10.42)
1 1 0 0 0
0 0 1 0 0
from which expressions for hi and Hi (i = 1, 2, 3) follow. For brevity, only the derivation of the
tangent stiffness matrix will be described. Begin by forming the vectors
     
1 + g1 0 g3
 g2   0   1 + g4 
b1 =   , b2 =   , b3 =  . (10.43)
0 g3 1 + g1
0 1 + g4 g2
Then from (10.29) we get the core stiffness

S = E i j bi biT + si Hi = S M + SG , (10.44)

where si = si0 + E i j e j , (i, j = 1, 2, 3), are the PK2 stresses in the current configuration.
In full and using the abbreviations a1 = 1 + g1 , a4 = 1 + g4 we get

 E a 2 + 2E a g + E g 2 E 11 a1 g2 + E 13 (a1 a4 + g2 g3 ) + E 33 a4 g3
11 1 13 1 3 33 3
 E 11 g22 + 2E 13 a4 g2 + E 33 a42
SM = 

symm
(10.45)
E 12 a1 g3 + E 13 a12 + E 23 g32 + E 33 a1 g3 E 12 a1 a4 + E 13 a1 g2 + E 23 a4 g3 + E 33 g2 g3 
E 12 g2 g3 + E 13 a1 g2 + E 23 a4 g3 + E 33 a1 a4 E 12 a4 g2 + E 13 g22 + E 23 a42 + E 33 a4 g2 

E 22 g32 + 2E 23 a1 g3 + E 33 a12 E 22 a4 g3 + E 23 (a1 a4 + g2 g3 ) + E 33 a1 g2
E 22 a42 + 2E 23 a4 g2 + E 33 g22

10–15
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–16

 
s1 0 s3 0
 s1 0 s3 
SG =  . (10.46)
s2 0
symm s2

§10.5.3. Plate Bending

This is similar to the previous example in that the structure is a flat thin plate but now motion
in 3D space {X, Y, Z } is allowed. With this increased freedom the plate is capable of membrane
stretching and bending. For the latter a Kirchhoff mathematical model is assumed. The three
energy-contributing GL strains are now functions of six gradients. Consequently n d = 3, n s = 3
and n g = 6. The contributing gradients are arranged as
   
g1 ∂u X /∂ X
 g2   ∂u Y /∂ X 
   
 g3   ∂u Z /∂ Z 
g= =  (10.47)
 g4   ∂u X /∂Y 
   
g5 ∂u Y /∂Y
g6 ∂u Z /∂Y
The three GL strains are defined as
 T  
1 1 0 0 0 0 0
 
0 0 1 0 0 0 0
   
0 1 T 0 0 1 0 0 0
e1 = e X X = g1 + 2 (g1 + g2 + g3 ) =   g + 2 g 
1 2 2 2
 g, (10.48)
0 0 0 0 0 0 0
   
0 0 0 0 0 0 0
0 0 0 0 0 0 0
 T  
0 0 0 0 0 0 0
0 0 0 0 0 0 0
   
0 1 T 0 0 0 0 0 0
e2 = eY Y = g5 + 2 (g4 + g5 + g6 ) =   g + 2 g 
1 2 2 2
 g, (10.49)
0 0 0 0 1 0 0
   
1 0 0 0 0 1 0
0 0 0 0 0 0 1
 T  
0 0 0 0 1 0 0
1 0 0 0 0 1 0
   
0 0 0 0 0 0 1
e3 = e X Y + eY X = g2 + g4 + g1 g4 + g2 g5 + g3 g6 =   g + 12 gT   g,
1 1 0 0 0 0 0
   
0 0 1 0 0 0 0
0 0 0 1 0 0 0
(10.50)
which define hi and Hi , i = 1, 2, 3. When one reaches this level of bookkeeping it is more expedient
and less error-prone to obtain the core matrices through symbolic manipulation. For example, the
following Macsyma program forms S M and SG in matrices SM and SG, respectively:

10–16
10–17 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

h1: matrix([1],[0],[0],[0],[0],[0])$
h2: matrix([0],[0],[0],[0],[1],[0])$
h3: matrix([0],[1],[0],[1],[0],[0])$
g: matrix([g1],[g2],[g3],[g4],[g5],[g6])$
HH1:matrix([1,0,0,0,0,0],[0,1,0,0,0,0],[0,0,1,0,0,0],
[0,0,0,0,0,0],[0,0,0,0,0,0],[0,0,0,0,0,0])$
HH2:matrix([0,0,0,0,0,0],[0,0,0,0,0,0],[0,0,0,0,0,0],
[0,0,0,1,0,0],[0,0,0,0,1,0],[0,0,0,0,0,1])$
HH3:matrix([0,0,0,1,0,0],[0,0,0,0,1,0],[0,0,0,0,0,1],
[1,0,0,0,0,0],[0,1,0,0,0,0],[0,0,1,0,0,0])$
b1:h1+HH1.g$ b2:h2+HH2.g$ b3:h3+HH3.g$
SM:E11* b1.transpose(b1)+E22*b2.transpose(b2)+E33*b3.transpose(b3)
+ E12*(b1.transpose(b2)+b2.transpose(b1))
+ E13*(b1.transpose(b3)+b3.transpose(b1))
+ E23*(b2.transpose(b3)+b3.transpose(b2))$
ratvars(g6,g5,g4,g3,g2,g1,a5,a1,E11,E12,E13,E22,E23,E33)$
SM:ratsimp(SM)$
SG:ratsimp(s1*HH1+s2*HH2+s3*HH3)$

These matrices may be automatically converted to TEX by appropriate Macsyma statements (not
shown above). That output was reformatted by hand for inclusion here. For the core tangent
stiffness this semi-automated process yields

10–17
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–18

SM (1, 1) = E 33 g42 + 2E 13 (1 + g1 )g4 + E 11 (1 + g1 )2

SM (1, 2) = E 13 ((1 + g1 )(1 + g5 ) + g2 g4 ) + E 33 g4 (1 + g5 ) + E 11 (1 + g1 )g2
SM (1, 3) = E 13 ((1 + g1 )g6 + g3 g4 ) + E 33 g4 g6 + E 11 (1 + g1 )g3
SM (1, 4) = E 23 g42 + E 33 (1 + g1 )g4 + E 12 (1 + g1 )g4 + E 13 (1 + g1 )2
SM (1, 5) = E 12 ((1 + g1 )(1 + g5 )) + E 23 g4 (1 + g5 ) + E 33 g2 g4 + E 13 (1 + g1 )g2
SM (1, 6) = E 12 (1 + g1 )g6 + E 23 g4 g6 + E 33 g3 g4 + E 13 (1 + g1 )g3
SM (2, 2) = E 33 (1 + g5 )2 + 2E 13 g2 (1 + g5 ) + E 11 g22
SM (2, 3) = E 33 (1 + g5 )g6 + E 13 (g2 g6 + g3 (1 + g5 )) + E 11 g2 g3
SM (2, 4) = E 33 ((1 + g1 )(1 + g5 )) + E 23 g4 (1 + g5 ) + E 12 g2 g4 + E 13 (1 + g1 )g2
SM (2, 5) = E 23 (1 + g5 )2 + E 33 g2 (1 + g5 ) + E 12 g2 (1 + g5 ) + E 13 g22
SM (2, 6) = E 23 (1 + g5 )g6 + E 12 g2 g6 + E 33 g3 (1 + g5 ) + E 13 g2 g3 (10.51)
SM (3, 3) = E 33 g62 + 2E 13 g3 g6 + E 11 g32
SM (3, 4) = E 33 (1 + g1 )g6 + E 23 g4 g6 + E 12 g3 g4 + E 13 (1 + g1 )g3
SM (3, 5) = E 23 (1 + g5 )g6 + E 33 g2 g6 + E 12 g3 (1 + g5 ) + E 13 g2 g3
SM (3, 6) = E 23 g62 + E 33 g3 g6 + E 12 g3 g6 + E 13 g32
SM (4, 4) = E 22 g42 + 2E 23 (1 + g1 )g4 + E 33 (1 + g1 )2
SM (4, 5) = E 23 ((1 + g1 )(1 + g5 ) + g2 g4 ) + E 22 g4 (1 + g5 ) + E 33 (1 + g1 )g2
SM (4, 6) = E 23 ((1 + g1 )g6 + g3 g4 ) + E 22 g4 g6 + E 33 (1 + g1 )g3
SM (5, 5) = E 22 (1 + g5 )2 + 2E 23 g2 (1 + g5 ) + E 33 g22
SM (5, 6) = E 22 (1 + g5 )g6 + E 23 (g2 g6 + g3 (1 + g5 )) + E 33 g2 g3
SM (6, 6) = E 22 g62 + 2E 23 g3 g6 + E 33 g32

(which can be further compacted by introducing the auxiliary symbols a1 = 1 + g1 and a5 = 1 + g5

as done in §10.5.3) and
 
s1 0 0 s3 0 0
0 s1 0 0 s3 0
 
0 0 s1 0 0 s3 
SG = SG P = . (10.52)
 s3 0 0 s2 0 0
 
0 s3 0 0 s2 0
0 0 s3 0 0 s2

Remark 10.6. If the plate element to which the particle belong has (as usual) rotational freedoms, an additional
geometric stiffness (the complementary geometric stiffness) appears in the transformation phase. Because of
this, the core geometric stiffness (10.52) has been relabeled as SG P , where subscript P means “principal.”

Remark 10.7. The core stiffness matrices may also be used for part of the formulation of thin-shell facet
elements, with the proviso that global reference axes {X, Y, Z } are to be replaced by a local coordinate system
{ X̄ , Ȳ , Z̄ } with Z̄ normal to the element midplane.

10–18
10–19 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

θ2
θ

θ1 T
Rζ

x u u0

X ζ
1 2 X
X0

Figure 10.3. Kinematics of 2D Timoshenko beam element

§10.5.4. 2D Timoshenko Beam

Consider next an isotropic Timoshenko plane beam that moves in the (X ,Y ) plane. For notational
simplicity it is assumed that the longitudinal axis of the beam is aligned with X . The only PK2
stresses that contribute to the strain energy are the axial stress s1 ≡ s X X and the mean shear stress
s2 ≡ s X Y . The corresponding GL strains are the axial strain e1 ≡ e X X and the section-averaged
shear strain e2 ≡ γ X Y = e X Y +eY X . The constitutive equations are s1 = s10 +Ee1 and s2 = s20 +G e2 ,
where E and G are the Young’s modulus and shear modulus, respectively, of the material. The
treatment outlined below is slightly modified from that of a course term project by Alexander, de
la Fuente and Haugen. [11]
The finite displacements are described in a local coordinate system that is attached to the initial
position of the beam, as illustrated in Figure 1. Under the usual kinematic assumptions of the
Timoshenko beam model (plane sections remain plane but not necessarily normal to the deformed
centroidal axis) the coordinates of a particle in the underformed and deformed configurations may
be written    
X 0
X = X0 + ζ, X0 = , ζ= , (10.53)
0 Y
   
X + u 0X cos θ − sin θ
x = x0 + R ζ,
T
x0 = , R =T
, (10.54)
u 0Y sin θ cos θ
where u 0X and u 0Y are the components of the centroidal displacement vector u0 . Subtracting (10.53)

10–19
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–20

from (10.54) gives the element displacement field

   
uX u 0X − Y sin θ
u=x−X= = . (10.55)
uY u 0Y + Y (cos θ − 1)
Four displacement gradients contribute to the GL strains. Thus for this case we have n d = 2, n s = 2
and n g = 4. The four contributing displacement gradients are arranged in the usual pattern:
   
g1 ∂u X /∂ X
 g   ∂u /∂ X 
g= 2= Y . (10.56)
g3 ∂u X /∂Y
g4 ∂u Y /∂Y
For future use in Chapter 11 we note that the gradients can be written in terms of generalized section
freedoms as    
g1  − Y κ cos θ
 g   γ − Y κ sin θ 
g= 2= , (10.57)
g3 − sin θ
g4 cos θ − 1
in which  = ∂u 0X /∂ X is a generalized axial strain, γ = ∂u 0Y /∂ X a generalized shear strain, and
κ = ∂θ/∂ X is the beam curvature.
The matrix form of the GL strains is
 T  
1 1 0 0 0
0 0 1 0 0
e1 = g1 + 12 (g12 + g2 )2 =   g + 12 gT   g, (10.58)
0 0 0 0 0
0 0 0 0 0
 T  
0 0 0 1 0
1 0 0 0 1
e2 = g2 + g3 + g1 g3 + g2 g4 =   g + 12 gT   g, (10.59)
1 1 0 0 0
0 0 1 0 0
which define h1 , h2 , H1 and H2 . On introducing the auxiliary vectors
    1 + 1g   1 
1 + g1 g3 2 1
g
2 3
 g2   1 + g4   12 g2   1 + 1 g4 
b1 =   , b2 =   , c1 =   , c2 =  
 1 + 1 g1  ,
2 (10.60)
0 1 + g1 0 2
0 g2 0 1
g
2 2
the spectral core stiffness matrices and internal force vector can be written
SU = Ec1 c1T + G c2 c2T + s10 H1 + s20 H2 , (10.61)
Sr = Ec1 c1T + G c2 c2T + 12 (s10 + s1 )H1 + 12 (s20 + s2 )H2 , (10.62)
S = S M + SG P , S M = Eb1 b1T + Gb2 b2T , SG P = s1 H1 + s2 H2 , (10.63)
Φ = s1 b1 + s2 b2 . (10.64)
Because beam elements have rotational freedoms, a complementary geometric stiffness matrix
appears when carrying out the transformation phase. This term is considered in the subsequent
GCCF treatment of this element in Chapter 11.

10–20
10–21 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

a2

a1

a3
n2 x0
u0

X, n1
X0

Z n3

Figure 10.4. Kinematics of 3D Timoshenko beam element.

§10.5.5. 3D Timoshenko Beam: Kinematics

The last example of derivation of core equations involve a TL 3D Timoshenko beam capable of
arbitrarily large rotations. The following material is largely extracted from a recent paper by Crivelli
and Felippa [8] as well as Crivelli’s thesis [6] and is continued with the DGCCF transformation
phase in Chapter 11. The notation used in those references has been slightly edited to fit that of the
present article.
As in the 2D case, the beam is isotropically elastic with Young’s modulus E and shear modulus G.
The reference configuration of the beam is straight and prismatic although not necessarily stress
free. A local reference frame ni is attached to it, with n1 directed along the longitudinal axis (the
locus of cross section centroids). Axes n2 and n3 are in the plane of the left-end cross section; these
will be eventually aligned with the principal inertia axes to simplify some algebraic expressions.
Along these axes we attach the coordinate system {X, Y, Z }. This description is schematically
shown in Figure 2. We further define a set of moving frames, denoted by {a1 , a2 , a3 }, parametrized
by the longitudinal coordinate X . Initially these frames coincide with {n1 , n2 , n3 }, and displace
rigidly attached to the cross-sections of the moving current configuration.
A beam particle originally at (X, Y, Z ) displaces to

x(X) = x0 (X ) + RT (X )ζ(Y, Z ), ζT = [ 0 Y Z ], (10.65)

where x0 describes the position of the centroid of the given cross-section, R is a 3-by-3 orthogonal
matrix function that orients the displaced cross section, and ζ is a cross-section position vector. The
displacement field is
u = x − X = u0 + (RT − I)ζ. (10.66)
where u0 (X ) = x0 (X ) − X0 (X ) is the centroidal displacement (see Figure 2).

10–21
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–22

In the sequel 3 × 3 skew-symmetric matrices are consistently denoted by placing a tilde over their
axial 3-vector symbol; for example
 
0 a3 −a2 a1
ã = spin (a) = −a3 0 a1 , a = a2 = axial (ã). (10.67)
a2 −a1 0 a3

The skew-symmetric curvature matrix κ̃ is defined by κ̃ = R(dRT /d X ), which is the rate of change
of the orthogonal rotation matrix R with respect to the longitudinal coordinate. The curvature vector
is κ = axial (κ̃). We shall also require later the variation of angular orientation δΘ, defined as
the axial vector of the skew matrix R δRT :
 = R δRT = −δR RT ,
δΘ 
δΘ = axial (δΘ), (10.68)

All displacement gradients gi j appear in the GL strain measures. To maintain compactness the nine
gradients are partitioned into three 3-vectors:
  
∂u X /∂ X ∂u X /∂Y ∂u X /∂ Z
g1 = ∂u Y /∂ X , g2 = ∂u Y /∂Y , g3 = ∂u Y /∂ Z , (10.69)
∂u Z /∂ X ∂u Z /∂Y ∂u Z /∂ Z

The 9-component gradient vector is gT = [g1T g2T g3T ], but this symbol is not used directly here.
Also introduce the 3-vectors
  
1 0 0
h1 = 0 , h2 = 1 , h3 = 0 . (10.70)
0 0 1
With the help of these quantities, explicit expressions for the displacement gradient vectors g can
be given as
du0 du0 T
g1 = + RT κ̃ζ = + RT ζ̃ κ,
dX dX (10.71)
g2 = (R − I)h2 ,
T
g3 = (R − I)h3 .
T

The only nonzero components of the GL strain tensor can be written

e1 ≡ e11 = h1T g1 + 12 g1T Hg1 ,

e2 ≡ γ12 = 2e12 = h2T g1 + h1T g2 + 12 (g1T Hg2 + g2T Hg1 ), (10.72)
e3 ≡ γ13 = 2e13 = h3T g1 + h1T g3 + 12 (g1T Hg3 + g3T Hg1 ),

where H is here the 3 × 3 identity matrix. Note that from the orthogonality of the rotation matrix
R we find
e22 = h2T g2 + 12 g2T g2
 2 
= R22 − 1 + 12 R21 + (R22 − 1)2 + R23
2
= R22 − 1 + 12 (2 − 2R22 ) = 0,
(10.73)
2e23 = h2T g3 + h3T g2 + g2T g3
= R32 + R23 + R21 R31 + R22 R32 − R32 + R23 R33 − R23 = 0,

10–22
10–23 §10.5 CORE STIFFNESS DERIVATION EXAMPLES

and similarly e33 = 0. This confirms that the only nonzero strains are (10.72).
The strains (10.72) may be rewritten in a more physically suggestive form:

e1 = e11 = eb + e f , γ = γ12 + γ13 ,



du0 T 1 du0 T
eb = h1 + 2 , e2 = γ12 = γ2 + γ 2 = h2T φ + h2T ζ̃ κ, (10.74)
dX dX
T
T
e f = ζT Ke + 12 κT ζ̃ζ̃ κ  ζT Ke , e3 = γ13 = γ3 + γ 3 = h3T φ + h3T ζ̃ κ.

Here eb , e f are stretching and flexural normal strains, γ2 and γ3 represent bending-induced shear
strains, and γ 2 , γ 3 are torsion-induced shear strains. The last term in e f represents a squared-
curvature contribution to flexure, which can usually be neglected (cf. Remark 9.2). The strain
energy stored in the current configuration is
 
 
U= U d A d X, with U = 12 Ee12 + 12 G e22 + e32 + s10 e1 + s20 e12 + s30 e3 . (10.75)
L 0 A0

§10.5.6. 3D Timoshenko Beam: Core equations

The PK2 stresses associated with the GL strains (10.72) are s1 ≡ s11 = s X X , s2 ≡ s12 = s X Y and
s3 ≡ s13 = s X Z . The constitutive equations are s1 = s10 + Ee1 , s2 = s20 + Ge2 and s3 = s30 + Ge3 .
The spectral core stiffnesses can be compactly expressed in terms of the vectors ci = hi + 12 Hgi
and bi = hi + Hgi for i = 1, 2, 3, where no subscript is needed in H ≡ I. Applying the spectral
formulas of §10.4.4 we obtain for the 9 × 9 core energy stiffness
  0 
ESU1 + G(SU2 + SU3 ) GSU4 GSU5 s1 H s20 H s30 H
SU =  
T
GSU4 GSU1 0 + s20 H 0 0 , (10.76)
GSU5
T
0 GSU1 s30 H 0 0

where SU1 = c1 c1T , SU2 = c2 c2T , SU3 = c3 c3T , SU4 = c2 c1T and SU5 = c3 c1T . At the residual level
we obtain for Sr a form similar to (10.76) except that the prestresses si0 , i = 1, 2, 3 have to
be replaced by the midpoint stresses 12 (si0 + si ). The internal force vector conjugate to δg is
Φ = Sr g + Φ0 = Φσ + Φτ , in which
 
s 1 b1 s2 b2 + s3 b3
Φσ = 0 Φτ = s2 b1 , (10.77)
0 s3 b1

represent the contribution of the normal and shear stresses, respectively.

The principal core tangent stiffness matrix S = S M + SG P is obtained from (10.29). The material
stiffness is 
ES1 + G(S2 + S3 ) GS4 GS5
SM = GS4T GS1 0 , (10.78)
GS5T 0 GS1

10–23
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–24

where S1 = b1 b1T , S2 = b2 b2T , S3 = b3 b3T , S4 = b2 b1T and S5 = b3 b1T . The principal geometric
stiffness is
 0 
s1 H s2 H s3 H (s1 + Ee1 )H (s20 + Ge2 )H (s20 + Ge3 )H
S G P = s2 H 0 0 = (s20 + Ge2 )H 0 0 . (10.79)
s3 H 0 0 (s2 + Ge3 )H
0
0 0
 T
The contribution of δ 2 g Φ to the complementary geometric stiffness depends on the target
variables in the ensuing transformation phase. Because this transformation requires the DGCCF, it
is taken up in Chapter 11.
§10.6. References

[1] C. A. Felippa, L. A. Crivelli and B. Haugen, A Survey of the Core-Congruential Formulation for Geo-
metrically Nonlinear TL Finite Elements, Archives of Computational Methods in Engineering, 1, 1994,
pp. 1–48
[2] S. Rajasekaran and D. W. Murray, Incremental finite element matrices, J. Str. Div. ASCE, 99, pp. 2423–
2438, 1973.
[3] R. H. Mallet and P. V. Marcal, Finite element analysis of nonlinear structures, J. Str. Div. ASCE, 94, pp.
2081–2105, 1968.
[4] D. W. Murray, Finite element nonlinear analysis of plates, Ph. D. Dissertation, Dept. of Civil Engineering,
University of California, Berkeley, California, 1967.
[5] C. A. Felippa, Discussion of Reference 1, J. Str. Div. ASCE, 100, pp. 2519–2521, 1974.
[6] L. A. Crivelli, A Total-Lagrangian beam element for analysis of nonlinear space structures, Ph. D.
Dissertation, Dept. of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, 1990.
[7] C. A. Felippa and L. A. Crivelli, A congruential formulation of nonlinear finite elements, in Nonlinear
Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner, Springer-Verlag,
Berlin, pp. 283–302, 1991.
[8] L. A. Crivelli and C. A. Felippa, A three-dimensional non-linear Timoshenko beam element based on
the core-congruential formulation, Int. J. Numer. Meth. Engrg., 36, pp. 3647–3673, 1993.
[9] K. Mathiasson, A. Bengtsson and A. Samuelsson, On the accuracy and efficiency of numerical algorithms
for geometrically nonlinear structural analysis, in Finite Element Methods for Nonlinear Problems, ed.
by P. G. Bergan, K. J. Bathe and W. Wunderlich, Springer-Verlag, Berlin, pp. 3–24, 1986.
[10] P. G. Bergan and K. M. Mathisen, Large displacement analysis of highly flexible offshore structures,
in Nonlinear Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner,
Springer-Verlag, Berlin, pp. 303–331.
[11] S. Alexander, H. M. de la Fuente and B. Haugen, Correspondence between CC-TL and C-TL formulations
A 2D Timoshenko beam element using the Total-Lagrangian Core Congruential Formulation, in Term
Projects in Nonlinear Finite Element Methods, ed. by C. A. Felippa, Report CU-CSSC-91-12, Center
for Space Structures and Controls, University of Colorado, Boulder, CO, May 1991.

10–24
10–25 Exercises

Homework Exercises for Chapter 10

Not assigned

EXERCISE 10.1
Show that the “core” forms gT SU g and Sr g, with SU and Sr given by (10.19), are independent of the coefficients
α, β, φ and ψ. It is sufficient to work this exercise for for i, j = 1, that is, only one stress, strain and modulus
is considered. Keep, however, g, h and H generic.
Note: This is a good exercise in matrix gymnastics. If you are “rusty” in matrix magic, reading the Addendum
below is strongly recommended.

Addendum: Matrix Product Properties

In carrying out the manipulations required by Exercise 10.1, the following properties of matrix products ought
to be kept in mind.

(1) Let x and y be two conforming vectors. Since their inner product is a scalar, obviously

xT y = yT x.

(2) If x and y are two conforming vectors and A a conforming matrix, the quadratic form xT Ay is also a
scalar; consequently
xT Ay = yT AT x.
Furthermore, if A is square and symmetric, xT Ay = yT Ax.

(3) If x and y are conforming vectors, xyT and yxT are rank-one square matrices (the transpose of each
other). Furthermore, xxT is a symmetric matrix.

(4) Scalars can be moved to any position within a matrix product. If the scalar is in itself the result of a
vector or matrix product, the components may be transposed as per rules (1) and (2). For example, if A,
B and C are conforming matrices, and x and y are conforming vectors,

xT yAB = AxT yB = ABxT y = AByT x, etc.

xT AyBC = yT AT xBC = BxT AyC = ByT AT xC, etc.

10–25
 9
.

The TL
Timoshenko
Plane Beam
Element

9–1
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–2

TABLE OF CONTENTS

Page
§9.1. Introduction 9–3
§9.2. Beam Models 9–3
§9.2.1. Basic Concepts and Terminology . . . . . . . . . . . 9–3
§9.2.2. Mathematical Models: Classical and Timoshenko . . . . . 9–4
§9.2.3. Finite Element Models . . . . . . . . . . . . . . 9–5
§9.2.4. Bernoulli-Euler versus Timoshenko Beam Elements . . . . 9–7
§9.3. X -Aligned Reference Configuration 9–9
§9.3.1. Element Description . . . . . . . . . . . . . . . 9–9
§9.3.2. Motion . . . . . . . . . . . . . . . . . . . 9–9
§9.3.3. Displacement Interpolation . . . . . . . . . . . . . 9–12
§9.3.4. Strain-Displacement Relations . . . . . . . . . . . 9–12
§9.3.5. *Consistent Linearization . . . . . . . . . . . . . 9–13
§9.4. Arbitrary Reference Configuration 9–14
§9.4.1. Strain-Displacement Matrix . . . . . . . . . . . . 9–14
§9.4.2. Constitutive Equations . . . . . . . . . . . . . . 9–15
§9.4.3. Strain Energy . . . . . . . . . . . . . . . . . 9–16
§9.5. The Internal Force 9–16
§9.6. The Stiffness Matrix 9–17
§9.6.1. The Material Stiffness Matrix . . . . . . . . . . . . 9–17
§9.6.2. Eliminating Shear Locking by RBF . . . . . . . . . 9–18
§9.6.3. The Geometric Stiffness Matrix . . . . . . . . . . . 9–19
§9.7. A Commentary on the Element Performance 9–21
§9.8. Summary 9–22
§9. Exercises . . . . . . . . . . . . . . . . . . . . . . 9–23

9–2
9–3 §9.2 BEAM MODELS

§9.1. Introduction

In the present Chapter the Standard Formulation of Total Lagrangian (TL) kinematics is used to
derive the finite element equations of a two-node Timoshenko plane beam element. This derivation
is more typical of the general case. It is still short, however, of the enormous complexity involved,
for instance, in the FEM analysis of nonlinear three-dimensional beams or shells. In fact the latter
are still doctoral thesis topics.
In the formulation of the bar element in Chapter 8, advantage was taken of the direct expression of
the axial strain in terms of reference and current element lengths. That shortcut bypasses the use
of displacement gradients and coordinate transformations. The simplification works equally well
for bars in three-dimensional space.
A more systematic but lengthier procedure is unavoidable with more complicated elements. The
procedure requires going through the displacement gradients to construct a strain measure. Some-
times this measure is too complex and must be simplified while retaining physical correctness.
Then the stresses are introduced and paired with strains to form the strain energy function of the
element. Repeated differentiations with respect to node displacements yield the expressions of the
internal force vector and tangent stiffness matrix. Finally, a transformation to the global coordinate
system may be required.
In addition to giving a better picture of the general procedure, the beam element offers an illustration
of the treatment of rotational degrees of freedom.

§9.2. Beam Models

§9.2.1. Basic Concepts and Terminology

Beams represent the most common structural component found in civil and mechanical structures.
Because of their ubiquity they are extensively studied, from an analytical viewpoint, in Mechanics
of Materials courses. Such a basic knowledge is assumed here. The following material recapitulates
definitions and concepts that are needed in the finite element formulation.
A beam is a rod-like structural member that can resist transverse loading applied between its
supports. By “rod-like” it is meant that one of the dimensions is considerably larger than the other
two. This dimension is called the longitudinal dimension and defines the longitudinal direction
or axial direction. Directions normal to the longitudinal directions are called transverse. The
intersection of planes normal to the longitudinal direction with the beam are called cross sections,
just as for bar elements. The beam longitudinal axis is directed along the longitudinal direction and
passes through the centroid of the cross sections.1 .
Beams may be used as isolated structures. But they can also be combined to form framework
structures. This is actually the most common form of high-rise building construction. Individual
beam components of a framework are called members, which are connected at joints. Frameworks

1 If the beam is built of several materials, as in the case of reinforced concrete, the longitudinal axis passes through the
centroid of a modified cross section. The modified-area technique is explained in elementary courses of Mechanics of
Materials

9–3
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–4

reference configuration

motion

current configuration

Figure 9.1. A geometrically nonlinear plane framework structure.

can be distinguished from trusses by the fact that their joints are sufficiently rigid to transmit bending
moments between members.
In practical structures beam members can take up a great variety of loads, including biaxial bending,
transverse shears, axial forces and even torsion. Such complicated actions are typical of spatial
beams, which are used in three-dimensional frameworks and are subject to forces applied along
arbitrary directions.
A plane beam resists primarily loading applied in one plane and has a cross section that is symmetric
with respect to that plane. Plane frameworks, such as the one illustrated in Figure 9.1, are assemblies
of plane beams that share that symmetry. Those structures can be analyzed with two-dimensional
idealizations.
A beam is straight if the longitudinal direction is a straight line. A beam is prismatic if the cross
section is uniform. Only straight, prismatic, plane beams will be considered in this Chapter.

§9.2.2. Mathematical Models: Classical and Timoshenko

Beams are actually three-dimensional solids. One-dimensional mathematical models of plane
beams are constructed on the basis of beam theories. All such theories involve some form of
approximation that describes the behavior of the cross sections in terms of quantities evaluated at
the longitudinal axis. More precisely, the element kinematics of a plane beam is completely defined
if the following functions are given: the axial displacement u X (X ), the transverse displacement
u Y (X ) and the cross section rotation θ Z (X ) ≡ θ(X ), where X denotes the longitudinal coordinate
in the reference configuration. See Figure 9.2.
Two beam models are in common use in structural mechanics:
Euler-Bernoulli (EB) Model. This is also called classical beam theory or the engineering beam
theory and is the one covered in elementary treatments of Mechanics of Materials. This model
accounts for bending moment effects on stresses and deformations. Transverse shear forces are
recovered from equilibrium but their effect on beam deformations is neglected. Its fundamental
assumption is that cross sections remain plane and normal to the deformed longitudinal axis. This

9–4
9–5 §9.2 BEAM MODELS

θ Z (X ) ≡ θ(X )

current configuration Current

cross section

Y, y
motion u Y (X )

X, x
u X (X )

reference configuration Reference

X cross section

Figure 9.2. Definition of beam kinematics in terms of the three displacement functions
u X (X ), u Y (X ) and θ (X ). The figure actually depicts the EB model kinematics.
In the Timoshenko model, θ (X ) is not constrained by normality (see next figure).

rotation occurs about a neutral axis that passes through the centroid of the cross section.
Timoshenko Model. This model corrects the classical beam theory with first-order shear deformation
effects. In this theory cross sections remain plane and rotate about the same neutral axis as the EB
model, but do not remain normal to the deformed longitudinal axis. The deviation from normality
is produced by a transverse shear that is assumed to be constant over the cross section.
Both the EB and Timoshenko models rest on the assumptions of small deformations and linear-
elastic isotropic material behavior. In addition both models neglect changes in dimensions of the
cross sections as the beam deforms. Either theory can account for geometrically nonlinear behavior
due to large displacements and rotations as long as the other assumptions hold.

§9.2.3. Finite Element Models

To carry out the geometrically nonlinear finite element analysis of a framework structure, beam
members are idealized as the assembly of one or more finite elements, as illustrated in Figure 9.3.
The most common elements used in practice have two end nodes. The i th node has three degrees of
freedom: two node displacements u Xi and u Y i , and one nodal rotation θi , positive counterclockwise
in radians, about the Z axis. See Figure 9.4.
The cross section rotation from the reference to the current configuration is called θ in both models.
In the BE model this is the same as the rotation ψ of the longitudinal axis. In the Timoshenko

9–5
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–6

current configuration finite element idealization

of current configuration

motion

reference configuration finite element idealization

of reference configuration

Figure 9.3. Idealization of a geometrically nonlinear beam member (as taken,

for example, from a plane framework structure like
the one in Figure 9.1) as an assembly of finite elements.

(a) C1 (BE) model θ2 (b) C 0 (Timoshenko) model θ2

θ1 θ1

Y, y uY 2 uY 2

uY 1 uY 1

X, x
uX 1 uX 2 uX 1 uX 2

1 2 1 2

Figure 9.4. Two-node beam elements have six DOFs, regardless of the model used.

model, the difference γ̄ = ψ − θ is used as measure of mean shear distortion.2 These angles are
illustrated in Figure 9.5.
Either the EB or the Timoshenko model may be used as the basis for the element formulation.
Superficially it appears that one should select the latter only when shear effects are to be considered,
as in “deep beams” whereas the EB model is used for ordinary beams. But here a “twist” appear
because of finite element considerations. This twist is one that has caused significant confusion
among FEM users over the past 25 years.

2 It is ψ − θ instead of θ − ψ because of sign convention, to make e X Y positive.

9–6
9–7 §9.2 BEAM MODELS

normal to reference
θ beam axis X

ψ
−θ
normal to deformed
_ =ψ beam axis
γ
direction of deformed ψ
cross section
90◦ _
// X (X = X)

ds
_
Note: in practice γ << θ; typically 0.1% or less. The magnitude
_
of γ is grossly exaggerated in the figure for visualization convenience.

Figure 9.5. Definition of total section rotation θ and BE section rotation ψ in the
Timoshenko beam model. The mean shear deformation is γ̄ = θ − ψ, which
is constant over the cross section. For small deformations of typical engineering
materials γ̄ << 1; for example, typical values for |γ̄ | are O(10−4 ) radians
whereas rotations ψ and θ may be of the order of several radians.

Although the Timoshenko beam model appears to be more complex because of the inclusion of
shear deformation, finite elements based on this model are in fact simpler to construct! Here are
the two main reasons:
(i) Separate kinematic assumptions on the variation of cross-section rotations are possible, as
made evident by Figure 9.5. Mathematically: θ(X ) may be assumed independenltly of u X (X )
and u Y (X ). As a consequence, two-node Timoshenko elements may use linear variations in
both displacement and rotations. On the other hand a two-node EB model requires a cubic
polynomial for u Y (X ) because the rotation θ(X ) is not independent.
(ii) The linear transverse displacement variation matches that commonly assumed for the axial
deformation (bar-like behavior). The transverse and axial displacements are then said to be
consistent.
The simplicity is even more important in geometrically nonlinear analysis, as strikingly illustrated
in the two-node elements depicted in Figure 9.6. Although as shown in that figure both of these
elements have six degrees of freedom, the internal kinematics of the Timoshenko model is far
simpler.

§9.2.4. Bernoulli-Euler versus Timoshenko Beam Elements

In the FEM literature, a BE-based model such as the one shown in Figure 9.4(a) is called a C 1
beam because this is the kind of mathematical continuity achieved in the longitudinal direction
when a beam member is divided into several elements (cf. Figure 9.3). On the other hand the
Timoshenko-based element shown in Figure 9.4(b) is called a C 0 beam because both transverse
displacements, as well as the rotation, preserve only C 0 continuity.

9–7
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–8

2-node C 1 (cubic) element 2-node C 0 linear-displacement-and-rotations

for Euler-Bernoulli beam model: element for Timoshenko beam model:
plane sections remain plane and plane sections remain plane but not
normal to deformed longitudinal axis normal to deformed longitudinal axis

(a) (b)
C 1 element
with same DOFs

Figure 9.6. Sketch of the kinematics of two-node beam finite element models based on
(a) Euler-Bernoulli beam theory, and (b) Timoshenko beam theory. These
models are called C 1 and C 0 beams, respectively, in the FEM literature.

What would be the first reaction of an experienced but old-fashioned (i.e, “never heard about
FEM”) structural engineer on looking at Figure 9.6? The engineer would pronounce the C 0 element
unsuitable for practical use. And indeed the kinematics looks strange. The shear distortion implied
by the drawing appears to grossly violate the basic assumptions of beam behavior. Furthermore, a
huge amount of shear energy would be require to keep the element straight as depicted.
The engineer would be both right and wrong. If the two-node element of Figure 9.6(b) were
constructed with actual shear properties and exact integration, an overstiff model results. This
phenomenom is well known in the FEM literature and receives the name of shear locking. To avoid
locking while retaining the element simplicity it is necessary to use certain computational devices.
The most common are:
1. Selective integration for the shear energy.
2. Residual energy balancing.
These devices will be used without explanation in some of the derivations of this Chapter. For
detailed justification the curious reader may consult advanced FEM books such as Hughes’.3
In this Chapter the C 0 model will be used to illustrated the TL formulation of a two-node, geomet-
rically nonlinear beam element.
Remark 9.1. As a result of the application of the aforementioned devices the beam element behaves like a BE
beam although the underlying model is Timoshenko’s. This represents a curious paradox: shear deformation

3 T. J. R. Hughes, The Finite Element Method, Prentice-Hall, 1987.

9–8
9–9 §9.3 X -ALIGNED REFERENCE CONFIGURATION

is used to simplify the kinematics, but then most of the shear is removed to restore the correct stiffness.4 As a
result, the name “C 0 element” is more appropriate than “Timoshenko element” because capturing the actual
shear deformation is not the main objective.

Remark 9.2. The two-node C 1 beam element is used primarily in linear structural mechanics. (It is in fact
the beam model used in the “Introduction to FEM” course.) This is because some of the easier-construction
advantages cited for the C 0 element are less noticeable, while no artificial devices to eliminate locking are
needed. The C 1 element is also called the Hermitian beam element because the shape functions are cubic
polynomials specified by Hermite interpolation formulas.

§9.3. X -Aligned Reference Configuration

§9.3.1. Element Description

We consider a two-node, straight, prismatic C 0 plane beam element moving in the (X, Y ) plane, as
depicted in Figure 9.7(a). For simplicity in the following derivation the X axis system is initially
aligned with the longitudinal direction in the reference configuration, with origin at node 1. This
assumption is relaxed in the following section, once invariant strain measured are obtained.
The reference element length is L 0 . The cross section area A0 and second moment of inertia I0
with respect to the neutral axis5 are defined by the area integrals
  
A0 = d A, Y d A = 0, I0 = Y 2 d A, (9.1)
A0 A0 A0

In the current configuration those quantities become A, I and L, respectively, but only L is frequently
used in the TL formulation. The material remains linearly elastic with elastic modulus E relating
the stress and strain measures defined below.
As in the previous Chapter the identification superscript (e) will be omitted to reduce clutter until
it is necessary to distinguish elements within structural assemblies.
The element has the six degrees of freedom depicted in Figure 9.4. These degrees of freedom and
the associated node forces are collected in the node displacement and node force vectors
   
u X1 f X1
 uY 1   fY 1 
   
 θ1   fθ 1 
u= , f= . (9.2)
 u X2   f X2 
   
uY 2 fY 2
θ2 fθ 2

The loads acting on the nodes will be assumed to be conservative.

4 The FEM analysis of plates and shells is also rife with such paradoxes.
5 For a plane prismatic beam, the neutral axis at a particular section is the intersection of the cross section plane X =
constant with the plane Y = 0.

9–9
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–10

(a) (b)
L
_
Y, y θ=ψ−γ
θ(X)
2
2
x P(x,y) ψ

C(xC ,yC ) C(X+uX ,u Y )

xC
C
y uYC
1 uYP 1
yC
uY (X) = u YC
X uXC uX (X) = u XC
X
Po (X,Y)
uXP
Y X, x 1 Co (X) 2
1 Co (X,0) 2
C0
Lo

Figure 9.7. Lagrangian kinematics of the C 0 beam element with X -aligned reference
configuration: (a) plane beam moving as a two-dimensional body; (b) reduction
of motion description to one dimension measured by coordinate X .

§9.3.2. Motion
The kinematic assumptions of the Timoshenko model element have been outlined in §9.2.2. Ba-
sically they state that cross sections remain plane upon deformation but not necessarily normal to
the deformed longitudinal axis. In addition, changes in cross section geometry are neglected.
To analyze the Lagrangian kinematics of the element shown in Figure 9.6(a), we study the motion of
a particle originally located at P0 (X, Y ). The particle moves to P(x, y) in the current configuration.
The projections of P0 and P along the cross sections at C0 and C upon the neutral axis are called
C0 (X, 0) and C(xC , yC ), respectively. We shall assume that the beam cross section dimensions do
not change, and that the shear distortion γ << 1 so that cos γ can be replaced by 1. Then
.
x = xC − Y (sin ψ + sin γ cos ψ) = xC − Y [sin(ψ + γ ) + (1 − cos γ ) sin ψ] = xC − Y sin θ,
.
y = yC + Y (cos ψ − sin γ sin ψ) = yC + Y [cos(ψ + γ ) + (1 − cos γ ) cos ψ] = yC + Y cos θ.
(9.3)
But xC = X + u XC and yC = u XC . Consequently x = X + u XC − Y sin θ and y = u Y C + Y cos θ .
From now we shall call u XC and u Y C simply u X and u Y , respectively, so that the Lagrangian
representation of the motion is

x X + u X − Y sin θ
= (9.4)
y u Y + Y cos θ

9–10
9–11 §9.3 X -ALIGNED REFERENCE CONFIGURATION

in which u X , u Y and θ are functions of X only. This concludes the reduction to a one-dimensional
model, as sketched in Figure 9.7(b).
For future use below it is convenient to define an “extended” internal displacement vector w, and
its gradient or X derivative:
     
u X (X ) dw du X /d X uX

w = u Y (X ) , w = = du Y /d X = u Y , (9.5)
θ(X ) d X dθ/d X θ

in which primes denote derivatives with respect to X . The derivative θ is also as κ, which has the
meaning of beam curvature in the current configuration. Also useful are the following differential
relations
ds

1 + u X = s cos ψ, u Y = s sin ψ, s = = (1 + u X )2 + (u Y )2 , (9.6)
dX
in which ds is the differential arclength in the current configuration; see Figure 9.5. If u X and u Y
are constant over the element,

s = L/L 0 , 1 + u X = L cos ψ/L 0 , u Y = L sin ψ/L 0 . (9.7)

Remark 9.3. The replacement of 1 − cos γ by zero in (9.3) is equivalent to saying that Y (1 − cos γ ) can be
neglected in comparison to other cross section dimensions. This is consistent with the uncertainty in the cross
section changes, which would depend on the normal stress and Poisson’s ratio effects. The motion expression
(9.4) has the virtue of being purely kinematic and of leading to (exactly) eY Y = 0.

9–11
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–12

§9.3.3. Displacement Interpolation

For a 2-node C 0 element it is natural to express the displacements and rotation functions as linear
in the node displacements:
 
u X1
     uY 1 
u X (X ) 1−ξ 0 0 1+ξ 0 0  
 θ1 
w = u Y (X ) = 2 1
0 1−ξ 0 0 1+ξ 0   = N u,
u 
θ(X ) 0 0 1−ξ 0 0 1 + ξ  X2 
uY 2
θ2
(9.8)
in which ξ = (2X/L 0 ) − 1 is the isoparametric coordinate that varies from ξ = −1 at node 1 to
ξ = 1 at node 2. Differentiating this expression with respect to X yields the gradient interpolation:
 
u X1
   u 
u X −1 0 0 1 0 0  Y1 
1  θ 
w = u Y = 0 −1 0 0 1 0  1  = G u. (9.9)
L0 u 
θ 0 0 −1 0 0 1  X2 
uY 2
θ2

§9.3.4. Strain-Displacement Relations

The deformation matrix of the motion (9.4) is
 ∂x ∂x 

∂ X ∂Y 1 + u X − Y κ cos θ − sin θ
F= = (9.10)
∂y ∂y u Y − Y κ sin θ cos θ
∂ X ∂Y
where primes denote derivatives with respect to X , and κ = θ . The displacement gradient matrix
is6

u X − Y κ cos θ − sin θ
GF = F − I = , (9.11)
u Y − Y κ sin θ cos θ − 1
from which the Green-Lagrange (GL) strain tensor follows:

eX X eX Y
e= = 12 (FT F − I) = 12 (G F + GTF ) + 12 GTF G F
eY X eY Y

1 2(u X − Y κ cos θ) + (u X − Y κ cos θ) + (u Y − Y κ sin θ)

2 2
−(1 + u X ) sin θ + u Y cos θ
=2
−(1 + u X ) sin θ + u Y cos θ 0
(9.12)
It is seen that the only nonzero strains are the axial strain e X X and the shear strain e X Y +eY X = 2e X Y ,
whereas eY Y vanishes. Through the consistent-linearization techniques described in the subsection

6 This is denoted by G F to avoid clash with the shape function gradient matrix G = N introduced in (9.9).

9–12
9–13 §9.3 X -ALIGNED REFERENCE CONFIGURATION

below, it can be shown that under the small-strain assumptions made precise therein, the axial strain
e X X can be replaced by the simpler form

e X X = (1 + u X ) cos θ + u Y sin θ − Y κ − 1, (9.13)

in which all quantities appear linearly except θ. The nonzero axial and shear strains will be arranged
in the strain vector

e1 eX X (1 + u X ) cos θ + u Y sin θ − Y θ − 1 e − Yκ
e= = = = . (9.14)
e2 2e X Y −(1 + u X ) sin θ + u Y cos θ γ

The three strain quantities introduced in (9.14):

e = (1 + u X ) cos θ + u Y sin θ − 1, γ = −(1 + u X ) sin θ + u Y cos θ, κ = θ , (9.15)

characterize axial strains, shear strains and curvatures, respectively. These are collected in the
following generalized strain vector:

hT = [ e γ κ] (9.16)

Because of the assumed linear variation in X of u X (X ), u Y (X ) and θ(X ), e and γ only depend on
θ whereas κ is constant over the element. Making use of the relations (9.7) one can express e and
γ in the geometrically invariant form

L cos γ̄ L sin γ̄
1 + e = s cos(θ − ψ) = , γ = −s sin(θ − ψ) = (9.17)
L0 L0

In theory one could further reduce e to L/L 0 and γ to L γ̄ /L 0 , but these “simplifications” actually
complicate the strain variations taken in the following Section.
§9.3.5. *Consistent Linearization
The derivation of the consistent linearization (9.14) is based on the following study, known in continuum
mechanics as a polar decomposition analysis of the deformation gradient. Introduce the matrix
 
cos α − sin α
Ω(α) = (9.18)
sin α cos α
which represents a two-dimensional rotation (about Z ) through an angle α. Since Ω is an orthogonal matrix,
ΩT = Ω−1 . The deformation gradient (9.9) can be written
   
s 0 −Y θ 0
F = Ω(ψ) + Ω(θ) . (9.19)
0 0 0 1
where s is defined in (9.6). Premultiplying both sides of (9.19) by Ω(−θ ) gives the modified deformation
gradient  
s cos(θ − ψ) − Y θ 0
F̄ = Ω(−θ )F = (9.20)
−s sin(θ − ψ) 1
Now the GL strain tensor 2e = FT F − I does not change if F is premultiplied by an orthogonal matrix Ω
T
because FT ΩT ΩF = FT F. Consequently 2e = F̄ F̄ − I. But if the strains remain small, as it is assumed in
the Timoshenko model, the following are small quantities:

9–13
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–14

(i) s − 1 = (L/L 0 ) − 1 because the axial strains are small;

(ii) Y θ = Y κ because the curvature κ is of order 1/R, R being the radius of curvature, and |R| >> |Y |
according to beam theory because Y can vary only up to the cross section in-plane dimension;
(iii) γ̄ = ψ − θ, which is the mean angular shear deformation.
Then F̄ = I + L + higher order terms, where L is a first-order linearization in the small quantities s − 1, Y θ
and γ̄ = ψ − θ . It follows that
e = 12 (L + LT ) + higher order terms (9.21)
Carrying out this linearization one finds that e X Y and eY Y do not change, but that e X X simplifies to (9.13). It
.
can also be shown that 2e X Y = γ = γ̄ within the order of approximation of (9.21).

§9.4. Arbitrary Reference Configuration

In the general case the reference configuration C0 of the element is not aligned with X . The
longitudinal axis X̄ forms an angle ϕ with X , as illustrated in Figure 9.8. The six degrees of
freedom of the element are indicated in that Figure. Note that the section rotation angle θ is
measured from the direction Ȳ , normal to X̄ , and no longer from Y as in Figure 9.6.
Given the node coordinates (X 1 , Y1 ) and (X 2 , Y2 ), the reference angle ϕ is determined by
cos ϕ = X 21 /L 0 , sin ϕ = Y21 /L 0 , X 21 = X 2 − X 1 , Y21 = Y2 − Y1 , L 20 = X 21
2
+ Y21
2
. (9.22)
The angle φ = ψ +ϕ formed by the current longitudinal axis with X (see Figure 9.12) is determined
by
cos φ = cos(ψ + ϕ) = x21 /L , sin φ = sin(ψ + ϕ) = y21 /L , with
x21 = x2 − x1 = X 21 + u X 2 − u X 1 , y21 = y2 − y1 = Y21 + u Y 2 − u Y 1 , L 2 = x21
2
+ y21
2
.
(9.23)
Solving the trigonometric relations (9.22)-(9.23) for ψ gives
X 21 x21 + Y21 y21 X 21 (X 21 + u X 2 − u X 1 ) + Y21 (Y21 + u Y 2 − u Y 1 )
cos ψ = = ,
L L0 L L0
(9.24)
X 21 y21 − Y21 x21 X 21 (Y21 + u Y 2 − u Y 1 ) − Y21 (X 21 + u X 2 − u X 1 )
sin ψ = = .
L L0 L L0
It follows that L sin ψ and L cos ψ are exactly linear in the translational node displacements. This
property simplifies considerably the calculations that follow.

§9.4.1. Strain-Displacement Matrix

For the generalized strains it is convenient to use the invariant form (9.17), which does not depend
on ϕ. The variations δe, δγ and δκ with respect to nodal displacement variations are required in the
formation of the strain displacement relation δh = B δu. To form B we take partial derivatives of e,
γ and κ with respect to node displacements. Here is a sample of the kind of calculations involved:
∂e ∂[L cos(θ − ψ)/L 0 − 1] ∂[L(cos θ cos ψ + sin θ sin ψ)/L 0 − 1]
= =
∂u X 1 ∂u X 1 ∂u X 1
(9.25)
−X 21 cos θ + Y21 sin θ − cos ϕ cos θ + sin ϕ sin θ cos ω
= 2
= =−
L0 L0 L0

9–14
9–15 §9.4 ARBITRARY REFERENCE CONFIGURATION

_
// γY
φ = ψ+ϕ _
θ2 ψ
// γX
ϕ
2(x2 ,y2) // γX

_ C
// γY
uY 2
θ1

_
1(x1 ,y1) Xγ
ϕ
2(X2, Y2 )
uY 1
Y, y _ uX 2
Yγ
C0
X, x
uX
1(X1, Y1 )

Figure 9.8. Plane beam element with arbitrarily oriented reference configuration.

in which ω = θ + ϕ, and where use is made of (9.24) in a key step. These derivatives were checked
with Mathematica. Collecting all of them into matrix B:
 
1 − cos ω − sin ω L 0 N1 γ cos ω sin ω L 0 N2 γ
B= sin ω − cos ω −L 0 N1 (1 + e) − sin ω cos ω −L 0 N2 (1 + e) . (9.26)
L0 0 0 −1 0 0 1

Here N1 = (1 − ξ )/2 and N2 = (1 + ξ )/2 are abbreviations for the element shape functions
(caligraphic symbols are used to lessen the chance of clash against axial force symbols).

§9.4.2. Constitutive Equations

Because the beam material is assumed to be homogeneous and isotropic, the only nonzero PK2
stresses are the axial stress s X X and the shear stress s X Y . These are collected in a stress vector s
related to the GL strains by the linear elastic relations


0
0

sX X s1 s1 + Ee1 s1 E 0 e1
s= = = 0 = 0 + = s0 + Ee, (9.27)
sX Y s2 s2 + Ge2 s2 0 G e2

where E is the modulus of elasticity and G is the shear modulus. We introduce the prestress
resultants   
N =
0 0
s1 d A, V =
0 0
s2 d A, M =
0
−Y s10 d A, (9.28)
A0 A0 A0

9–15
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–16

N
M
V

M0
N0
0
V
C0

Figure 9.9. Beam stress resultants (internal forces)

in the reference and current configurations.

which define the axial forces, transverse shear forces and bending moments, respectively, in the
reference configuration. We also define the stress resultants

N = N 0 + E A0 e, V = V 0 + G A0 γ , M = M 0 + E I0 κ. (9.29)

These represent axial forces, tranverse shear forces and bending moments in the current configura-
tion, respectively, defined in terms of PK2 stresses. See Figure 9.9 for signs. These are collected
in the stress-resultant vector
zT = [ N V M ] . (9.30)

§9.4.3. Strain Energy

As in the case of the bar element, the total potential energy  = U − P is separable because
P = λqT u. The strain (internal) energy is given by
  
 0 T   0 
U= (s ) e + 2 e Ee d V =
1 T
(s1 e1 + s20 e2 ) + 12 (Ee12 + Ge22 ) d A d X̄ . (9.31)
V0 A0 L0

Carrying out the area integrals while making use of (9.27) through (9.30), U can be written as the
sum of three length integrals:
  
U= (N e + 2 E A0 e ) d X̄ +
0 1 2
(V γ + 2 G A0 γ ) d X̄ +
0 1 2
(M 0 κ + 12 E I0 κ 2 ) d X̄ ,
L0 L0 L0
(9.32)
The three terms in (9.32) define the energy stored through bar-like axial deformations, shear dis-
tortion and pure bending, respectively.

9–16
9–17 §9.6 THE STIFFNESS MATRIX

§9.5. The Internal Force

The internal force vector can be obtained by taking the first variation of the internal energy with
respect to the node displacements. This can be compactly expressed as
  
 
δU = N δe + V δγ + M δκ d X̄ = z δh d X̄ =
T
zT B d X̄ δu. (9.33)
L0 L0 L0

Here h and B are defined in equations (9.16) and (9.26) whereas z collects the stress resultants in
C as defined in (9.28) through (9.30). Because δU = p T δu, we get

p= BT z d X̄ . (9.34)
L0

This expression may be evaluated by a one point Gauss integration rule with the sample point at
ξ = 0 (beam midpoint). Let θm = (θ1 + θ2 )/2, ωm = θm + ϕ, cm = cos ωm , sm = sin ωm ,
em = L cos(θm − ψ)/L 0 − 1, γm = L sin(ψ − θm )/L 0 , and
 
−cm −sm − 12 L 0 γm cm sm − 12 L 0 γm
1 
Bm = B|ξ =0 = sm −cm 21 L 0 (1 + em ) sm −cm 12 L 0 (1 + em )  (9.35)
L0
0 0 −1 0 0 1
where subscript m stands for “beam midpoint.” Then
 T  
−cm −sm 1
L γ
2 0 m
cm sm 1
L γ
2 0 m N
p = L 0 BmT z =  sm −cm − 2 L 0 (1 + em ) −sm
1
cm − 12 L 0 (1 + em )  V (9.36)
0 0 −1 0 0 1 M

§9.6. The Stiffness Matrix

The first variation of the internal force vector (9.34) defines the tangent stiffness matrix

 T 
δp = B δz + δBT z d X̄ = (K M + KG ) δu = K δu. (9.37)
L0

This is again the sum of the material stifness K M and the geometric stiffness KG .

§9.6.1. The Material Stiffness Matrix

The material stiffness comes from the variation δz of the stress resultants while keeping B fixed.
This is easily obtained by noting that
    
δN E A0 0 0 δe
δz = δV = 0 G A0 0 δγ = S δh, (9.38)
δM 0 0 E I0 δκ
where S is the diagonal constitutive matrix with entries E A0 , G A0 and E I0 . Because δh = B δu,
the term BT δz becomes BT SB δu = K M δu whence the material matrix is

KM = BT SB d X̄ . (9.39)
L0

9–17
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–18

This integral is evaluated by the one-point Gauss rule at ξ = 0. Denoting again by Bm the matrix
(9.35), we find

KM = BmT SBm d X̄ = KaM + KbM + KsM (9.40)
L0

where KaM , KbM and KsM are due to axial (bar), bending, and shear stiffness, respectively:
 
cm2 cm sm −cm γm L 0 /2 −cm2 −cm sm −cm γm L 0 /2
 cm sm 2
sm −γm L 0 sm /2 −cm sm −sm2 −γm L 0 sm /2 
E A0  −c
 m γm L 0 /2 −γm L 0 sm /2 γm2 L 20 /4 cm γm L 0 /2 γm L 0 sm /2

γm2 L 20 /4 
KaM =  
L0  −cm2 −cm sm cm γm L 0 /2 cm2 cm sm cm γm L 0 /2 
 
−cm sm −sm2
γm L 0 sm /2 cm sm sm2 γm L 0 sm /2
−cm γm L 0 /2 −γm L 0 sm /2 γm2 L 20 /4 cm γm L 0 /2 γm L 0 sm /2 γm2 L 20 /4
  (9.41)
0 0 0 0 0 0
0 0 0 0 0 0 
 
E I 0 0 0 1 0 0 −1 
KbM =   (9.42)
L0  0 0 0 0 0 0 
 
0 0 0 0 0 0
0 0 −1 0 0 1
 
sm2 −cm sm −a1 L 0 sm /2 −sm2 c m sm −a1 L 0 sm /2
 −cm sm cm2
cm a1 L 0 /2 cm sm −cm2 cm a1 L 0 /2 
G A0 
 −a1 L 0 sm /2 cm a1 L 0 /2 a1 L 0 /4
2 2
a1 L 0 sm /2 −cm a1 L 0 /2 a12 L 20 /4 

KsM =  
L0  −sm2 cm sm a1 L 0 sm /2 sm2 −cm sm a1 L 0 sm /2 
 
cm sm −cm2 −cm a1 L 0 /2 −cm sm cm2 −cm a1 L 0 /2
−a1 L 0 sm /2 cm a1 L 0 /2 a1 L 0 /4
2 2
a1 L 0 sm /2 −cm a1 L 0 /2 a12 L 20 /4
(9.43)
in which a1 = 1 + em .

§9.6.2. Eliminating Shear Locking by RBF

How good is the nonlinear material stiffness (9.42)-(9.43)? If evaluated at the reference configura-
tion aligned with the X axis, cm = 1, sm = em = γm = 0, and we get
 E A0 
L0 0 0 − ELA0 0 0
0
 G A0 
 0
 L0
1
2
G A0 0 − GLA0 1
2

G A0

 0

 0 1
G A0 E I0 + 1 G A L 0 − E I
− 2 G A0
10 + 1
G A L 
 2 L0 4 0 0 L0 4 0 0 
KM =  EA 
− 0 0 0 E A 0 0 0 
 L0 L0 
 
 0 − LG A 0 − 2 G A0
1
0 G A 0 − 1
G A 
 0 L0 2 0 
0 1
2
G A0 − ELI0 + 14 G A0 L 0 0 − 12 G A0 E I0 + 1 G A L
L0 4 0 0
0
(9.44)
This is the well known linear stiffness of the C 0 beam. As noted in the discussion of Section
9.2.4, this element does not perform as well as the C 1 beam when the beam is thin because too

9–18
9–19 §9.6 THE STIFFNESS MATRIX

much strain energy is taken by shear. The following substitution device, introduced by MacNeal,7
removes that deficiency in a simple way. The shear rigidity G A0 is formally replaced by 12E I0 /L 20 ,
and magically (9.44) becomes
 EA 
L
0 0 0 − ELA0 0 0
 0 0 
 0 12E I0 6E I0 − 12E3 I0 6E I0 
 L 30 L 20
0
L 20 
 L0 
 6E I0 4E I0 
 0
 L0 0 − 2I0
6E 2E I0
L0


L 20 L0
K̂ M =
 E A0 E A0
.
 (9.45)
− L 0 0 0 0 
 0 L0 
 
 0 − 12E3 I0 − 6E2I0 0 12E I0 6E
− 2 I 0
 L0 L0 L 30 L0 
 
0 6E I0 2E I0 0 − 6E2I0 4E I0
L 20 L0 L0 L0

This is the well known linear stiffness matrix of the C 1 (Hermitian) beam based on the Euler-
Bernoulli model. That substitution device is called the residual bending flexibility (RBF) correc-
tion.8 Its effect is to get rid of the spurious shear energy due to the linear kinematic assumptions. If
the RBF is formally applied to the nonlinear material stiffness one gets K̂ M = KaM + K̂ Mb , where
KaM is the same as in (9.43) (because the axial stiffness if not affected by the substitution), whereas
KbM and KsM merge into
 
12sm2 −12cm sm 6a1 L 0 sm −12sm2 12cm sm 6a1 L 0 sm
 −12cm sm 12cm2 −6cm a1 L 0 12cm sm −12cm2 −6cm a1 L 0 
b EI  6a L s −6cm a1 L 0 a2 L 20 −6a1 L 0 sm 6cm a1 L 0 a3 L 20 

K̂ M = 3  1 0 2m 
L 0  −12sm 12cm sm −6a1 L 0 sm 12sm2 −12cm sm −6a1 L 0 sm 
 
12cm sm −12cm2 6cm a1 L 0 −12cm sm 12cm2 6cm a1 L 0
6a1 L 0 sm −6cm a1 L 0 a3 L 20 −6a1 L 0 sm 6cm a1 L 0 a2 L 20
(9.46)
in which a1 = 1 + em , a2 = 4 + 6em + 3em2 and a3 = 2 + 6em + 3em2 .

Remark 9.4. MacNeal actually proposed the more refined substitution

1 1 L 20
replace by + (9.47)
G A0 G As 12E I0

where G As is the actual shear rigidity; that is, As is the shear-reduced cross section studied in Mechanics
of Materials. The result of (9.47) is the C 1 Hermitian beam corrected by shear deformations computed
from equilibrium considerations.9 If the shear deformation is negligible, the right hand side of (9.47) is
approximately L 20 /(12E I0 ), which leads to the substitution used above.

7 R. H. MacNeal, A simple quadrilateral shell element, Computers and Structures, 8, 1978, pp. 175-183.
8 RBF can be rigurously justified through the use of a mixed variational principle, or through a flexibility calculation.
9 See, e.g., J. Przemieniecki, Theory of Matrix Structural Analysis, Dover, New York, 1968.

9–19
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–20

§9.6.3. The Geometric Stiffness Matrix

The geometric stiffness KG comes from the variation of B while the stress resultants in z are kept
fixed. To get a closed form expression it is convenient to pass to indicial notation, reverting to
matrix notation later upon “index contraction.” Let the entries of KG , B, u and z be denoted as
K Gi j , Bki , u j and z k , where indices i, j and k range over 1–6, 1–6, and 1–3, respectively. Call
A j = ∂B/∂u j , j = 1, . . . 6. Then using the summation convention,
  
∂ Bki j
K Gi j δu j = δB z d X =
T
δu j z k d X = Aki z k d X δu j , (9.48)
L0 L0 ∂u j L0

whence 
j
K Gi j = z k Aki d X̄ , (9.49)
L0

Note that in carrying out the derivatives in (9.49) by hand one must use the chain rule because B is
a function of e, γ and θ , which in turn are functions of the node displacements u j . To implement
this scheme we differentiate B with respect to each node displacement in turn, to obtain:

 
∂B 1 0 0 N1 sin ω 0 0 N2 sin ω
A1 = = 0 0 N1 cos ω 0 0 N2 cos ω ,
∂u X 1 L0 0 0 0 0 0 0
 
∂B 1 0 0 −N1 cos ω 0 0 −N2 cos ω
A2 = = 0 0 N1 sin ω 0 0 N2 sin ω ,
∂u Y 1 L0 0 0 0 0 0 0
 
∂B N1 sin ω − cos ω −N1 L 0 (1 + e) − sin ω cos ω −N2 L 0 (1 + e)
A3 = = cos ω sin ω −N1 L 0 γ − cos ω − sin ω −N2 L 0 γ ,
∂θ1 L0 0 0 0 0 0 0
  (9.50)
∂B 1 0 0 −N1 sin ω 0 0 −N2 sin ω
A4 = = 0 0 −N1 cos ω 0 0 −N2 cos ω ,
∂u X 2 L0 0 0 0 0 0 0
 
∂B 1 0 0 N1 cos ω 0 0 N2 cos ω
A5 = = 0 0 −N1 sin ω 0 0 −N2 sin ω ,
∂u Y 2 L0 0 0 0 0 0 0
 
∂B N2 sin ω − cos ω −N1 L 0 (1 + e) − sin ω cos ω −N2 L 0 (1 + e)
A6 = = cos ω sin ω −N1 L 0 γ − cos ω − sin ω −N2 L 0 γ .
∂θ2 L0 0 0 0 0 0 0

To restore matrix notation it is convenient to define

j j j
W N i j = A1i , WV i j = A2i , W Mi j = A3i , (9.51)

as the entries of three 6 × 6 “weighting matrices” W N , WV and W M that isolate the effect of the
stress resultants z 1 = N , z 2 = V and z 3 = M. The first, second and third row of each A j becomes

9–20
9–21 §9.7 A COMMENTARY ON THE ELEMENT PERFORMANCE

the j th column of W N , WV and W M , respectively. The end result is

 
0 0 N1 sin ω 0 0 N2 sin ω
 0 0 −N1 cos ω 0 0 −N2 cos ω 
1  N1 sin ω −N1 cos ω −N1 L 0 (1 + e)
2
−N1 sin ω N1 cos ω −N1 N2 L 0 (1 + e) 

WN =  
L0  0 0 −N1 sin ω 0 0 −N2 sin ω 
 
0 0 N1 cos ω 0 0 N2 cos ω
N2 sin ω −N2 cos ω −N1 N2 L 0 (1 + e) −N2 sin ω N2 cos ω −N22 L 0 (1 + e)
 (9.52)

0 0 N1 cos ω 0 0 N2 cos ω
 0 0 N1 sin ω 0 0 N2 sin ω 
1  N
 1 cos ω N sin ω −N 2
L γ −N cos ω −N sin ω −N

1 N2 L 0 γ 
WV =  1 1 0 1 1

L0  0 0 −N1 cos ω 0 0 −N2 cos ω 
 
0 0 −N1 sin ω 0 0 −N2 sin ω
N2 cos ω N2 sin ω −N1 N2 L 0 γ −N2 cos ω −N2 sin ω −N22 L 0 γ
(9.53)
and W M = 0. Notice that the matrices must be symmetric, since KG derives from a potential. Then

KG = (W N N + WV V ) d X̄ = KG N + KGV . (9.54)
L0

Again the length integral should be done with the one-point Gauss rule at ξ = 0. Denoting again
quantities evaluated at ξ = 0 by an m subscript, one obtains the closed form
 
0 0 sm 0 0 sm
 0 0 −cm 0 0 −cm 
Nm  sm −cm − 2 L 0 (1 + em ) −sm
1 
cm − 2 L 0 (1 + em ) 
1
KG =  
2  0 0 −sm 0 0 −sm 
 
0 0 cm 0 0 cm
sm −cm − 2 L 0 (1 + em ) −sm
1
cm − 2 L 0 (1 + em )
1

  (9.55)
0 0 cm 0 0 cm
 0 0 sm 0 0 sm 
Vm  cm sm − 2 L 0 γm −cm −sm
1
− 2 L 0 γm 
1 
+  .
2  0 0 −cm 0 0 −cm 
 
0 0 −sm 0 0 −sm
cm sm − 12 L 0 γm −cm −sm − 12 L 0 γm
in which Nm and Vm are N and V evaluated at the midpoint.

§9.7. A Commentary on the Element Performance

The material stiffness of the present element works fairly well once MacNeal’s RBF device is done.
On the other hand, simple buckling test problems, as in Exercise 9.3, show that the geometric
stiffness is not so good as that of the C 1 Hermitian beam element.10 Unfortunately a simple

10 In the sense that one must use more elements to get equivalent accuracy.

9–21
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–22

Node Displacements u

Eqs (9.5), (9.8)

Element displacement field w = [uX , uY , θ]T

Eq. (9.9)

Displacement gradients w' = [u'X , u'Y , θ' ]

T

Eq. (9.15)

Generalized strains h = [ e , γ, κ ]
T

Eq. (9.29) Strain energy U

Stress resultants z = [ N ,V, M ]T

∫ _
vary U: δU = L zT BT dX δu = pT δu
0
Eq. (9.34)

Internal forces p

vary p: δp = ∫ L0
_
(B T δz + δBT z) dX δu = (KM + KG )

Eq. (9.40)

Tangent stiffness matrix K = KM + KG

Figure 9.10. Main steps in the derivation of the C 0 plane beam element.

substitution device such as RBF cannot be used to improve KG , and the problem should be viewed
as open.
An intrinsic limitation of the present element is the restriction to small axial strains. This was
done to facilitate close form derivation. The restriction is adequate for many structural problems,
particularly in Aerospace (example: deployment). However, it means that the element cannot model
correctly problems like the snap-through and bifurcation of the arch example used in Chapter 8, in
which large axial strains prior to collapse necessarily occur.

§9.8. Summary

Figure 9.10 is a roadmap that summarizes the key steps in the derivation of the internal force and
tangent stiffness matrix for the C 0 plane beam element.

9–22
9–23 Exercises

Homework Exercises for Chapter 9

The Plane Beam Element

EXERCISE 9.1 [A+N:20] Consider a plane prismatic beam of length L 0 , cross section area A0 and second
moment of inertia I0 . In the reference configuration the beam extends from node 1 at (X = Y = 0) to node
2 at (X = L 0 , Y = 0). The beam is under axial prestress force N 0 = 0 in the reference configuration, while
both V 0 and M0 vanish.
(a) Obtain the internal force p, the material stiffness matrix K M and the geometric stiffness matrix KG in
the reference configuration, for which u X 1 = u Y 1 = u X 2 = u Y 2 = θ1 = θ2 = 0.
(b) The beam rotates 90◦ rigidly to a current configuration for which u X 1 = u Y 1 = 0, u X 2 = −L 0 , u Y 2 = L 0 ,
θ1 = θ2 = 90◦ . Check that the stress resultants do not change (that is, N = N0 and V = M = 0 because
e = γ = κ = 0), and obtain K M and KG in that configuration.

EXERCISE 9.2

;; ;
[N:20] Analyze the pure bending of a cantilever discretized into an arbitrary number of elements.

;
reference configuration C0 Buckling sketch
P
L
Pcr = -λcr EI/L2

fixed end P
FEM discretization into
Ne equal length elements

Figure E9.1. Structure for Exercises 9.3 and 9.4.

EXERCISE 9.3
[N:25] This exercise and the next one pertain to the simple structure shown in Figure E9.1. It is a cantilever,
plane beam-column of length L, modulus E, area A and inertia I , loaded by an axial force P as shown. The
structure is discretized into Ne plane beam elements of equal length. The structures moves in the plane of the
figure. The objective is to compute the classical buckling load Pcr and compare with the known analytical
value.
In the classical buckling analysis, deformations prior to buckling are neglected. The structure stiffness K =
K M + KG is evaluated on the reference configuration, with KG evaluated from the internal force state. This
stability model is called linearized prebuckling or LPB, and is studied in detail in later Chapters. Under the
LPB assumptions K M is constant while KG linear in the applied forces. Critical point analysis leads to a linear
eigensystem called the buckling eigenproblem. The smallest eigenvalue characterizes the critical load, which
for the LPB model can be shown to produce bifurcation.
The classical buckling load (also called Euler load) for the configuration of Figure E9.1 is Pcr = λcr E I /L 2
with λcr = −π 2 /4 (negative because P has to compress the beam-column to achieve buckling). FEM results
for Ne = 1, 2, . . . can be obtained by running the code of Figure A9.7. This script assembles K taking in 0 with
E = I = A = L = 1 to simplify computations while leaving P = λ symbolic. The only nonzero internal
force is N0 = P. The determinant det(K) is formed explicitly as a polynomial in P, called the characteristic
polynomial. All of its roots are computed via NSolve. The root closest to zero defines the critical load.

9–23
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–24

As answers to the exercise run the script until the largest Ne that can be reasonably handled by your Mathematica
version (this will depend on memory and CPU speed). Termination can be controlled by adjusting the loop
on k, which doubles Ne on each pass.
Report on the following items:
(a) Your computed λcr ’s for the Ne you were able to run. Comment on whether they converge toward −π 2 /4.
Is the convergence monotonic?
(b) Why does NSolve give 2Ne roots?11
(c) Why are the roots real?12
(d) Why do the computations get rapidly very slow as Ne increases?13

EXERCISE 9.4 [N:20] The brute-force technique used to find λcr in Exercise 9.3 is easy to implement but it
is extremely inefficient as the number of elements increase.14
A more effective tecnique is implemented in the script shown in Figure A9.8 of the Addendum. On Mathematica
this works fine on my Mac up to 128 elements, beyond which point memory and CPU time requirements grow
too large. As answer to this exercise:
(a) Describe what is going on in the script.
(b) Discuss why the solution is more efficient and robust than the previous script.15

11 Only one of which is of practical interest.

12 Assuming that exact integer arithmetic is used to form the determinant. If you form the determinant in floating-point
complex roots will likely emerge because of numerical imprecision.
13 If you can’t guess, try printing the expanded determinant.
14 Forming the characteristic polynomial by expanding det(K) is frowned upon by numerical analysts for several reasons,
one of which is that the polynomial coefficients tend to get enormously large. This requires either extended precision if
done in integer arithmetic (as in the script) or rapidly lead to overflow if done in floating point.
15 It is still far, however, from the optimal way to compute λcr . The practical technique used in production FEM codes is a
variation on that script, using an eigensolution method called inverse power iteration.

9–24
9–25 Exercises

Addendum for Homework Exercises 9.3 and 9.4

The Mathematica Notebook PlaneBeam.nb contains several modules and scripts that support Exercises 9.3
and 9.4. The Notebook file is posted on the course Web site. The important pieces of code are briefly described
in this Addendum.
Figure A9.1 shows module FormIntForceC0TwoNodePlaneBeam. This forms the internal force vector p of
a two-node, C 0 , plane beam formulated with the TL description. Although this module is not directly called
in Exercises 9.3 and 9.4, it is used later in the course. It is also used in the verifivation of the stiffness matrix
modules by finite differences.

FormIntForceC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=Module[
{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,x21,y21,
thetam,ctheta,stheta,Lcpsi,Lspsi,N0,V0,M0,EA0,GA0,EI0,
cphi,sphi,cm,sm,Nm,Vm,Mm,kappa,p},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
kappa=(theta2-theta1)/L0;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0;
Nm=Simplify[N0+EA0*em]; Vm=Simplify[V0+GA0*gm];
Mm=Simplify[M0+EI0*kappa];
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
Bm= (1/L0)*{{-cm,-sm, L0*gm/2, cm,sm, L0*gm/2 },
{ sm,-cm,-L0*(1+em)/2,-sm,cm,-L0*(1+em)/2},
{ 0, 0, -1, 0, 0, 1 }};
p=L0*Transpose[Bm].{{Nm},{Vm},{Mm}};
Return[Simplify[p]] ];

p=FormIntForceC0TwoNodePlaneBeam[{0,0},{L/Sqrt[2],L/Sqrt[2]},
{0,0,Pi/2},{-2*L/Sqrt[2],0,Pi/2},
{EA,GA0,EI},{0,V0,0}];
Print["p=",p];

Figure A9.1. Evaluation of internal force vector p for plane beam element.

The 6 arguments of this module are XY1, XY2, u1, u2, S0 and z0. XY1 lists the coordinates {X 1 , Y2 } of node
1 whereas XY2 lists the coordinates {X 1 , Y2 } of node 2, in the reference configuration. u1 passes the three
displacements: {u X 1 , u Y 1 , θ1 } of node 1, and u2 does the same for node 2. S0 collects the section integrated
constitutive properties {E A0 , G A0 , E I0 }. To apply MacNeal’s RBF, G A0 should be replaced by 12E I0 /L 20
as discussed in Section 9.5. Finally z0 passes the internal forces {N0 , V0 , M0 } in the reference configuration;

9–25
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–26

FormMatStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,x21,y21,
thetam,ctheta,stheta,Lcpsi,Lspsi,N0,V0,M0,EA0,GA0,EI0,
cphi,sphi,cm,sm,a1,KM},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0; a1=1+em;
KM = (EA0/L0)*{
{ cm^2,cm*sm,-cm*gm*L0/2,-cm^2,-cm*sm,-cm*gm*L0/2},
{ cm*sm,sm^2,-gm*L0*sm/2,-cm*sm,-sm^2,-gm*L0*sm/2},
{-cm*gm*L0/2,-gm*L0*sm/2,gm^2*L0^2/4,
cm*gm*L0/2,gm*L0*sm/2,gm^2*L0^2/4},
{-cm^2,-cm*sm,cm*gm*L0/2,cm^2,cm*sm,cm*gm*L0/2},
{-cm*sm,-sm^2,gm*L0*sm/2,cm*sm,sm^2,gm*L0*sm/2},
{-cm*gm*L0/2,-gm*L0*sm/2,gm^2*L0^2/4,
cm*gm*L0/2,gm*L0*sm/2,gm^2*L0^2/4}}+
(EI0/L0)*{{0,0,0,0,0,0}, {0,0,0,0,0,0},
{0,0,1,0,0,-1}, {0,0,0,0,0,0},
{0,0,0,0,0,0}, {0,0,-1,0,0,1}}+
(GA0/L0)*{
{sm^2,-cm*sm,-a1*L0*sm/2,-sm^2,cm*sm,-a1*L0*sm/2},
{-cm*sm,cm^2,cm*a1*L0/2,cm*sm,-cm^2,cm*a1*L0/2},
{-a1*L0*sm/2,cm*a1*L0/2, a1^2*L0^2/4,
a1*L0*sm/2,-cm*a1*L0/2, a1^2 L0^2/4},
{-sm^2,cm*sm, a1*L0*sm/2,sm^2,-cm*sm,a1*L0*sm/2},
{ cm*sm,-cm^2,-cm*a1*L0/2,-cm*sm,cm^2,-cm*a1*L0/2},
{-a1*L0*sm/2,cm*a1*L0/2,a1^2*L0^2/4,
a1*L0*sm/2,-cm*a1*L0/2, a1^2*L0^2/4}};
Return[KM] ];

Figure A9.2. Evaluation of material stiffness matrix K M for plane beam element.

these forces are assumed to be constant along the element. Thmodule forms the internal force vector p with a
one point integration rule as discussed earlier, and returns p as function value.
Figures A9.2 and A9.3 show modules FormMatStiffC0TwoNodePlaneBeam and
FormGeoStiffC0TwoNodePlaneBeam. As their name suggest, these form the material and geometric stiffness
components, respectively, of the plane beam element. They have been separated for convenience although for
most applications they are combined to form the tangent stiffness matrix. The 6 arguments are exactly the

9–26
9–27 Exercises

FormGeoStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,
x21,y21,thetam,ctheta,stheta,Lcpsi,Lspsi,
N0,V0,M0,EA0,GA0,EI0,cphi,sphi,cm,sm,Nm,Vm,KG},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
kappa=(theta2-theta1)/L0;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0;
Nm=Simplify[N0+EA0*em]; Vm=Simplify[V0+GA0*gm];
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
KG = Nm/2*{{ 0, 0, sm, 0, 0, sm}, {0, 0, -cm, 0, 0,-cm},
{sm, -cm, -(L0/2)*(1+em), -sm, cm,-(L0/2)*(1+em)},
{0, 0, -sm, 0, 0, -sm}, {0, 0, cm, 0, 0, cm},
{sm, -cm, -(L0/2)*(1+em), -sm, cm,-(L0/2)*(1+em)}}+
Vm/2*{{0, 0, cm, 0, 0, cm}, {0, 0, sm, 0, 0, sm},
{cm, sm, -(L0/2)*gm, -cm, -sm,-(L0/2)*gm},
{0, 0, -cm, 0, 0, -cm}, {0, 0, -sm, 0, 0, -sm},
{cm, sm, -(L0/2)*gm, -cm, -sm,-(L0/2)*gm}};
Return[KG] ];

Figure A9.3. Evaluation of geometric stiffness matrix KG for plane beam element.

FormTanStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{},
KM=FormMatStiffC0TwoNodePlaneBeam[XY1,XY2,u1,u2,S0,z0];
KG=FormGeoStiffC0TwoNodePlaneBeam[XY1,XY2,u1,u2,S0,z0];
Return[KM+KG]];

ClearAll[EA,GA,EI];
KM=FormMatStiffC0TwoNodePlaneBeam[{0,0},{10,0},{0,0,0},{0,0,0},
{EA,GA,EI},{0,0,0}]; Print[KM];
KG=FormGeoStiffC0TwoNodePlaneBeam[{0,0},{4,3},{0,0,0},{0,0,0},
{1,1,1},{10,30,20}]; Print[KG];
Print[Chop[Eigenvalues[N[KG]]]];

Figure A9.4. Evaluation of tangent stiffness matrix K for plane beam element,
along with test statements.

9–27
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–28

MergeElemIntoMasterIntForce[pe_,eftab_,pm_]:=
Module[{i,ii,nf=Length[eftab],p}, p=pm;
For[i=1, i<=nf, i++, ii=eftab[[i]];
If [ii>0,p[[ii,1]]+=pe[[i,1]]]
]; Return[p]
];

MergeElemIntoMasterStiff[Ke_,eftab_,Km_]:=
Module[{i,j,ii,jj,nf=Length[eftab],K}, K=Km;
For[i=1, i<=nf, i++, ii=eftab[[i]]; If[ii==0,Continue[]];
For[j=i, j<=nf, j++, jj=eftab[[j]];
If [ii>0 && jj>0,
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]]
]
]; Return[K]
];

Figure A9.5. Merge modules for master internal force

and master stiffness of a plane beam structure.

same as for the internal force module, Both matrices are formed with the one-point integration rule.and are
return as function values. The correctness of the implementation is checked with a finite difference technique
implemented in a script not listed here.

Figure A9.4 lists module FormTanStiffC0TwoNodePlaneBeam, which returns the tangent stiffness matrix
K = K M + KG . It simple calls the previous two modules and returns the matrix sum as function value.

Figure A9.5 lists two modules that merge the internal force vector and stiffness matrix, respectively, of one
individual element into the corresponding master quantities for the entire structure.

Figure A9.6 lists two modules: AssembleMasterStiffOfCantBeam and AssembleMasterIntForceOfCantBeam

that assemble p and K, respectively, for one specific structure. This structure is the cantilever beam-column
under axial force P shown in Figure E9.1, discretized into Ne ≥ 1 plane beam elements.

Figures A9.7 and A9.8 lists two scripts for use in Exercises 9.3 and 9.4, respectively.

9–28
9–29 Exercises

AssembleMasterIntForceOfCantBeam[{L_,P_,Ne_},{Em_,A0_,I0_},u_]:=
Module[{e,numnod,numdof,u1,u2,u3,Le,X1,X2,S0,z0,eftab,
pe,p},
numnod=Ne+1; numdof=3*numnod; Le=L/Ne;
p=Table[0,{numdof},{1}];
z0={P,0,0}; S0={Em*A0,12*Em*I0/Le^2,Em*I0}; (* RBF *)
For [e=1, e<=Ne, e++,
X1=(e-1)*Le; X2=e*Le;
If [e==1,u1={0,0,0}; u2=Take[u,{3*e-2,3*e}];
eftab={0,0,0,3*e-2,3*e-1,3*e}];
If [e>1, u1=Take[u,{3*e-5,3*e-3}]; u2=Take[u,{3*e-2,3*e}];
eftab={3*e-5,3*e-4,3*e-3,3*e-2,3*e-1,3*e}];
pe=FormIntForceC0TwoNodePlaneBeam[{X1,0},{X2,0},u1,u2,S0,z0];
(*Print["pe=",pe];*)
p=MergeElemIntoMasterIntForce[pe,eftab,p];
];
Return[Simplify[p]]
];

AssembleMasterStiffOfCantBeam[{L_,P_,Ne_},{Em_,A0_,I0_},u_]:=
Module[{e,numnod,numdof,u1,u2,u3,Le,X1,X2,S0,z0,eftab,
pe,p},
numnod=Ne+1; numdof=3*numnod-3; Le=L/Ne;
K=Table[0,{numdof},{numdof}];
z0={P,0,0}; S0={Em*A0,12*Em*I0/Le^2,Em*I0}; (* RBF *)
For [e=1, e<=Ne, e++,
X1=(e-1)*Le; X2=e*Le;
If [e==1,u1={0,0,0}; u2=Take[u,{3*e-2,3*e}];
eftab={0,0,0,3*e-2,3*e-1,3*e}];
If [e>1, u1=Take[u,{3*e-5,3*e-3}]; u2=Take[u,{3*e-2,3*e}];
eftab={3*e-5,3*e-4,3*e-3,3*e-2,3*e-1,3*e}];
Ke=FormTanStiffC0TwoNodePlaneBeam[{X1,0},{X2,0},u1,u2,S0,z0];
(*Print["Ke=",Ke];*)
K=MergeElemIntoMasterStiff[Ke,eftab,K];
];
Return[Simplify[K]]
];

Figure A9.6. Assembly modules for cantilevered beam-column under axial load.

9–29
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–30

ClearAll[L,P,Em,A0,I0]; Em=A0=I0=L=1; Ne=1;

For [k=1, k<=6, k++,
Print["Number of elements=",Ne];
LPNe={L,P,Ne}; EAI={Em,A0,I0}; u=Table[0,{3*Ne+3}];
K=AssembleMasterStiffOfCantBeam[LPNe,EAI,u];
(*Print["K=",K//MatrixForm];*)
detK=Det[K];
(*Print["det(K)=",detK//InputForm];*)
roots=NSolve[detK==0,P];
Print["roots of stability det=",roots]; Ne=2*Ne;
];
Print["exact buckling load coeff is ",-N[Pi^2/4]];

Figure A9.7. Script for Exercise 9.3. It uses assembly,

merge and element formation modules.

ClearAll[L,P,Em,A0,I0]; Em=A0=I0=L=1; Ne=1;

For [k=1, k<=8, k++,
Print["Number of elements=",Ne];
LPNe={L,P,Ne}; EAI={Em,A0,I0}; u=Table[0,{3*Ne+3}];
K=AssembleMasterStiffOfCantBeam[LPNe,EAI,u];
KM=Coefficient[K,P,0]; KG=Coefficient[K,P,1];
SG=LinearSolve[N[KM],N[KG]];
(*Print["KM=",KM//MatrixForm]; Print["KG=",KG//MatrixForm];*)
(*Print["SG=",SG//MatrixForm];*)
emax=-Max[Eigenvalues[SG]]; Print["FEM lambda cr=",1/emax];
Ne=2*Ne;
];
Print["exact buckling lambda coeff is ",-N[Pi^2/4]//InputForm];

Figure A9.8. Script for Exercise 9.4. It uses assembly,

merge and element formation modules.

9–30
 8
.

The TL
Bar Element

8–1
Chapter 8: THE TL BAR ELEMENT 8–2

TABLE OF CONTENTS

Page
§8.1. Introduction 8–3
§8.2. The Two-Dimensional Bar Element 8–3
§8.2.1. Element Kinematics . . . . . . . . . . . . . . . 8–3
§8.2.2. Strain Measure . . . . . . . . . . . . . . . . 8–4
§8.2.3. Stress Measure . . . . . . . . . . . . . . . . . 8–7
§8.2.4. Total Potential Energy and Residual Force Equations . . . . 8–8
§8.2.5. The Tangent Stiffness Matrix . . . . . . . . . . . . 8–8
§8.3. FEM Coding Using Mathematica 8–10
§8.3.1. The Example Structure . . . . . . . . . . . . . . 8–10
§8.3.2. Forming the Internal Force . . . . . . . . . . . . . 8–11
§8.3.3. The Equilibrium Equations . . . . . . . . . . . . 8–14
§8.3.4. Plotting the Equilibrium Paths . . . . . . . . . . . . 8–15
§8.3.5. Having Some Fun: Snapshots and Animation . . . . . . 8–15
§8.3.6. Forming the Tangent Stiffness Matrix . . . . . . . . . 8–18
§8.3.7. Critical Point Study . . . . . . . . . . . . . . . 8–18
§8. Exercises . . . . . . . . . . . . . . . . . . . . . . 8–25

8–2
8–3 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT

§8.1. Introduction
In the present Chapter the key concepts of nonlinear continuum mechanics reviewed in Chapter 7
are applied to the development of governing equations of bar (truss) elements based on the Total
Lagrangian (TL) kinematic description. For brevity these will be referred to as “TL bar elements.”
There are two ways to construct TL elements:
1. The Standard Formulation (SF)
2. The Core Congruential Formulation (CCF).
The first method is easier to describe and will be presented in this Chapter through examples. The
second one is more flexible and powerful but it is more difficult to teach because it proceeds in
stages. As such it will be relegated to Chapters 10-11.
§8.2. The Two-Dimensional Bar Element
The element developed in this Chapter is a prismatic bar element that can be used to model pin-
jointed plane truss structures of the type sketched in Figure 8.1. These structures undergo large
displacements and rotations but their strains are assumed to remain small so that the material
behavior stays in the linear elastic range. These assumptions allows us to consider only geometric
nonlinear effects.
A two-node bar element appropriate to model members of such truss structures is shown in Figure
8.2. The element moves in the (X, Y ) plane. In the reference (base) configuration the element has
cross section area A0 (constant along the element) and length L 0 . In the current configuration the
cross section area and length become A and L, respectively. The material has an elastic modulus
E that links the axial-stress and axial-strain measures defined below.
Because this Chapter deals primarily with the formulation of an individual element, the identification
superscript (e) will be omitted to reduce clutter until assemblies are considered.
The element has four node displacements and associated node forces. These quantities are collected
in the vectors    
u X1 fX1
u   f 
u =  Y1  , f =  Y1  , (8.1)
u X2 f X2
uY 2 fY 2
The loads acting on the nodes will be assumed to be conservative.
§8.2.1. Element Kinematics
In accordance with bar theory, to describe the element motion it is sufficient to consider a generic
point of coordinates X located on the longitudinal axis of the reference configuration C0 . That point
maps to point x in the current configuration C. The bar remains straight in any configuration. These
coordinates can be parametrically interpolated from the end nodes as
 1
X (ξ ) 2
(1 − ξ )X 1 + 12 (1 + ξ )X 2
X(ξ ) = = 1
Y (ξ ) (1 − ξ )Y1 + 12 (1 + ξ )Y2
2
 1 (8.2)
x(ξ ) (1 − ξ )x 1 + 1
(1 + ξ )x 2
x(ξ ) = = 21 2
y(ξ ) (1 − ξ )y1 + 2 (1 + ξ )y2
1
2

8–3
Chapter 8: THE TL BAR ELEMENT 8–4

Reference configuration
(same as base)

Motion

Current configuration

Figure 8.1. A plane truss structure undergoing large displacements

while its material stays in the linear elastic range.

Here ξ is the usual isoparametric coordinate that varies from −1 at node 1 to +1 at node 2. The
displacement field is obtained by subtracting the foregoing position vectors:
1

u X (ξ ) 2
(1 − ξ )u X 1 + 12 (1 + ξ )u X 2
u(ξ ) = x(ξ ) − X(ξ ) = = 1 , (8.3)
u Y (ξ ) (1 − ξ )u Y 1 + 12 (1 + ξ )u Y 2
2

which expressed in matrix form is

 
 1 u X1
u X (ξ ) (1 − ξ ) 0 1
(1 + ξ) 0  uY 1 
u(ξ ) = = 2 2
  = N(ξ ) u.
u Y (ξ ) 0 1
2
(1 − ξ) 0 1
2
(1 + ξ ) u X 2
uY 2
(8.4)
The element kinematic defined by these equations is depicted in Figure 8.3.

§8.2.2. Strain Measure

As discussed in Chapter 7, in the Total Lagrangian (TL) description the Green-Lagrange (GL)
strains and the second Piola-Kirchhoff (PK2) stresses are frequently used as conjugate measures in
the formulation of the internal energy. The only GL strain that appears in the energy expression is

8–4
8–5 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT

Reference configuration C0 2(X 2 , Y2 )

PK2 stress s0 , GL strain e0 = 0

Area A0 , length L 0

Y, y 1(X 1 , Y1 )
uY 2

uY 1
fY 1
X, x Area A, length L
uX1
f X1 fY 2
1(x1 , y1 )

uX2
f X2
2(x2 , y2 )
Current configuration C
PK2 stress s, GL strain e

Figure 8.2. The geometrically nonlinear, two-node, two-dimensional bar element

in Total Lagrangian description. This element may be used to
model members of a plane truss such as that shown in Figure 8.1.

the axial strain e1 ≡ e which is most expediciously defined using the length change as

L 2 − L 20
e= , (8.5)
2L 20

rather than through displacement gradients. Because of the linear displacement assumptions (8.4)
the strain e is constant over the element.1
This expression can be maneuvered into a matrix function of the node displacements as follows. Let
X 21 = X 2 − X 1 , Y21 = Y2 − Y1 , a X = X 21 /L 0 , aY = Y21 /L 0 , u X 21 = u X 2 − u X 1 , u Y 21 = u Y 2 − u Y 1 ,
u mX = (u X 2 − u X 1 )/2 and u m
Y = (u Y 2 − u Y 1 )/2. Some of these quantities can be geometrically
interpreted as illustrated in Figure 8.4. Then

1 This is in fact the only use of the displacement interpolation (8.4) in the following derivations.

8–5
Chapter 8: THE TL BAR ELEMENT 8–6

_
X

Bar longitudinal axis

2(X 2 , Y2 )
C0


1 
X (ξ ) (1 − ξ )X 1 + 12 (1 + ξ )X 2
X(ξ ) = = 2

1(X 1 , Y1 ) Y (ξ ) 1
2
(1 − ξ )Y1 + 21 (1 + ξ )Y2
Y, y


1 
u X (ξ ) (1 − ξ )v X 1 + 12 (1 + ξ )v X 2
u(ξ ) = x(ξ )−X(ξ ) = = 2
u Y (ξ ) 1
2
(1 − ξ )vY 1 + 12 (1 + ξ )vY 2

X, x


1 
x(ξ ) (1 − ξ )x1 + 12 (1 + ξ )x2
x(ξ ) = = 2
y(ξ ) 1
2
(1 − ξ )y1 + 12 (1 + ξ )y2
1(x1 , y1 )
C

2(x2 , y2 )

Figure 8.3. The definition of displacement field for the

two-dimensional TL bar element.

L 2 = (X 21 + u X 21 )2 + (Y21 + u Y 21 )2 ,
L 2 − L 20 1 1  2
e= = (a X u X 21 + a Y u Y 21 ) + u X 21 + u 2
Y 21
2L 20 L0 2L 20
(8.6)
1 1
= [ −a X −aY a X aY ] u + 2 [ −u mX −u m Y u mX u mY ]u
L0 L0

= Bl + Bn (u) u.

Observe that the GL strain e has been separated into two parts: e = el + en , where
1. el = Bl u, where Bl is constant, depends linearly in the node displacements u. This is called
the linear part of the strain.
2. en = Bn u, in which Bn is a function of the node displacements, depends quadratically on
the displacements. This is called the nonlinear part of the strain, because Bn (u) vanishes if
u → 0. Note that both matrices are constant over the element.

8–6
8–7 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT

_
X

2(X 2 , Y2 ) ψ0
X 21

C0
Y21 X 21
aX = = cos ψ0
L0
L0 Y21
aY = = sin ψ0
L0
Y, y 1(X 1 , Y1 )

x21 L
ax = = cos ψ
L0 L0
y21 L
X, x 1(x1 = X 1 + uX 1 , y1 = Y1 + uY 1 ) ay = = sin ψ
L0 L0
x21 = X 21 + u X 21

ψ
y21 = Y21 + uY 21
C
L 2(x2 = X 2 + uX 2 , y2 = Y2 + uY 2 )

Figure 8.4. Geometric interpretation of quantities used in the study of element kinematics.

The variation of e induced by variations δ u in the node displacements is

δe = Bl δu + δ(Bn u) = B δu, (8.7)

The matrix B that links δe to δu is

∂(B + Bn u) 1
B= = [ −ax −a y ax ay ] (8.8)
∂u L0

in which
u X 21 x2 − x1 x21 u Y 21 y2 − y1 y21
ax = a X + = = , a y = aY + = = . (8.9)
L0 L0 L0 L0 L0 L0

For a geometric interpretation of ax and a y see Figure 8.4.

8–7
Chapter 8: THE TL BAR ELEMENT 8–8

§8.2.3. Stress Measure

The stress measure conjugate to GL strains is the second Piola-Kirchhoff (PK2) stress tensor. The
only component that appears in the internal energy is the axial stress s, which is related to e through
the constitutive equation
s = s0 + Ee, (8.10)
where s0 is the axial stress in the reference configuration, and E is the elastic modulus.
The axial force based on this stress is
N = A0 s (8.11)
Note that this is not the true axial force in the current configuration C , which would be
Ntr ue = Aσ (8.12)
where σ is the true or Cauchy stress in C and A is the actual area.
§8.2.4. Total Potential Energy and Residual Force Equations
In what follows it is assumed that the element is subjected only to node forces f that are conservative
and proportional, so that f = λq. The Total Potential Energy (TPE) of the element in the current
configuration is
 
=U−P = (s0 e + 2 Ee ) d V0 − f u =
1 2 T
A0 (s0 e + 12 Ee2 ) d X̄ − λqT u. (8.13)
V0 L0

where X̄ is directed along the bar longitudinal axis in C0 , as shown in Figure 8.4. This energy
expression is separable because the internal energy depends only on u through e and not on λ.
The finite element residual equations are obtained by differentiating this equation or, equivalently,
making (8.13) stationary with respect to virtual displacement variations δ u:

δ = δU − δ P = ( p − f)T δu = 0. (8.14)
The first variation of U is given by
  
δU = p δu =
T
A0 (s0 δe + Eeδe) d X̄ = A0 s δe d X̄ = N0 B δu d X̄ = N L 0 B δu
L0 L0 L0
(8.15)
where N = A0 s is the axial force in the current configuration measured per area of the reference
configuration.2 It follows that the internal force vector is
 
−ax
 −a 
p = N0 L 0 BT = N  y  . (8.16)
ax
ay
This equation admits of a simple geometric interpretation; see Figure 8.5. The relation between
N0 = A0 s and the true axial force N = Aσ can be worked out from inspection of this diagram.
The load potential variation δ P simply generates f = λq as can be expected.
2 This is the PK2 axial force; cf (8.11).

8–8
8–9 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT

f Y 1 = −N0 a y

f X 1 = −N0 ax 1
N0
2 f X 2 = N0 a x

f Y 2 = N0 a y

Figure 8.5. Geometrical interpretation of the internal force vector.

The axial force N0 = A0 s would be positive as shown.

§8.2.5. The Tangent Stiffness Matrix

Because the residual equations are separable the tangent stiffness matrix is obtained simply by
differentiating the internal force with respect to the node displacements u:

∂p ∂(N L 0 BT ) T ∂s ∂BT
K= = = A0 L 0 B + A0 L 0 s = K M + KG . (8.17)
∂u ∂u ∂u ∂u
The above expression shows that K splits naturally into two parts: K M and KG , which are called
the material stiffness matrix and geometric stiffness matrix, respectively, in the FEM literature.
To get K M note that
∂s ∂(s0 + Ee) ∂e
= =E = EB. (8.18)
∂u ∂u ∂u
Consequently
K M = E A0 L 0 BT B. (8.19)
Inserting the expression (8.8) for B yields

ax2 ax a y −ax2 −ax a y 
E A0  ax a y a y2 −ax a y −a y2 
KM =   (8.20)
L0 −ax2 −ax a y ax2 ax a y
−ax a y −a y2 ax a y a y2

This component of K looks formally similar to the stiffness matrix of a linear bar element,3 except
that B now depends on u. The dependence of K M on the material properties (here the elastic
modulus E) explains the name “material stiffness” given in the FEM literature.

3 To which it reduces if u = 0. In that case ax and a y become the sine and cosine of the angle ψ0 shown in Figure 8.4.

8–9
Chapter 8: THE TL BAR ELEMENT 8–10

The other component can be obtained by differentiating B with respect to the node displacements,
the result being a constant 4 × 4 matrix:
 
1 0 −1 0
∂B T
1  0 1 0 −1 
= 2 . (8.21)
∂u L 0 −1 0 1 0
0 −1 0 1

Inserting this into (8.17) one gets

 
1 0 −1 0
N  0 1 0 −1 
KG =   . (8.22)
L 0 −1 0 1 0
0 −1 0 1

This component of K depends only on the stress state in the current configuration, because N = A0 s.
No material properties appear. Thus the name “geometric stiffness” applied to KG .4

Remark 8.1. Assuming that E, A0 and L 0 are nonzero, the rank of K M is obviosly one because B is a 1 × 4
matrix. On the other hand the rank of the numerical matrix in (8.21) is 2 (because its eigenvalues are 2, 2, 0, 0).
Consequently KG has rank 2 if s is nonzero and 0 otherwise. Combining these results it can be shown that the
rank of K = K M + KG is 1 if the configuration is stressed and 2 otherwise. In other words, the rank deficiency
is 3 and 2, respectively. The implications of this property in the analysis of stability are studied later.

Remark 8.2. The addition of KG increases the bar stiffness if the current configuration is in tension (s > 0),
but it reduces it if the current configuration is in compression (s < 0). This is in accord with physical intuition.
The main effect of this stiffness is on the rotational rigid-body motions of the bar about the Z axis.

§8.3. FEM Coding Using Mathematica

This section illustrates the use of Mathematica as a quick way to do some simple problems with the TL bar
element. The use of a computer symbolic system as prototyping language has several advantages:
(i) The code is very compact, because of higher lever array notation. Because of its compactness, it can be
easily translated into other high-level languages such as, for example, Matlab or Maple.
(ii) Debugging is often quicker than with programming languages because of the interpretative nature of the
language front end. This is of course platform dependent; the implementation of Mathematica on a
Macintosh or PC is friendler than on a Unix workstation.
(iii) Graphic output is part of the language. On the Macintosh, graphics can be transported into documents
such as this one via PostScript and Adobe Illustrator.
(iv) The language can manipulate both algebraic (symbolic) and numerical expressions. The later may be
done exactly or in floating-point arithmetic. This flexibility is convenient for many situations, as the
examples below illustrate.

4 In the pre-1970 FEM literature, the name “initial stress stiffness” was used for KG by some authors.

8–10
8–11 §8.3 FEM CODING USING MATHEMATICA

fY = λ
Here λ > 0 means crown
load is applied upwards
uY
2 uX

E, A0 (1) (2) E, A0
Y, y

;; ;;
H
α
3

;; ;;
1 X, x

Figure 8.6. An arch as a two-bar FE model. This model has two degrees of
freedom: u X and u Y , which are the displacements of node 2. This
structure will be used as example for use of Mathematica in §8.3.

§8.3.1. The Example Structure

The two-spring arch studied in Exercise 6.3 will be used as a structure that illustrates the use of Mathematica
in conjunction with finite elements. The structure is shown in Figure 8.6. It is modeled with two TL bar
elements labeled (1) and (2), which join nodes 1-2 and 2-3, respectively. The modulus of elasticity and the
bar areas are unity. The arch geometry is defined by the span S and height H . The reference configuration is
assumed stress free.
Because nodes 1 and 3 are pinned to the ground, the only degrees of freedom are the horizontal and vertical
displacements of node 2. To avoid clutter these will be denoted in the sequel as u X and u Y rather than u X 2
and u Y 2 , respectively. The vertical motion u Y is positive upwards. The node-2 components of the applied
and internal force vector will be also denoted simply by ( f X , f Y ) and ( p X , pY ), respectively. Displacement and
force components for nodes 1 and 3 will be explicitly removed from the governing equations.
The arch is loaded by a vertical force f Y = λ applied at node 2, positive upwards. The applied horizontal
force f X will be always zero. Recall that in Exercise 6.3 only symmetric motions of the arch under a vertical
load were considered. That constraint effectively reduces the model to one degree of freedom: u Y , and allows
only limit point behavior. When two degrees of freedoms are considered, however, a richer set of possibilities
emerges; notably lateral bifurcation under a vertical load.

§8.3.2. Forming the Internal Force

The Mathematica module FormIntForce2DTwoNodeBar listed in Figure 8.7 forms the internal force vector
of an individual 2D bar finite element handled by the TL description. The arguments of this module are
XY1 Coordinates {X 1 , Y1 } of end node 1 in C0 .
XY1 Coordinates {X 2 , Y2 } of end node 2 in C0 .
uX1 Displacements {u X 1 , u Y 1 } of end node 1.

8–11
Chapter 8: THE TL BAR ELEMENT 8–12

FormIntForce2DTwoNodeBar[XY1_,XY2_,uXY1_,uXY2_,Em_,A0_,s0_]:=
Module[{X1,Y1,X2,Y2,X21,Y21,uX1,uY1,uX2,uY2,uX21,uY21,L0,L,
e,s,ax,ay,ifv},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1}=uXY1; {uX2,uY2}=uXY2; uX21=uX2-uX1; uY21=uY2-uY1;
L0=Sqrt[X21^2+Y21^2]; L=Sqrt[(X21+uX21)^2+(Y21+uY21)^2];
e=(L-L0)*(L+L0)/(2*L0^2); s=s0+Em*e;
ax=(X21+uX21)/L0; ay=(Y21+uY21)/L0;
ifv=A0*s*{{-ax}, {-ay}, {ax}, {ay}};
Return[Simplify[ifv]]
];
pe=FormIntForce2DTwoNodeBar[{2,3},{5,7},{1,0},{1,0}, 20,12, 0];
Print["Internal elem force pe=",pe];
pe=FormIntForce2DTwoNodeBar[{2,3},{5,7},{1,0},{1,0}, 20,12, s0];
Print["Internal elem force pe=",pe];

Internal elem force pe={{0}, {0}, {0}, {0}}

-36 s0 -48 s0 36 s0 48 s0
Internal elem force pe={{------}, {------}, {-----}, {-----}}
5 5 5 5

Figure 8.7. A Mathematica module that computes the internal force vector p of an individual
2-node 2D bar element. Output from two test statements is shown.

uX2 Displacements {u X 2 , u Y 2 } of end node 2.

Em Elastic modulus E.
A0 Cross sectional area A0 .
s0 PK2 stress in C0
Two test statements, one with purely numeric arguments and the other one with one symbolic argument (s0
for the bar prestress), are shown in the figure along with their output.
Figure 8.8 shows two modules, AssembleMasterIntForceOfTwoBarArch and
MergeElemIntoMasterIntForce that, together with FormIntForce2DTwoNodeBar, form the internal force
vector p for the two-bar example structure. The arguments of the former are the element internal force vector
eif, the element freedom table eftab that maps element-level freedom to master-level freedoms and the
master internal force vector mif before the element is merged. The function returns the updated internal force
vector. The arguments of the second module are:
span Span S
height Height H .
Em Elastic modulus E.
A0 Cross sectional area A0 .
uX,uY Crown displacements u X , u Y .

8–12
8–13 §8.3 FEM CODING USING MATHEMATICA

MergeElemIntoMasterIntForce[pe_,eftab_,p_]:=
Module[{i,ii,neldof,fmaster}, fmaster=p;
neldof=Dimensions[eftab][[1]];
For[i=1, i<=neldof, i++, ii=eftab[[i]];
If [ii>0,fmaster[[ii,1]]+=pe[[i,1]]]
]; Return[fmaster]
];

AssembleMasterIntForceOfTwoBarArch[span_,height_,Em_,A0_,uX_,uY_]:=
Module[{f1,f2,fmaster},
fmaster=Table[0,{2},{1}];
f1=FormIntForce2DTwoNodeBar[{-span/2,0},{0,height},
{0,0},{uX,uY},Em,A0,0];
fmaster=MergeElemIntoMasterIntForce[f1,{0,0,1,2},fmaster];
f2=FormIntForce2DTwoNodeBar[{0,height},{span/2,0},
{uX,uY},{0,0},Em,A0,0];
fmaster=MergeElemIntoMasterIntForce[f2,{1,2,0,0},fmaster];
Return[Simplify[fmaster]]
];

ClearAll[Em,A0,S,H,uX,uY,f];
p=AssembleMasterIntForceOfTwoBarArch[2,2.5,10,.75,-0.4,0.25];
Print["Master Int Force p= ",p];
p=AssembleMasterIntForceOfTwoBarArch[S,H,Em,A0,uX,uY];
Print["Master Int Force p= ",p];

Master Int Force p= {{-0.5336499821957073}, {1.555758744891104}}

Master Int Force p= {{(4*A0*Em*uX*
(S^2 + 2*uX^2 + 4*H*uY + 2*uY^2))/(4*H^2 + S^2)^(3/2)},
{(8*A0*Em*(H + uY)*(uX^2 + 2*H*uY + uY^2))/
(4*H^2 + S^2)^(3/2)}}

Figure 8.8. Two Mathematica modules that assemble the internal force vector p
for the two-bar arch structure of Figure 8.6. Two test statements, one
numeric and the other symbolic, are shown along with output.

Two test statements are also shown in the figure. The first one is purely numeric and sets the following
properties for the structure:
S = 2, H = 2.5, E = 10, A0 = 0.75, u X = 0.4, u Y = −0.25 (8.23)
and the module AssembleMasterIntForceOfTwoBarArch returns
   
pX −0.53365
p= = (8.24)
pY 1.55576
The second one specifies symbolic values for all of the arguments: S for span S, H for the height H , etc. The
result (displayed in InputForm so it can be cut and pasted easily) gives the internal force for an arbitrary

8–13
Chapter 8: THE TL BAR ELEMENT 8–14

structure:  
pX 4E A0 u X (S 2 + 2u 2X + 4H u Y + 2u 2Y )
p= = . (8.25)
pY (4H 2 + S 2 )3/2 2(H + u Y )(u 2X + 2H u Y + u 2Y )

§8.3.3. The Equilibrium Equations

The residual equilibrium equations are of course r = p − f = 0 or

   
pX 0
p= =f= (8.26)
pY λ

It follows that p X = 0 and pY = λ. Expression (8.26) shows that p X = 0 is a cubic equation that has the trivial
root u X = 0 in addition to the two roots of the quadratic S 2 + 2u 2X + 4H u Y + 2u 2Y = 0. Rather than solving
these by hand, let us exemplify the use of Mathematica to solve a polynomial equation using the built-in
Solve function. These roots are then substituted, using the /. operator, into the other equation λ = pY to get
the equations of the 3 equilibrium paths. The appropriate statements and output are shown in Figure 8.9.

ClearAll[Em,A0,S,H,uX,uY,pX,pY,lambda];
pX = (4*A0*Em*uX*(S^2+2*uX^2+4*H*uY+2*uY^2))/(4*H^2+ S^2)^(3/2);
pY = (8*A0*Em*(H+uY)*(uX^2+2*H*uY+uY^2))/(4*H^2+S^2)^(3/2);
roots=Solve[pX==0, uX];
Print["roots of pX=0 are ",roots];
Print["lambda=",Simplify[pY/.roots]];

roots of pX=0 are {{uX -> 0},

{uX -> -((-S^2 - 4*H*uY - 2*uY^2)^(1/2)/2^(1/2))},
{uX -> (-S^2 - 4*H*uY - 2*uY^2)^(1/2)/2^(1/2)}}
lambda={(8*A0*Em*uY*(H + uY)*(2*H + uY))/(4*H^2 + S^2)^(3/2),
(-4*A0*Em*S^2*(H + uY))/(4*H^2 + S^2)^(3/2),
(-4*A0*Em*S^2*(H + uY))/(4*H^2 + S^2)^(3/2)}

Figure 8.9. Illustration of use of Mathematica built-in Solve function to obtain the
roots of p X = 0 in terms of u X , where p X is given by (8.25). These roots
are then substituted into pY = λ with pY from the same equation,
to get the expression of the 3 equilibrium paths in the form λ = λ(u Y ).

The definition of pX in Figure 8.9 was directly set via cut-and-paste from the output shown in Figure 8.8. From
the printed solution it is obvious that the three roots in terms of u X are

p

u X = 0, u s1,s2
X = ± −S 2 /2 − 2H u Y − u 2Y , (8.27)
p
The solution u X , where superscript p stands for “primary,” corresponds to a symmetric deformation of the
arch in its primary equilibrium path. This represents the kind of response studied in Exercise 6.3.

8–14
8–15 §8.3 FEM CODING USING MATHEMATICA

A nonzero u X is possible if the roots u s1 s2

X and u X , where supercripts stand for “secondary,” given by (8.27) are
real. This can occur only if
 = −S 2 /2 − 2H u Y − u 2Y > 0. (8.28)
Because both S and H are positive, condition (8.28) can only happen for negative u Y , that is, downward
displacement of the arch. Furthermore, given S and H the condition (8.28) requires (the analysis using
Mathematica is not shown) that
 
−H + H2 − 1 2
2
S ≥ u Y ≥ −H − H 2 − 12 S 2 , (8.29)

which obviously can only happen if H 2 ≥ S 2 /2, or

H √
= tan α ≥ 2. (8.30)
S/2
p
√
If (8.29)-(8.30) are verified, the equilibrium equation p X = 0 has three real roots: u X = 0, u s1,s2
X = ± .
We shall see that the last two pertain to a secondary equilibrium path that intersects the primary equilibrium
path u X = 0 at two bifurcation points.

§8.3.4. Plotting the Equilibrium Paths

The Mathematica statements shown in Figure 8.10 produces the four response √ plots of λ versus u Y shown in
◦
Figure
√ √ For all plots E = A0 = 1 and S = 2. The height H is set to 3/3 (α = 30 , as in Exercise 6.3),
8.11.
2/2, 3 and 3. These are labeled cases 1 through 4 respectively. All plots show the symmetric-response
(u X = 0) primary path with two limit points passing through λ = 0 at u Y = 0, u X = −H and u Y = −2H .
The straight line is the projection of the plane on which the secondary equilibrium path lies. This path has a
closed elliptical shape quite similar to that sketched in Figure 5.3 in Chapter 5; see also Figure 8.12 below. A
study of the location of the secondary path with respect to the primary path leads to the following conclusions:
√
(a) If H < 2, or equivalently α < 54.74◦ , the secondary path does not interesect the primary path. The
structure has only two critical points, both of which are limit points. This is exemplified by Case 1, for
which α = 30◦ .
√ √
(b) If 2 ≤ H ≤ 3, the secondary path intersects the primary path at two bifurcation points, which
occur between the limit points. When the arch is loaded downward from λ = 0, the first limit point
is encountered first, then the first bifurcation point. The arch will then follow the secondary path as it
snaps through unsymmetrically, until it reaches the second bifurcation point, and proceeds to deform
symmetrically from then on. The structure has four critical points: two limit and two bifurcation, but
the load capacity is determined by the limit points. Cases 2 and 3 in Figure 8.11 illustrates the minimum
and maximum heights for which this kind of behavior occur.
√
(c) If H > 3, which correponds to α = 60◦ , the structure has again four critical points: two limit and
two bifurcation, but the limit points are now “inside” the bifurcation ones. Physically a gradually loaded
arch will bifurcate first and snap through unsymmetrically until it reaches the second bifurcation point
and continues to deform symmetrically from now on. Consequently the limit points cannot be physically
reached. Case 4 in Figure 8.11, in which H = 3, exemplifies this kind of behavior.
The transition between (b) and (c) occurs at α = 60◦ , which is interesting because the limit and bifurcation
points coalesce, as shown in Case 3 of Figure 8.11.
Plotting the equilibrium paths in three-dimensional state-control space (u X , u Y , λ) require far more program-
mming effort because the graphic capabilities of Mathematica begin to be seriously strained. Figure 8.12
illustrates the kind of complicated program needed to get decent results. The logic will not be explained here
because it would not be intelligible unless the reader has substantial
√ expertise in advanced graphics. The
results of running this procedure for the interesting case H = 3 or α = 60◦ are shown in Figure 8.13. This
graph clearly displays the coalescence of the limit and bifurcation points.

8–15
Chapter 8: THE TL BAR ELEMENT 8–16

lambdaP=(8*A0*Em*uY*(H + uY)*(2*H + uY))/(4*H^2 + S^2)^(3/2);

lambdaS=(-4*A0*Em*S^2*(H + uY))/(4*H^2 + S^2)^(3/2);
case1={H->Sqrt[3]/3,S->2,A0->1,Em->1};
case2={H->Sqrt[2], S->2,A0->1,Em->1};
case3={H->Sqrt[3], S->2,A0->1,Em->1};
case4={H->3, S->2,A0->1,Em->1};
Plot[{lambdaP/.case1,lambdaS/.case1},{uY,0,-2*H/.case1},
AxesLabel->{"uY","lambda"},PlotLabel->"Case 1: H=Sqrt[3]/3"];
Plot[{lambdaP/.case2,lambdaS/.case2},{uY,0,-2*H/.case2},
AxesLabel->{"uY","lambda"},PlotLabel->"Case 2: H=Sqrt[2]"];
Plot[{lambdaP/.case3,lambdaS/.case3},{uY,0,-2*H/.case3},
AxesLabel->{"uY","lambda"},PlotLabel->"Case 3: H=Sqrt[3]"];
Plot[{lambdaP/.case4,lambdaS/.case4},{uY,0,-2*H/.case4},
AxesLabel->{"uY","lambda"},PlotLabel->"Case 4: H=3"];

Figure 8.10. Statements to generate λ = λ(u Y ) response plots for four geometries.
√
For all cases E = A√ ◦
0 = 1 and
√ S = 2. The height H is set to 3/3 (α = 30 ,
as in Exercise 6.3), 2/2, 3 and 3.

lambda lambda
Case 1: H=Sqrt[3]/3 Case 2: H=Sqrt[2]
0.2
0.4
0.1
0.2

vY vY
-1 -0.8 -0.6 -0.4 -0.2 -2.5 -2 -1.5 -1 -0.5
-0.2
-0.1
-0.4
-0.2

lambda lambda
Case 3: H=Sqrt[3] Case 4: H=3

0.4 0.3

0.2
0.2
0.1

vY vY
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 -6 -5 -4 -3 -2 -1
-0.1
-0.2
-0.2

-0.4 -0.3

Figure 8.11. Plots of the arch equilibrium paths projected on the λ versus u Y plane
for the four cases described in Figure 8.10.

8–16
8–17 §8.3 FEM CODING USING MATHEMATICA

ClearAll[S,H,Em,A0,uX,uY]; H=Sqrt[3.];
lambdamax[S_,H_,Em_,A0_]:=3.0792*A0*Em*H^3/(4*H^2+S^2)^(3/2);
primarypath[S_,H_,Em_,A0_,uY_]:=
{0,uY,8*A0*Em*uY*(H+uY)*(2*H+uY)/(4*H^2+S^2)^(3/2)};
secondarypath1[S_,H_,Em_,A0_,uY_]:=Module[{c,uYB1,uYB2},
uYB1=Re[N[-H+Sqrt[H^2-S^2/2]]]; uYB2=Re[N[-H-Sqrt[H^2-S^2/2]]];
c=4*A0*Em*S^2/(4*H^2+S^2)^(3/2);
If [N[uY]>=uYB1, Return[{0,uYB1,-c*(H+uYB1)}]];
If [N[uY]<=uYB2, Return[{0,uYB2,-c*(H+uYB2)}]];
Return[{Sqrt[-S^2/2-2*H*uY-uY^2],uY,-c*(H+uY)}]
];
secondarypath2[S_,H_,Em_,A0_,uY_]:=Module[{c,uYB1,uYB2},
uYB1=Re[N[-H+Sqrt[H^2-S^2/2]]]; uYB2=Re[N[-H-Sqrt[H^2-S^2/2]]];
c=4*A0*Em*S^2/(4*H^2+S^2)^(3/2);
If [N[uY]>=uYB1, Return[{0,uYB1,-c*(H+uYB1)}]];
If [N[uY]<=uYB2, Return[{0,uYB2,-c*(H+uYB2)}]];
Return[{-Sqrt[-S^2/2-2*H*uY-uY^2],uY,-c*(H+uY)}]
];
lambdarange=1.1*{-lambdamax[2,H,1,1],lambdamax[2,H,1,1]};
pp=ParametricPlot3D[primarypath[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
ps1=ParametricPlot3D[secondarypath1[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
ps2=ParametricPlot3D[secondarypath2[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
Show[pp,ps1,ps2,ViewPoint->{3,1,2},DisplayFunction->$DisplayFunction];

Figure 8.12. Statements used to generate response plots in the three-dimensional state-control
space (u X , u Y , λ), such as the one shown in the next figure.

§8.3.5. Having Some Fun: Snapshots and Animation

To illustrate what happens to the arch as it traverses its equilibrium path, configuration “snapshots” and
particularly animations are invaluable and fun to do. The program shown in 8.14 does the first kind of display,
showing configurations as the arch traverses the equilibrium path by incrementing u Y . The program also
depicts (very roughly) the magnitude of the applied vertical force. The result of running this program on an
arch with S = 2, H = 3, E = A0 = 1 is shown in Figure 8.15.
A plot such as that of Figure 8.15 is a bit confusing because all configurations are in the same cell, which
appears on the screen as a flash. More instructive is the use of animations. Mathematica animations are
much like movies or videos: they are a sequence of pictures that, when displayed in rapid succession, appear
to move. In Mathematica versions endowed with a Notebook interface, any sequence of graphic cells can

8–17
Chapter 8: THE TL BAR ELEMENT 8–18

-1
uX 0

0.2

λ 0

-0.2

-3
-2
-1
uY
0

Figure 8.13. The arch equilibrium paths for α = 60◦ , plotted in

the three-dimensional state-control space (u X , u Y , λ).

be animated by simply doubly clicking the mouse on one of the pictures (any one will do). The cell group
containing the pictures will be selected and the animation started automatically.
The program shown in Figure 8.16 is a variant of the program of Figure 8.14. The logic is rearranged to
produce a sequence of graphic cells, one per configuration, and so make the animation possible. Unfortunately
the result cannot be shown on the old-fashioned medium of these Notes. But it should be possible in the next
millennium.

§8.3.6. Forming the Tangent Stiffness Matrix

The tangent stiffness matrix derived in §8.2 is computed by the Mathematica module shown in Figure 8.17,
along with some test statements. The arguments of this module, called FormTangentStiff2DTwoNodeBar
are the same as those of the internal force subroutine shown in Figure 8.7. The module returns the 4 × 4
K = K M + KG matrix for the element. Figure 8.18 lists two modules: MergeElemIntoMasterStiff and
AssembleMasterStiffOfShallowArch, which together assemble the 2 × 2 master stiffness matrix for the
example arch structure.
The main use of the stiffness matrix in this and subsequent Chapters is the investigation of critical points,
because solution argorithms based on K have not yet been described.

§8.3.7. Critical Point Study

A detailed symbolic analysis, not reported here, for arbitrary S, H , E, A0 , u X and u Y shows that under a
vertical load the determinant of K can only vanish if u X = 0. Thus it is enough to restrict consideration to
symmetric motions. The stiffness matrix for u X = 0 and its determinant are produced by the test statements
shown in Figure 8.19 along with their output. It is seen that

8–18
8–19 §8.3 FEM CODING USING MATHEMATICA

ClearAll[S,H,Em,A0,uX,uY,a,c,d]; S=2.;H=3.;A0=1; Em=1;

nmax =20; c= N[4*A0*Em/(4*H^2+S^2)^(3/2)];
plotelem1=Table[0,{nmax+1}];plotelem2=Table[0,{nmax+1}];
plotload =Table[0,{nmax+1}];plotarrow=Table[0,{nmax+1}];
Do [uY=-2.5*n*H/nmax; d= N[-S^2/2-2*H*uY-uY^2];
If[d <=0., uX=0; lambda= N[2*c*uY*(H+uY)*(2*H+uY)],
uX=Sqrt[d]; lambda= N[-c*S^2*(H+uY)]];
plotelem1[[n+1]]=Graphics[Line[{{-S/2,0},{uX,H+uY}}]];
plotelem2[[n+1]]=Graphics[Line[{{ uX,H+uY},{S/2,0}}]];
plotload[[n+1]] =Graphics[Line[{{ uX,H+uY},{uX,H+uY-lambda}}]];
a = Min[.5*Abs[lambda],0.2*H];
If [lambda>0, plotarrow[[n+1]]=Graphics[{
Line[{{ uX,H+uY},{uX-a/2,H+uY-a}}],
Line[{{ uX,H+uY},{uX+a/2,H+uY-a}}]} ],
plotarrow[[n+1]]=Graphics[{
Line[{{ uX,H+uY},{uX-a/2,H+uY+a}}],
Line[{{ uX,H+uY},{uX+a/2,H+uY+a}}]} ] ],
{n,0,nmax}];
Show[Graphics[Thickness[.004]],plotelem1,plotelem2,
Graphics[Thickness[.005]],Graphics[RGBColor[1,0,1]],plotload,plotarrow,
PlotRange->{{-2*H,2*H},{-2*H,2*H}},AspectRatio->1];

Figure 8.14. A Mathematica program to generate and plot configuration snapshots

of the arch structure as it traverses the equilibrium paths.

Figure 8.15. Results of running the program of Figure 8.14.

8–19
Chapter 8: THE TL BAR ELEMENT 8–20

ClearAll[S,H,Em,A0,uX,uY,a,c,d]; S=2.;H=1.5;A0=1; Em=1;

nmax =20; c= N[4*A0*Em/(4*H^2+S^2)^(3/2)];
Do [uY=-2.5*n*H/nmax; d= N[-S^2/2-2*H*uY-uY^2];
If[d <=0., uX=0; lambda= N[2*c*uY*(H+uY)*(2*H+uY)],
uX=Sqrt[d]; lambda= N[-c*S^2*(H+uY)]];
plotelem1=Graphics[Line[{{-S/2,0},{uX,H+uY}}]];
plotelem2=Graphics[Line[{{ uX,H+uY},{S/2,0}}]];
plotload =Graphics[Line[{{ uX,H+uY},{uX,H+uY-lambda}}]];
a = Min[.5*Abs[lambda],0.2*H];
If [lambda>0, plotarrow=Graphics[{
Line[{{ uX,H+uY},{uX-a/2,H+uY-a}}],
Line[{{ uX,H+uY},{uX+a/2,H+uY-a}}]} ],
plotarrow=Graphics[{
Line[{{ uX,H+uY},{uX-a/2,H+uY+a}}],
Line[{{ uX,H+uY},{uX+a/2,H+uY+a}}]} ] ];
Show[Graphics[Thickness[.004]],plotelem1,plotelem2,
Graphics[Thickness[.005]],Graphics[RGBColor[1,0,1]],plotload,plotarrow,
PlotRange->{{-2*H,2*H},{-2*H,2*H}},AspectRatio->1],
{n,0,nmax}];

Figure 8.16. A variant of the program of Figure 8.14. This program generates a
sequence of configuration plots (frames) that can be used for animation.

 
8E A0 S 2 /2 + 2H u Y + u 2Y 0
K(S, H, E, A0 , u X = 0, u Y ) =  , (8.31)
(4H 2 + S 2 )3 0 2H 2 + 6H u Y + 3u 2Y

Therefore 
det(K) = 32E 2 A20 (S 2 + 4H u Y + 2u 2Y )(2H 2 + 6H u Y + 3u 2Y )/ (4H 2 + S 2 )3 . (8.32)
Because K is diagonal if u X = 0, det(K) factors out into
√ two quadratic polynomials in u Y . Consequently there
are four critical points (C.P.s). The roots −1 ± H/ 3 of 2H 2 + 6H u Y + 3u 2Y = 0 are always real and can
be easily checked to correspond to limit points. The other two, which are roots of S 2 /2 + 2H u Y + u 2Y = 0,
correspond to bifurcation points and are real if 2H 2 ≥ S 2 , as already announced in §8.3.4.
The expression of the four roots of det(K ) = 0 in terms of u Y is symbolically obtained using the statements
listed in Figure 8.19. These roots are substituted, again using the /. operator, into the equations λ(u Y ) of the
primary equilibrium path to get the four values of λ shown in the output cell of Figure 8.19:
√ 
16E A0 H 3 2 2 E A0
λ L1 = −λ L2 = − √  , λ B1 = −λ B2 = −  S 2 2H 2 − S 2 . (8.33)
3 3 (4H 2 + S 2 )3 (4H 2 + S 2 )3

The subscripts L and B stand here for limit and bifurcation point, respectively. That classification is readily
done given the diagonal nature of K shown in (8.31).
Figure 8.20 shows statements used to produce plots of λ L1 and λ B1 versus H upon setting E = A0 = 1
and S = 2. Thus H = tan α. Negative λ (downward load) is plotted upwards for convenience. The plot

8–20
8–21 §8.3 FEM CODING USING MATHEMATICA

FormTangentStiff2DTwoNodeBar[XY1_,XY2_,uXY1_,uXY2_,Em_,A0_,s0_]:=
Module[{X1,Y1,X2,Y2,X21,Y21,uX21,uY21,L0,L,e,s,ax,ay,Ke},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1}=uXY1; {uX2,uY2}=uXY2; uX21=uX2-uX1; uY21=uY2-uY1;
L0=Sqrt[X21^2+Y21^2]; L=Sqrt[(X21+uX21)^2+(Y21+uY21)^2];
e=(L-L0)*(L+L0)/(2*L0^2); s=s0+Em*e;
ax=(X21+uX21)/L0; ay=(Y21+uY21)/L0;
Ke=(Em*A0/L0)*{{ ax*ax, ax*ay,-ax*ax,-ax*ay},
{ ax*ay, ay*ay,-ay*ax,-ay*ay},
{-ax*ax,-ay*ax, ax*ax, ay*ax},
{-ax*ay,-ay*ay, ay*ax, ay*ay}} +
(A0*s/L0)*{{1,0,-1,0},{0,1,0,-1},{-1,0,1,0},{0,-1,0,1}};
Return[Simplify[Ke]]
];

Ke=FormTangentStiff2DTwoNodeBar[{0,-S+H},{0,H},{0,0},{0,uY},Em,A0,0];
Ke=Simplify[Ke/.{Sqrt[S^2]->S,(1/Sqrt[S^2])->1/S}];
Print["Ke=",Ke//InputForm];
Ke=FormTangentStiff2DTwoNodeBar[{-4,0},{0,3},{0,0},{0,0}, Em,A0,s0];
Print["Ke=",Ke];
Print[Eigenvalues[Ke]];

Ke={{(A0*Em*uY*(2*S + uY))/(2*S^3), 0, -(A0*Em*uY*(2*S + uY))/(2*S^3),

0}, {0, (A0*Em*(2*S^2 + 6*S*uY + 3*uY^2))/(2*S^3), 0,
-(A0*Em*(2*S^2 + 6*S*uY + 3*uY^2))/(2*S^3)},
{-(A0*Em*uY*(2*S + uY))/(2*S^3), 0, (A0*Em*uY*(2*S + uY))/(2*S^3),
0}, {0, -(A0*Em*(2*S^2 + 6*S*uY + 3*uY^2))/(2*S^3), 0,
(A0*Em*(2*S^2 + 6*S*uY + 3*uY^2))/(2*S^3)}}
2 A0 s0 2 A0 (Em + s0)
{0, 0, -------, --------------}
5 5

Figure 8.17. A module that generates the tangent stiffness matrix K of an

individual bar element, along with test statements and output.

is shown in Figure 8.21 upon some labeling √ and beautification via Adobe Illustrator. The coalescence
of the bifurcation and limit points for H = 3 or α = 60◦ is clear in the plot. This figure also shows that
for fixed S, E and A0 the maximum vertical load carrying capacity for this structure is obtained for that rise
angle. This optimality criterion (coalescence of critical points gives the strongest structure) is typical of those
encountered in optimal design with stability constraints.

8–21
Chapter 8: THE TL BAR ELEMENT 8–22

MergeElemIntoMasterStiff[Ke_,eftab_,Km_]:=
Module[{i,j,ii,jj,neldof,K}, K=Km;
neldof=Length[eftab];
For[i=1, i<=neldof, i++, ii=eftab[[i]];
For[j=i, j<=neldof, j++, jj=eftab[[j]];
If [ii>0 && jj>0,
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]]
]
]; Return[K]
];

AssembleMasterStiffOfShallowArch[S_,H_,Em_,A0_,uX_,uY_]:=
Module[{K1,K2,K},
K=Table[0,{2},{2}];
K1=FormTangentStiff2DTwoNodeBar[{-S/2,0},{0,H},
{0,0},{uX,uY},Em,A0,0]; (*Print[K1//TableForm];*)
K=MergeElemIntoMasterStiff[K1,{0,0,1,2},K];
K2=FormTangentStiff2DTwoNodeBar[{0,H},{S/2,0},
{uX,uY},{0,0},Em,A0,0]; (*Print[K2//TableForm];*)
K=MergeElemIntoMasterStiff[K2,{1,2,0,0},K];
Return[Simplify[K]]
];

ClearAll[S,H,Em,A0,uX,uY];
K=AssembleMasterStiffOfShallowArch[S,H,Em,A0,uX,uY];
Print["Master stiffness matrix = ",K//InputForm];

Master stiffness matrix = {{(4*A0*Em*(S^2 + 4*H*uY + 2*uY^2))/

(4*H^2 + S^2)^(3/2), 0},
{0, (8*A0*Em*(2*H^2 + 6*H*uY + 3*uY^2))/(4*H^2 + S^2)^(3/2)}}

Figure 8.18. Two modules that assemble the tangent stiffness matrix K for the arch
structure of Figure 8.6.

8–22
8–23 §8.3 FEM CODING USING MATHEMATICA

ClearAll[H,S,Em,A0,uY];
detK=(32*A0^2*Em^2*(S^2 + 4*H*uY + 2*uY^2)*
(2*H^2 + 6*H*uY + 3*uY^2))/(4*H^2 + S^2)^3;
sol=Simplify[Solve[detK==0,uY]];
Print[sol//InputForm];
lambdaCP=(8*A0*Em*uY*(H + uY)*(2*H + uY))/(4*H^2 + S^2)^(3/2)/.sol;
lambdaCP=Simplify[Expand[lambdaCP]];
Print["lambdaCP=",lambdaCP//InputForm];

{{uY -> -((3 + 3^(1/2))*H)/3}, {uY -> ((-3 + 3^(1/2))*H)/3},

{uY -> (-4*H - (16*H^2 - 8*S^2)^(1/2))/4},
{uY -> (-4*H + (16*H^2 - 8*S^2)^(1/2))/4}}
lambdaCP={(16*A0*Em*H^3)/(3*3^(1/2)*(4*H^2 + S^2)^(3/2)),
(-16*A0*Em*H^3)/(3*3^(1/2)*(4*H^2 + S^2)^(3/2)),
(2*2^(1/2)*A0*Em*S^2*(2*H^2 - S^2)^(1/2))/(4*H^2 + S^2)^(3/2),
(-2*2^(1/2)*A0*Em*S^2*(2*H^2 - S^2)^(1/2))/(4*H^2 + S^2)^(3/2)}

Figure 8.19. Obtaining the critical points as roots of the tangent stiffness determinant.

8–23
Chapter 8: THE TL BAR ELEMENT 8–24

ClearAll[H,S,Em,A0,uY];
lambdaL=(16*A0*Em*H^3)/(3*3^(1/2)*(4*H^2 + S^2)^(3/2));
lambdaB=(2*2^(1/2)*A0*Em*S^2*(2*H^2 - S^2)^(1/2))/(4*H^2 + S^2)^(3/2);
set={Em->1,A0->1,S->2};
pL=Plot[lambdaL/.set,{H,0,5},DisplayFunction->Identity,
AxesLabel->{"H","lambda at CPs"},
PlotLabel->"lambdaCP vs H (S=2,E=A0=1)"];
pB=Plot[lambdaB/.set,{H,Sqrt[2.00001],5},DisplayFunction->Identity];

Figure 8.20. Program to produce the critical loads plots of next Figure.

lambda at CPs
lambdaCP vs H (S=2,E=A0=1)

0.35 √ λ L1
H= 3
0.3

0.25 Limit pt Bifurcation pt

first first
0.2

0.15
λ B1
λ L1
0.1 λ B1
0.05

H
1
√ 2 3 4 5
2

Figure 8.21. Plot of maximum load capacity as defined by λ (negative upwards) at the
critical points for arches with different heights H (S = 2, E = A0 = 1).

8–24
8–25 Exercises

Homework Exercises for Chapter 8

The TL Bar Element

EXERCISE 8.1 [D:10] The α = 30◦ (case 1) plot in Figure 8.11 shows a limit-point λ of about 0.048. Explain
why that value could be different from that of the “exact solution” plot of Figure E6.3 for the same structure.

EXERCISE 8.2 [A/C:20] A problem in optimal design. For the 2-bar arch example structure, the plot in
Figure 8.21 makes evident that, given the span S, modulus E and cross section area A0 , the√ largest vertical-
downward-load capacity is obtained if the rise angle is α = 60◦ , or H = (S/2) tan α = S 3/2. Suppose
that the design objective is to get the largest possible λmax = |λ| per unit volume of material, with respect to
the total volume. Mathematically: maximize λmax /V0 with respect to V0 , where λmax is the largest downward
load sustained before failure by snap-through or bifurcation. What would be the best α?

fY = λ

uY
2 uX

E, A0 (1) (2) E, A0

H Y, y

;; ;;
α
1 X, x 3

S
S
(3)
E, A0

;; 4

Figure E8.1. A 3-bar FEM model for Exercise 8.3.

EXERCISE 8.3 [A/C:25] You go to work as a nonlinear-FEM engineer for a car company. Your supervisor
assigns you the job of designing a component of a wheel suspension system that can be modeled by the 3-bar
structure depicted in Figure E8.1. The model has the dimensions and properties shown and is only subjected
to vertical loads at node 2. The length S, bar section areas A0 and elastic modulus E are known, but the rise
angle α > 0 is a design variable. Find the largest α for which bifurcation, which is bad for the wheel,
√ cannot
occur. (For the 2-bar arch example structure that maximum α was shown to be defined by tan α ≤ 2/2.)

8–25
Chapter 8: THE TL BAR ELEMENT 8–26

;;
Y, y

Bar element models

weightless string in
reference configuration
E, A0
L

(1)
Gravity field g

uY

1 uX X, x

Point mass m

Figure E8.2. Model of a classic pendulum for Exercise 8.4.

EXERCISE 8.4 [A:C:20] Although this course focuses on statics, this exercise deals with the effect of the
geometric stiffness on vibrations. Consider the pendulum configuration idealized in Figure E8.2. A lumped
mass m is suspended by a weightless elastic string. The string is modeled as a 2-node bar element. This
element is under a tensile prestress s0 = mg/A0 , where g is the accelaration of gravity. The tangent stiffness
matrix for the cable element in the reference configuration is K = K M + KG , which is 2 × 2 upon removing
the degrees of freedom at the fixed node 2. Because of the prestress the geometric stiffness does not vanish.
The order-2 vibration eigenproblem is

Kzi = ωi2 Mzi , i = 1, 2 (E8.1)

where i is the mode index, ωi is the i th circular frequency in radians per second, zi the associated eigenvector
that include the horizontal and vertical displacements of node 1, and the mass matrix is
 
m 0
M= (E8.2)
0 m
Compute the two frequencies ω1 and ω2 . One of them, say ω1 , describes pendulum motions while the other
one pertains to a “bar mode” associated with axial motions. Discuss what happens to ω1 and ω2 if E → ∞,
which characterizes
√ the “inextensional string” limit, and whether the classical pendulum small-oscillations
frequency ω P = g/L is correct.

EXERCISE 8.5 (Requires knowledge of continuum mechanics.) [A:15] Suppose that the bar-element material
is linear isotropic, with elastic modulus E and ν is Poisson’s ratio ν. Find the relation between the true (Cauchy)
axial stress σ = σx x in the bar and the PK2 axial stress s = s X X . Hint: study the change in cross section areas.

8–26
 7
.

Review of
Continuum
Mechanics

7–1
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–2

TABLE OF CONTENTS

Page
§7.1. The FEM Discretization 7–3
§7.2. Notation: Continuum vs. Discrete Mechanics 7–3
§7.3. Particles, Motions, Displacements, Configurations 7–4
§7.3.1. Distinguished Configurations . . . . . . . . . . . . . 7–4
§7.3.2. Kinematic Descriptions . . . . . . . . . . . . . . 7–6
§7.3.3. Coordinate Systems . . . . . . . . . . . . . . . . 7–7
§7.3.4. Configurations and Staged Analysis . . . . . . . . . . 7–8
§7.4. Kinematics 7–8
§7.4.1. Deformation and Displacement Gradients . . . . . . . . . 7–8
§7.4.2. Stretch and Rotation Tensors . . . . . . . . . . . . . 7–9
§7.4.3. Green-Lagrange Strain Measure . . . . . . . . . . . . 7–10
§7.4.4. Strain-Gradient Matrix Expressions . . . . . . . . . . 7–11
§7.4.5. Pull Forward and Pull Back . . . . . . . . . . . . . . 7–12
§7.5. Stress Measure 7–12
§7.6. Constitutive Equations 7–13
§7.7. Strain Energy Density 7–14
§7. Exercises . . . . . . . . . . . . . . . . . . . . . . . 7–15

7–2
7–3 §7.2 NOTATION: CONTINUUM VS. DISCRETE MECHANICS

§7.1. The FEM Discretization

In Chapters 3 through 6 we have studied some general properties of the governing force-equilibrium
equations of geometrically nonlinear structural systems with finite number of degrees of freedom (DOF).
The DOFs are collected in the state variable vector u. Those residual equations, being algebraic, are
well suited for numerical computation.
Mathematical models of real structures, however, possess an infinite number of DOFs. As such they
cannot be handled by numerical computations. The reduction to a finite number is accomplished by
discretization methods. As noted in Chapter 1, for nonlinear problems in solid and structural mechanics
the finite element method (FEM) is the most widely used discretization method.
This section provides background material for the derivation of geometrically nonlinear finite elements.
The material gives a review of kinematic, kinetic and constitutive concepts from the three-dimensional
continuum mechanics of an elastic deformable body, as needed in following Chapters. Readers familiar
with continuum mechanics should peruse it to grab notation.

§7.2. Notation: Continuum vs. Discrete Mechanics

Continuum mechanics deals with vector and tensor fields such as displacements, strains and stresses.
Four types of notation are in common use:
1. Indicial Notation. Also called component notation. This notation uses indexed components along
with abbreviation rules such as commas for partial derivatives and Einstein’s summation convention.
It is a powerful notation, and as such is preferred in journals and monographs. It has the advantage
of readily handling arbitrary tensors of any order, arbitrary coordinate systems and nonlinear relations.
It sharply distinguishes between covariant and contravariant quantities, which is necessary in non-
Cartesian coordinates. Tends to conceal or mask intrinsic properties, however, and as such is not
suitable for basic instruction.
2. Direct Notation. Sometimes called algebraic notation. Vectors and tensors are represented by single
symbols, usually bold letters. Has the advantage of compactness and quick visualization of intrinsic
properties. Some operations correspond to matrix notations while others do not. This fuzzy overlap
can lead to confusion in FEM work.
3. Matrix Notation. This is similar to the previous one, but entities are rearranged as appropriate
so that only matrix operations are used. It can be translated directly to discrete equations as well as
matrix-oriented programming languages such as Matlab. It has the disadvantage of losing contact with
the original physical entities along the way. For example, stress is a symmetric second-order tensor that
is rearranged as a 6-component vector for FEM developments. This change loses essential properties.
For instance it makes sense to talk of principal stresses as eigenvalues of the stress tensor. But those
get lost (as or least moved to the background) when rearranged as a stress vector.
4. Full Notation. In the full-form notation every term is spelled out. No ambiguities of interpretation
can arise; consequently this works well as a notation of last resort, and also as a “comparison template”
against one can check out the meaning of more compact expressions. It is also useful for programming
in low-order languages.
As an example, consider the well known dot product between two physical vectors in 3D space, a =
(a1 , a2 , a3 ) and b = (b1 , b2 , b3 ) done in the four different notations:

ai bi = 
a.b = 
aT b = a1 b1 + a2 b2 + a3 b3 . (7.1)
   
indicial direct matrix full

7–3
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–4

In the following review the direct, matrix and full notation are preferred, whereas the indicial notation
is avoided. Usually the expression is first given in direct form and confirmed by full form if feasible.
Then it is transformed to matrix notation for later used in FEM developments. The decision leads to
possible ambiguities against reuse of vector symbols in two contexts: continuum mechanics and FEM
discretizations. Such ambiguities are resolved in favor of keeping FEM notation simple.

§7.3. Particles, Motions, Displacements, Configurations

In the present section a structure is mathematically treated as a continuum body B. The body is
considered as being formed by a set of points P called particles, which are endowed with certain
mechanical properties. For FEM analysis the body is divided into elements.
Particles displace or move in response to external actions characterized by control parameters i or,
following the stage reduction discussed in Chapter 3, the single stage parameter λ. A one-parameter
series of positions occupied by the particles as they move in space is called a motion. The motion may
be described by the displacement u(P) ≡ u(x) of the particles with respect to a base or reference state
in which particle P is labelled P0 . [The underlining is used to distinguish the physical displacement
vector from the finite element node displacement array, which is a computational vector.]
The displacements of all particles u(x, y, z) such that x ≡ x, y, z ∈ B, constitutes the displacement
field.
The motion is said to be kinematically admissible if:

1. Continuity of particles positions is preserved so that no gaps or voids appear. (The mathematical
statement of this condition is given later.)

2. Kinematic constraints on the motion (for example, support conditions) are preserved.
A kinematically admissible motion along a stage will be called a staged motion. For one such motion
the displacements u(x) characterize the state and the stage control parameter λ characterizes the control
or action. Both will be generally parametrized by the pseudo-time t introduced in Chapter 3, so that a
staged motion can be generally represented by

λ = λ(t), u = u(x, t), x ∈ B. (7.2)

If in these equations we freeze t, we have a configuration of the structure. Thus a configuration

is formally the union of state and control. It may be informally viewed as a “snapshot” taken of the
structure and actions upon it when the pseudotime is frozen. If the configuration satisfies the equilibrium
equations, it is called an equilibrium configuration. In general, however, a randomly given configuration
is not in equilibrium unless artificial body and surface forces are applied to it.
A staged response, or simply response, can be now mathematically defined as a series of equilibrium
configurations obtained as λ is continuously varied.

§7.3.1. Distinguished Configurations

A particular characteristic of geometrically nonlinear analysis is the need to carefully distinguish among
different configurations of the structure.
As noted above, each set of kinematically admissible displacements u(x) plus a staged control parameter
λ defines a configuration. This is not necessarily an equilibrium configuration; in fact it will not usually

7–4
7–5 §7.3 PARTICLES, MOTIONS, DISPLACEMENTS, CONFIGURATIONS

Table 7.1 Distinguished Configurations in Nonlinear Analysis

Name Alias Definition Equilibrium Identification

Required?
Admissible A kinematically admissible configuration No C
Perturbed Kinematically admissible variation No C + δC
of an admissible configuration.
Deformed Current Actual configuration taken during No C D or C(t)
Spatial the analysis process. Contains
others as special cases.
Base∗ Initial The configuration defined as the Yes C 0 , C B or C(0)
Undeformed origin of displacements. Strain free but
Material not necessarily stress free.
Reference Configuration to which stepping TL,UL: Yes. TL: C 0 , UL: C n−1 ,
computations are referred CR: C no, C yes CR: C R and C 0
R 0

Iterated† Configuration taken at the k th No Ckn

iteration of the n th increment step
Target† Equilibrium configuration accepted Yes Cn
at the n th increment step
Corotated‡ Shadow Body or element-attached configuration No CR
Ghost obtained from C 0 through a rigid
body motion (CR description only)
Aligned Preferred A fictitious body or element configuration No CA
Directed aligned with a particular set of axes
(usually global axes)
∗
C 0 is often the same as the natural state in which body (or element) is undeformed and stress-free.
† Used only in the description of solution procedures.
‡ In dynamic analysis C 0 and C R are called the inertial and dynamic-reference configurations,
respectively, when they apply to the entire structure.

be one. It is also important to realize that an equilibrium configuration is not necessarily a physical
configuration assumed by the actual structure.1
Some configurations that are important in geometrically nonlinear analysis receive special qualifiers:
admissible, perturbed, deformed, base, reference, iterated, target, corotated, aligned

This terminology is collected in Table 7.1.

Remark 7.1. A great number of names can be found for these configurations in the literature in finite elements and
continuum mechanics. To further compound the confusion, here are some of these alternative names.
Perturbed configuration: adjacent, deviated, disturbed, incremented, neighboring, varied, virtual.

1 Recall the suspension bridge under zero gravity of Chapter 3.

7–5
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–6

Current configuration C = C(t).

P(x)

x = x(t)
Z, z

u=x−X
V
X, x Y, y X ≡ x0 P0 (X)
Base configuration C0
(often same as reference)

Figure 7.1. The geometrically nonlinear problem in a Lagrangian kinematic description:

coordinate systems, reference and current configurations, and displacements.
For many (but not all) problems, the base configuration, which is not shown
separately in the Figure, would be the same as the reference configuration.

Deformed configuration: arbitrary, current, distorted, moving, present, spatial, varying.

Base configuration: baseline, initial, material, global, natural, original, overall, undeformed, undistorted.
Reference configuration: fixed, frozen, known.
Iterated configuration: intermediate, stepped.
Target configuration: converged, equilibrated spatial, unknown.
Corotated configuration: attached, convected, ghost, phantom, shadow.
Aligned configuration: directed, preferred.
In FEM treatments of nonlinear analysis, confusion often reigns supreme. A common scenario is to identify base
and reference configurations in Total Lagrangian descriptions.

§7.3.2. Kinematic Descriptions

Three kinematic descriptions of geometrically nonlinear finite element analysis are in current use in
programs that solve nonlinear structural problems. They can be distinguished by the choice of reference
configuration.
1. Total Lagrangian description (TL). The reference configuration is seldom or never changed: often
it is kept equal to the base configuration throughout the analysis. Strains and stresses are measured
with respect to this configuration.
2. Updated Lagrangian description (UL). The last target configuration, once reached, becomes the
next reference configuration. Strains and stresses are redefined as soon as the reference configu-
ration is updated.
3. Corotational description (CR). The reference configuration is “split.” Strains and stresses are mea-
sured from the corotated configuration whereas the base configuration is maintained as reference
for measuring rigid body motions.

7–6
7–7 §7.3 PARTICLES, MOTIONS, DISPLACEMENTS, CONFIGURATIONS

Remark 7.2. The TL formulation remains the most widely used in continuum-based finite element codes. The CR
formulation is gaining in popularity for structural elements such as beams, plates and shells. The UL formulation
is primarily used in treatments of vary large strains and flow-like behavior.

§7.3.3. Coordinate Systems

Configurations taken by a body or element during the response analysis are linked by a Cartesian global
frame, to which all computations are ultimately referred.2 There are actually two such frames:
(i) The material global frame with axes {X i } or {X, Y, Z }.
(ii) The spatial global frame with axes {xi } or {x, y, z}.3
The material frame tracks the base configuration whereas the spatial frame tracks all others. This
distinction agrees with the usual conventions of classical continuum mechanics. In the present work
both frames are taken to be identical, as nothing is gained by separating them. Thus only one set of
global axes, with dual labels, is drawn in Figure 7.1.
In stark contrast to global frame uniqueness, the presence of elements means there are many local frames
to keep track of. More precisely, each element is endowed with two local Cartesian frames:
(iii) The element base frame with axes { X̃ i } or { X̃ , Ỹ , Z̃ }.
(iv) The element reference frame with axes {x̄i } or {x̄, ȳ, z̄}.
The base frame is attached to the base configuration. It remains fixed if the base is fixed. It is chosen
according to usual FEM practices. For example, in a 2-node spatial beam element, X̃ 1 is defined by the
two end nodes whereas X̃ 2 and X̃ 3 lie along principal inertia directions. The origin is typically placed
at the element centroid.
The meaning of the reference frame depends on the description chosen:
1. Total Lagrangian (TL). The reference frame and base frame coalesce.
2. Updated Lagrangian (UL). The reference frame is attached to the reference configuration, and
recomputed when the reference configuration (the previous converged solution) is updated. It
remains fixed during an iterative process.
3. Corotational description (CR). The reference frame is renamed corotated frame or CR frame. It
remains attached to the element and continuosly moves with it.
The transformation
x = X + u, (7.3)
maps the location of base particle P(X, Y, Z ) to P(x, y, z); see Figure 7.1. Consequently the particle
displacement vector is defined as
   
uX x−X
u = u Y = y − Y = x − X. (7.4)
uZ z−Z
in which (X, Y, Z ) and (x, y, z) pertain to the same particle.

2 In dynamic analysis the global frame may be moving in time as a Galilean or inertial frame. This is convenient to track the
motion of objects such as aircraft or satellites.
3 The choice between {X 1 , X 2 , X 3 } versus {X, Y, Z } and likewise {x1 , x2 , x3 } versus {x, y, z} is a matter of convenience. For
example, when developing specific finite elements it is preferable to use {X, Y, Z } or {x, y, z} so as to reserve coordinate
subscripts for node numbers.

7–7
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–8

Remark 7.3. Variations of this notation scheme are employed as appropriate to the subject under consideration.
For example, the coordinates of P in a target configuration Cn may be called (xn , yn , z n ).
Remark 7.4. In continuum mechanics, coordinates (X, Y, Z ) and (x, y, z) are called material and spatial coor-
dinates, respectively. In general treatments both systems are curvilinear and need not coalesce. The foregoing
relation (7.4) is restrictive in two ways: the base coordinate systems for the reference and current configurations
coincide, and that system is Cartesian. This assumption is sufficient, however, for the problems treated in this
course.
Remark 7.5. The dual notation (X, Y, Z ) ≡ (x 0 , y0 , z 0 ) is introduced on two accounts: (1) the use of (x 0 , y0 , z 0 )
sometimes introduces a profusion of additional subscripts, and (2) the notation agrees with that traditional in
continuum mechanics for the material coordinates as noted in the previous remark. The identification X ≡ x0 ,
Y ≡ y0 , Z ≡ z 0 will be employed when it is convenient to consider the reference configuration as the initial target
configuration (cf. Remark 7.1).
§7.3.4. Configurations and Staged Analysis
The meaning of some special configurations can be made more precise if the nonlinear analysis process is viewed
as a sequence of analysis stages, as discussed in Chapter 3. We restrict attention to the Total Lagrangian (TL) and
Corotational (CR) kinematic descriptions, which are the only ones covered in this course. In a staged TL nonlinear
analysis, two common choices for the reference configuration are:
(1) Reference ≡ base. The base configuration is maintained as reference configuration for all stages.
(2) Reference ≡ stage start. The configuration at the start of an analysis stage, i.e. at λ = 0, is chosen as reference
configuration.
A combination of these two strategies can be of course adopted. In a staged CR analysis the reference is split
between base and corotated. The same update choices are available for the base. This may be necessary when
rotations exceed 2π; for example in aircraft maneuvers.
The admissible configuration is a “catch all” concept that embodies all others as particular cases. The perturbed
configuration is an admissible variation from a admissible configuration. An ensemble of perturbed configurations
is used to establish incremental or rate equations.
The iterated and target configurations are introduced in the context of incremental-iterative solution procedures
for numerically tracing equilibrium paths. The target configuration is the “next solution”. More precisely, an
equilibrium solution (assumed to exist) which satisfies the residual equations for a certain value of the stage control
parameter λ. While working to reach the target, a typical solution process goes through a sequence of iterated
configurations that are not in equilibrium.
The corotated configuration is a rigid-body rotation of the reference configuration that “follows” the current
configuration like a “shadow”. It is used in the corotational (CR) kinematic description of nonlinear finite elements.
Strains measured with respect to the corotated configuration may be considered “small” in many applications, a
circumstance that allows linearization of several relations and efficient treatment of stability conditions.

§7.4. Kinematics
This section cover the essential kinematics necessary for finite displacement analysis.
§7.4.1. Deformation and Displacement Gradients
The derivatives of (x, y, z) with respect to (X, Y, Z ), arranged in Jacobian format, constitute the so-
called deformation gradient matrix:
 ∂x ∂x ∂x 
∂ X ∂Y ∂Z
∂(x, y, z)  ∂y ∂y ∂y 
F= = . (7.5)
∂(X, Y, Z )  ∂ X ∂Y ∂Z 
∂z ∂z ∂z
∂ X ∂Y ∂Z
7–8
7–9 §7.4 KINEMATICS

The inverse relation gives the derivatives of (X, Y, Z ) with respect to (x, y, z) as
 ∂X ∂X ∂X 
∂x ∂y ∂z
∂(X, Y, Z ) 


F−1 = =  ∂Y ∂Y ∂Y
. (7.6)
∂(x, y, z)  ∂x ∂y ∂z 
∂Z ∂Z ∂Z
∂x ∂y ∂z
These matrices can be used to relate the coordinate differentials
   
dx dX
dx = dy = F dY = F dX, dX = F−1 dx. (7.7)
dz dZ

Similarly, the displacement gradients with respect to the reference configuration can be presented as
the 3 × 3 matrix
 ∂x ∂x ∂ x   ∂u X ∂u X ∂u X 
∂X − 1 ∂Y ∂Z ∂X ∂Y ∂Z
 ∂y ∂y ∂y   
G=F−I=  =  ∂u Y ∂u Y ∂u Y  = ∇u. (7.8)
 ∂X ∂Y − 1 ∂Z   ∂X ∂Y ∂Z 
∂z ∂z ∂z ∂u Z ∂u Z ∂u Z
∂X ∂Y ∂Z − 1 ∂X ∂Y ∂Z
Displacement gradients with respect to the current configuration are given by
   ∂u X ∂u X ∂u X 
1 − ∂∂ X
x
∂X
∂y
∂Z
∂x ∂x ∂y ∂z
   
 ∂Y 1 − ∂Y ∂Y   ∂u Y ∂u Y ∂u Y 
J = I − F−1 =  ∂x ∂y ∂z  =  ∂x ∂y ∂z . (7.9)
   
∂Z ∂Z 1 − ∂∂zZ ∂u Z ∂u Z ∂u Z
∂x ∂y ∂x ∂y ∂z
For the treatment of the Total Lagrangian description it will found to be convenient to arrange the
displacement gradients of (7.8) as a 9-component vector (printed as row vector to save space):

gT = [ g1 g2 g3 g4 g5 g6 g7 g8 g9 ]
 ∂u Z  . (7.10)
= ∂u X ∂u Y ∂u Z ∂u X ∂u Y ∂u Z ∂u X ∂u Y
∂X ∂X ∂X ∂Y ∂Y ∂Y ∂Z ∂Z ∂Z

Remark 7.6. For arbitrary rigid-body motions (motions without deformations) FT F = FFT = I, that is, F is an
orthogonal matrix.

Remark 7.7. Displacement gradient matrices are connected by the relations

G = (I − J)−1 − I, J = I − (I + G)−1 . (7.11)

For small deformations G ≈ J−1 and J ≈ G−1 .

Remark 7.8. In nonlinear continuum mechanics, displacement gradients play an important role that is absent in
the infinitesimal theory. This is even more so in the Total-Lagrangian core-congruential formulation covered in
Chapters 8–11.

7–9
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–10

Remark 7.9. The ratio between infinitesimal volume elements d V = d x d y dz and d V0 = d X dY d Z in the current
and reference configuration appears in several continuum mechanics relations. Because of (7.7) this ratio may be
expressed as
dV ρ0
= = det F, (7.12)
d V0 ρ

where ρ and ρ0 denote the mass densities in the current and reference configuration, respectively. This equation
expresses the law of conservation of mass.

§7.4.2. Stretch and Rotation Tensors

Tensors F and G are the building blocks of various deformation measures used in nonlinear continuum
mechanics. The whole subject is dominated by the polar decomposition theorem: any particle defor-
mation can be expressed as a pure deformation followed by a rotation, or by a rotation followed by a
pure deformation. Mathematically this is written as multiplicative decompositions:

F = RU = VR. (7.13)

Here R is an orthogonal rotation tensor, whereas U and V are symmetric positive definite matrices
√ rotation, U = V = I.
called the right and left stretch tensors, respectively. If the deformation is a pure
Premultiplying (7.13) by FT = URT gives U2 = FT F and consequently√ U = FT F. Postmultiplying
(7.13) by FT = RT V gives V2 = FFT and consequently V = FFT . Upon taking the square roots, the
rotation is then computed as either R = FU−1 or R = V−1 F. Obviously U = RVRT and V = RT UR.
The combinations C R = FT F and C L = FFT are symmetric positive definite matrices that are called the
right and left Cauchy-Green strecht tensors, respectively. To get U and V as square roots it is necessary
to solve the eigensystem of C R and C L , respectively.
To convert a stretch tensor to a strain tensor one substracts I from it or takes its log, so as to have a
measure that vanishes for rigid motions. Either U − I or V − I represent proper strain measures. These
are difficult, however, to express analytically in terms of the displacement gradients because of the
intermediate eigenproblem. A more convenient strain measure is described next.

§7.4.3. Green-Lagrange Strain Measure

A convenient finite strain measure is the Green-Lagrange4 strain tensor. Its three-dimensional expression
in Cartesian coordinates is

 
  eX X eX Y eX Z
e= 1
2
FT F − I = 12 (G + GT ) + 12 GT G = eY X eY Y eY Z , (7.14)
eZ X eZ Y eZ Z

4 A more proper name would be Green-St.Venant strain tensor. In fact Lagrange never used it but his name appears because of
its strong connection to the Lagrangian kinematic description. Many authors call this measure simply the Green strain tensor.

7–10
7–11 §7.4 KINEMATICS

Identifying the components of FT F − I or 12 (G + GT ) + 12 GT G with the tensor conponents we get

 2  2  2 
∂u X ∂u X ∂u Y ∂u Z
eX X = + 12 + +
∂X ∂X ∂X ∂X
 2  2  2 
∂u Y ∂u X ∂u Y ∂u Z
eY Y = + 12 + +
∂Y ∂Y ∂Y ∂Y
 2  2  2 
∂u Z ∂u X ∂u Y ∂u Z
eZ Z = + 12 + +
∂Z ∂Z ∂Z ∂Z (7.15)
   
∂u Y ∂u Z ∂u X ∂u X ∂u Y ∂u Y ∂u Z ∂u Z
eY Z = 12 + + 12 + + = eZ Y ,
∂Z ∂Y ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z
   
∂u Z ∂u X 1 ∂u X ∂u X ∂u Y ∂u Y ∂u Z ∂u Z
eZ X = 21
+ +2 + + = eX Z ,
∂X ∂Z ∂Z ∂X ∂Z ∂X ∂Z ∂X
   
∂u X ∂u Y 1 ∂u X ∂u X ∂u Y ∂u Y ∂u Z ∂u Z
eX Y = 21
+ +2 + + = eY X .
∂Y ∂X ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y
If the nonlinear portion (that enclosed in square brackets) of these expressions is neglected, one obtains
the infinitesimal strains x x ,  yy , . . . zx = 12 γzx , x y = 12 γx y encountered in linear finite element
analysis. For future use in finite element work we shall arrange the components (7.15) as a 6-component
strain vector e constructed as follows:
     
e1 eX X eX X
 e2   eY Y   eY Y 
     
 e3   eZ Z   eZ Z 
e= = = . (7.16)
 e4   eY Z + e Z Y   2eY Z 
     
e5 eZ X + eX Z 2e Z X
e6 e X Y + eY X 2e X Y

Remark 7.10. Several other finite strain measures are used in nonlinear continuum mechanics. The common
characteristic of all measures is that they must predict zero strains for arbitrary rigid-body motions, and must
reduce to the infinitesimal strains if the nonlinear terms are neglected. This topic is further explored in Exercise
7.5.

§7.4.4. Strain-Gradient Matrix Expressions

For the development of the TL core-congruential formulation presented in following sections, it is useful
to have a compact matrix expression for the Green-Lagrange strain components of (7.16) in terms of
the displacement gradient vector (7.12). To that end, note that (7.15) may be rewritten as

e1 = g1 + 12 (g12 + g22 + g32 ),

e2 = g5 + 12 (g42 + g52 + g62 ),
e3 = g9 + 12 (g72 + g82 + g92 ),
(7.17)
e4 = g6 + g8 + g4 g7 + g5 g8 + g6 g9 ,
e5 = g3 + g7 + g1 g7 + g2 g8 + g3 g9 ,
e6 = g2 + g4 + g1 g4 + g2 g5 + g3 g6 .

7–11
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–12

These relations may be collectively embodied in the quadratic form

ei = hiT g + 12 gT Hi g, (7.18)
where hi are sparse 9 × 1 vectors:
           
1 0 0 0 0 0
 
0  
0  
0 0 0 1
           
0 0 0 0 1 0
           
0 0 0 0 0 1
           
h1 =  0  , h2 =  1  , h3 =  0  , h4 =  0  , h5 =  0  , h6 =  0  , (7.19)
           
0 0 0 1 0 0
           
0 0 0 0 1 0
           
0 0 0 1 0 0
0 0 1 0 0 0
and Hi are very sparse 9 × 9 symmetric matrices:
 
1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0
 
H1 =  0 0 0 0 0 0 0 0 0, etc. (7.20)
 
0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0

Remark 7.11. For strain measures other than Green-Lagrange’s, expressions similar to (7.14) may be constructed.
But although the hi remain the same, the Hi become complicated functions of the displacement gradients.

§7.4.5. Pull Forward and Pull Back

Most of the foregoing material is classical continuum mechanics as covered in dozens of scholarly
books. Next is a kinematic derivation scheme that is quintaessential FEM. Consider the motion of an
elastic bar element in the 2D plane as depicted in Figure 7.2.
(To be expanded, Chapter posted as is)
§7.5. Stress Measure
Associated with each finite strain measure is a corresponding stress measure that is conjugate to it
in the sense of virtual work. That corresponding to the Green-Lagrange strain is the second Piola-
Kirchhoff symmetric stress tensor, often abbreviated to “PK2 stress.” The three-dimensional component
expression of this tensor in Cartesian coordinates is
 
sX X sX Y sX Z
s = sY X sY Y sY Z , (7.21)
sZ X sZ Y sZ Z
in which s X Y = sY X , etc. As in the case of strains, for future use in finite element work it is convenient
to arrange the components (7.21) as a 6-component stress vector s:
s T = [ s1 s2 s3 s4 s5 s6 ] = [ s X X sY Y sZ Z sY Z sZ X sX Y ] . (7.22)

7–12
7–13 §7.6 CONSTITUTIVE EQUATIONS

~
Y ~ Deformed
X
Corotated _
θB y
Base
θ
Y
_
Globally X x
aligned

Figure 7.2. Rigid motion of bar in 2D illustrating concept of globally aligned configuration.

Remark 7.12. The physical meaning of the PK2 stresses is as follows: si j are stresses “pulled back” to the reference
configuration C 0 and referred to area elements there.

Remark 7.13. The PK2 stresses are related to the Cauchy (true) stresses σi j through the transformation
 ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X 
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
 ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y 
s    
 ∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂ x ∂ y  σX X
 
XX
 sY Y   ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z   σY Y 
 s Z Z  ρ0  ∂ x ∂ x ∂ y ∂ y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y   
     σZ Z 
 sY Z  = ρ  ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z   σY Z  , (7.23)
   ∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x  
∂ x ∂ y  σZ X 
sZ X 
 ∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X 
sX Y   σX Y
 ∂ x ∂ x ∂ y ∂ y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y 
∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
 ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x 
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
σ   ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y   
 
XX  ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y  s X X
 σY Y   ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z   sY Y 
 σZ Z  ρ  ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y  
 
 =   sZ Z  .
 σY Z  ρ0  ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z 
 
 YZ 
s
(7.24)
   ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y  s Z X 
σZ X  
σX Y  ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x  s X Y
 ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y 
∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
The density ratios that appears in these equations may be obtained from (7.13). If all displacement gradients are
small, both transformations reduce to the identity, and the PK2 and Cauchy stresses coalesce.

§7.6. Constitutive Equations

Throughout this course we restrict our attention to constitutive behavior in which conjugate strains and
stresses are linearly related. For the Green-Lagrange and PK2 measures used here, the stress-strain
relations will be written, with the summation convention implied,
si = si0 + E i j e j , (7.25)

7–13
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–14

where ei and si denote components of the strain and stress vectors defined by (7.16) and (7.22), respec-
tively, si0 are stresses in the reference configuration (also called prestresses) and E i j are constant elastic
moduli with E i j = E ji . In full matrix notation,
   0   
s1 s1 E 11 E 12 E 13 E 14 E 15 E 16 e1
 s2   s2   E 12
0
E 22 E 23 E 24 E 25 E 26   e2 
   0   
 s3   s3   E 13 E 23 E 33 E 34 E 35 E 36   e3 
 = 0+  , (7.26)
 s4   s4   E 14 E 24 E 34 E 44 E 45 E 46   e4 
   0   
s5 s5 E 15 E 25 E 35 E 45 E 55 E 56 e5
s6 s60 E 16 E 26 E 36 E 46 E 56 E 66 e6
or in compact form,
s = s0 + Ee. (7.27)

Remark 7.14. For an invariant reference configuration, PK2 and Cauchy (true) prestresses obviously coincide (see
Remark 7.14). Thus σ0 ≡ s0 in such a case. However if the reference configuration is allowed to vary often, as in
the UL description, things get more complicated.

§7.7. Strain Energy Density

We conclude this review by giving the expression of the strain energy density U in the current
configuration reckoned per unit volume of the reference configuration:

U = si0 ei + 12 (si − si0 )ei = si0 ei + 12 ei E i j e j , (7.28)

or, in matrix form

U = eT s0 + 12 eT Ee. (7.29)
If the current configuration coincides with the reference configuration, e = 0 and U = 0. It can be
observed that the strain energy density is quadratic in the Green-Lagrange strains. To obtain this density
in terms of displacement gradients, substitute (7.18) into (7.29) to get

 
U = si0 (hiT g + gT Hi g) + 1
2
(gT hi + 12 gT Hi g)E i j (hTj g + 12 gT H j g) . (7.30)

Because hi and Hi are constant, this relation shows that the strain energy density is quartic in the
displacement gradients collected in g.
The strain energy in the current configuration is obtained by integrating this energy density over the
reference configuration: 
U= U d X dY d Z . (7.31)
V0

This expression forms the basis for deriving finite elements based on the Total Lagrangian description.

7–14
7–15 Exercises

Homework Exercises for Chapter 7

Review of Continuum Mechanics

EXERCISE 7.1 [A:15] Obtain the expressions of H3 and H5 .

EXERCISE 7.2 [A:15] Derive (7.28) by integrating si dei from C0 (ei = 0) to C (ei = ei ) and using (7.25).

EXERCISE 7.3 [A:20] A bar of length L 0 originally along the X ≡ x axis (the reference configuration C0 ) is
rigidly rotated 90◦ to lie along the Y ≡ y axis while retaining the same length (the current configuration C). Node
1 at the origin X = Y = 0 stays at the same location.
(a) Verify that the motion from C0 to C is given by
x = −Y, y = X, z = Z. (E7.1)

(b) Obtain the displacement field u, the deformation gradient matrix F, the displacement gradient matrix G and
the Green-Lagrange axial strain e = e X X . Show that the Green-Lagrange measure correctly predicts zero
axial strain whereas the infinitesimal strain measure  =  X X = ∂u X /∂ X predicts the absurd value of −100%
strain.
EXERCISE 7.4 [A:20] Let L 0 and L denote the length of a bar element in the reference and current configurations,
respectively. The Green-Lagrange finite strain e = e X X , if constant over the bar, can be defined as
L 2 − L 20
e= . (E7.2)
2L 20
Show that the definitions (E7.2) and of e = e X X in (7.15) are equivalent. (Hint: express L 0 and L in terms of the
coordinates and displacements in the bar system.)
EXERCISE 7.5 [A:25] The Green-Lagrange strain measure is not the only finite strain measure used in structural
and solid mechanics. For the uniaxial case of a stretched bar that moves from a length L 0 in C0 to a length L in C,
some of the other measures are defined as follows:
(a) Uniaxial Almansi strain:
L 2 − L 20
eA = . (E7.3)
2L 2
(b) Uniaxial Hencky strain, also called logarithmic or “true” strain:
e H = log(L/L 0 ), (E7.4)
where log denotes the natural logarithm.
(c) Uniaxial midpoint strain5
L 2 − L 20
eM = . (E7.5)
2[(L + L 0 )/2]2
(d) Uniaxial engineering strain:
L − L0
eE =  = . (E7.6)
L0
If L = (1 + )L 0 , show by expanding e A , e H and e M in Taylor series in  (about  = 0) that these measures, as
well as the Green-Lagrange axial strain (E7.2), agree with each other to first order [i.e., they differ by O( 2 )] as
 → 0.

5 The midpoint strain tensor, which is a good approximation of the Hencky strain tensor but more easily computable, is frequently
used in finite element plasticity or viscoplasticity calculations that involve large deformations, for example in metal forming
processes.

7–15
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–16

EXERCISE 7.6 [A:30] (Advanced). Extend the definition of the Almansi, Hencky, midpoint and midpoint strains
to a three dimensional strain state. Hint: use the spectral decomposition of FT F and the concept of function of a
symmetric matrix.

EXERCISE 7.7 [A:35] (Advanced). Extend the definition of engineering strain to a three-dimensional strain
state. The resulting measures (there are actually two) are called the stretch tensors. Hint: use either the spectral
decomposition of FT F, or the polar decomposition theorem of tensor calculus.

EXERCISE 7.8 [A:40] (Advanced). Define the stress measures conjugate to the Almansi, Hencky, midpoint and
engineering strains.

7–16
 6
.

Conservative
Systems

6–1
Chapter 6: CONSERVATIVE SYSTEMS 6–2

TABLE OF CONTENTS

Page
§6.1. Introduction 6–3
§6.2. The Load Potential 6–3
§6.2.1. Concentrated Dead Loads . . . . . . . . . . . . . 6–4
§6.2.2. Distributed Dead Loads . . . . . . . . . . . . . 6–5
§6.3. The Internal Energy: A Linear Spring 6–6
§6.4. The Internal Energy: How Geometric Nonlinearities Arise 6–6
§6.5. Internal Energy: Additivity Property 6–8
§6. Exercises . . . . . . . . . . . . . . . . . . . . . . 6–9

6–2
6–3 §6.2 THE LOAD POTENTIAL

§6.1. Introduction

It was noted in previous Chapters that if a structure and the forces acting upon it collectively form a
conservative mechanical system, the residual force vector r may be expressed as the gradient of the
total potential energy  with respect to the state vector:

∂
r= . (6.1)
∂u

Furthermore, the decompositions  = U − P and r = p − f are related in the sense that

∂U ∂P
p= , f= . (6.2)
∂u ∂u
where p and f are the internal and external forces, respectively, U is the internal energy — which
reduces to the strain energy in the problems considered in this course — and P is the potential of the
applied loads, the negative of which is called the external work function W .
The force equilibrium equations r = 0 or f = p express the fact that the total potential energy is
stationary with respect to variations of the state vector when the structure is in static equilibrium.
Mathematically:
 
∂ T
δ = r δu =
T
δu = 0. (6.3)
∂u
where δu denotes a virtual displacement, δ being the variation symbol. Since δu is arbitrary, (6.3)
implies that r = 0.
If the structural system is conservative there are substantial advantages in taking advantage of that
property:
(1) If discrete force equilibrium equations are worked out by hand (either for complete structures
or finite elements) derivation from a potential is usually simpler than direct use of equilibrium,
because differentiation is a straightforward and less error prone operation, especially as regards
signs. Exercise 6.3 gives an example of this.
(2) The transformation of residual equations to different coordinate systems is simplified because of
the invariance properties of energy functions.
(3) The conventional finite element discretization method relies on the availability of an internal
energy functional.
(4) The tangent stiffness matrix is symmetric. Consequently equation solvers (and eigensolvers) can
take advantage of this property.
(5) Loss of stability can be assessed by the singular stiffness criterion, which is static in nature. If the
system is nonconservative, loss of stability may have to be tested by a dynamic criterion, which
is always more difficult and computationally expensive.
This Chapter introduces the concepts of internal and external potential for systems with finite degrees of
freedom. The presentation is not general in nature but relies on a few simple examples complemented
with exercises. The material is intended to serve as a “bridge” to the formulation of geometrically
nonlinear finite elements, which starts in the next Chapter.

6–3
Chapter 6: CONSERVATIVE SYSTEMS 6–4

Deformed

;
u

;;; ;
F

Undeformed
;;; No change in load
magnitude or direction

Figure 6.1. Structure under concentrated dead load F

§6.2. The Load Potential

The concept of load potential is the easiest to understand. This function, called P, is the potential of
the work done by the applied or prescribed forces working on the displacements of the points on which
those forces act. The negative of this potential W = −P is called the work function, but this function
will not be used in the present course.

§6.2.1. Concentrated Dead Loads

For a concrete example, consider a structure loaded by a single concentrated force F that does not
change in magnitude or direction as the structure displaces; see Figure 6.1). A force with these
properties is called a dead load.
If u is the deflection of the point of application of F in the direction of the force, then the work performed
is obviously Fu. Consequently,
P = Fu. (6.4)
If the structure is subjected to n loads Fk (k = 1, . . . n) and the corresponding deflections in the
direction of the forces are called u k , then

n
P= Fk u k . (6.5)
i=1

In general these forces will be defined by their three components along the axes x, y, z and are more
properly represented by vectors fk . For example, if at location k = 3 we have a force F3 acting in the
y-direction,
 
0
f3 = F3 . (6.6)
0
Likewise, the displacement of points of application of fk is denoted by vector uk . The vector general-
ization of (6.5) is then the sum of n inner products:

n
P= fkT uk . (6.7)
k=1

6–4
6–5 §6.2 THE LOAD POTENTIAL

f y (x)
Deformed A B
u y (x)

y
;
;;; ;
x

Undeformed
;;;
Figure 6.2. Structure under under line load f y (x)
(directed upward) over segment AB.

Finally, if all applied force components are collected in the external force vector f (augmented with
zero entries as necessary to be in one-to-one correspondence with the state vector u) then we have the
compact inner-product expression
P = fT u. (6.8)

§6.2.2. Distributed Dead Loads

For distributed forces invariant in magnitude and direction, a spatial integration process is necessary
to obtain P. These forces may include line loads, surface loads or volume loads (body forces).
For example, consider the structure of Figure 6.2, on which a dead line load f y (x) acts in the y direction
along segment AB of the x axis. Then
 xB
P= f y (x) u y (x) d x, (6.9)
xA

where u y (x) is the y-displacement component of points on segment (A,B). A similar technique can
be used for volume (body) forces as illustrated in Exercise 6.1.

6–5
Chapter 6: CONSERVATIVE SYSTEMS 6–6

; k δ

;
x
1 2

Figure 6.3. Linear spring of stiffness k deforming along its axis.

Remark 6.1. Substantial mathematical complications arise if some forces are functions of the displacements.
For example, in slender structures under aerodynamic pressure loads the change of direction of the forces as the
structure deflects may have to be considered in the stability analysis. These so-called “follower” forces, which
introduce force B.C. nonlinearities, are considered later in the course. Suffices to say here that no loads potential
P generally exist in such cases and the system is nonconservative.

§6.3. The Internal Energy: A Linear Spring

The internal energy, called U , is the recoverable mechanical work “stored” in the material of the
structure by virtue of its elastic deformation. When this work is expressed in terms of strains and
stresses, as in following Chapters, it is called the strain energy. Note that only flexible bodies can store
strain energy; a rigid body cannot.
We shall illustrate the internal energy concept here by considering the simplest of all structural elements
already encountered in linear finite element analysis: a linear spring of stiffness k, illustrated in
Figure 6.3. Generalization to more complicated structures and structural components will be made in
subsequent Chapters.

If the spring is undeformed, its internal energy U can be conventionally taken as zero (because an
energy function can be adjusted by an arbitrary constant without changing its gradients). Now let the
spring deform slowly (to avoid inertial effects) such that its two ends separate by a distance δ called
the elongation. The internal spring force f¯ for an intermediate elongation 0 ≤ δ̄ ≤ δ is f¯ = k δ̄. An
elementary result of mechanics is that the strain energy taken up by the spring in its deformed state is

 δ  δ
U= spring-force × d(elongation) = (k δ̄) d δ̄ = 12 kδ 2 . (6.10)
0 0

Suppose that the spring is fixed at end 1 and that end 2 can move only along the x axis, as in Figure
6.3. Call u the x displacement of end 2. Then δ = u − 0 = u and the strain energy is U = 12 ku 2 .
According to (6.2) the internal force, which in this case is just the spring axial force p, is the derivative
of U with respect to u:
∂U
p= = ku. (6.11)
∂u
This is linear in the displacement u so nothing has changed so far with respect to linear finite element
analysis.

6–6
6–7 §6.4 THE INTERNAL ENERGY: HOW GEOMETRIC NONLINEARITIES ARISE

Deformed

1(x1 ,y1) ux1 uy1

uy2

y k

x
ux2
Undeformed
2(x2 ,y2)

Figure 6.4. Linear spring of stiffness k displacing on the x, y plane.

§6.4. The Internal Energy: How Geometric Nonlinearities Arise

Now suppose that the spring can move arbitrarily on the plane x, y, as depicted in Figure 6.4. The
position of the deformed spring is completely defined by the four displacement components u x1 , u y1 ,
u x2 and u y2 , which we collect in the state vector

 
u x1
u 
u =  y1  . (6.12)
u x2
u y2

Let and d denote the spring lengths in the undeformed and deformed configurations, respectively.
The elongation δ is given by
 
δ = d − = ( x +
x ) + ( y +
y ) − 2x + 2y ,
2 2 (6.13)

where
x = u x2 − u x1 ,
y = u y2 − u y1 , x = x2 − x1 , y = y2 − y1 , in which x1 , y1 , x2 and y2 denote
the x, y coordinates of the end nodes of the undeformed spring. Consequently

U = 12 kδ 2 = 12 k( 2 + 2d − 2 d )
 (6.14)
= 12 k (2 2 + 2 x
x +
2x + 2 y
y +
2y − 2 ( x +
x )2 + ( y +
y )2 .

6–7
Chapter 6: CONSERVATIVE SYSTEMS 6–8

The components of the internal forces are

 ∂U 
∂u
 x1 
 ∂U 
∂U  ∂u 
 y1 
p= = . (6.15)
∂u  ∂U 
 ∂u x2 
 
∂U
∂u y2
The actual expressions of the components in (6.15), which are nonlinear functions of the displacements,
are worked out in Exercise 6.2.
The important points that emerge from this example are:

1. The internal forces are nonlinear functions of the displacements, al-

though the spring itself remains constitutively linear. This nonlinearity
comes in as a result of geometric effects, and is thus properly called
geometric nonlinearity.
2. The effect of geometric nonlinearities can be traced to the change in
direction of the spring. Because if the spring stretches along its original
axis the internal force remains linear in the displacements. This change
of direction is measured by rotations.

Even for this simple case the exact nonlinear equations are quite nasty, involving irrational functions
of the displacements. The second property, however, shows that approximations to the exact nonlinear
equations may be made when the change in direction is “small” in some sense. This feature is illustrated
in Exercise 6.3.

§6.5. Internal Energy: Additivity Property

If the structure consists of m linear springs, each of which absorbs an internal energy Uk , the total
internal energy is the sum of the individual spring energies:

U = U1 + U2 + . . . + Um . (6.16)

This additivity property is of course general because energies are scalar quantities. It applies to arbitrary
structures decomposed into structural components such as finite elements. Furthermore, (6.16) is not
affected by whether the structure is linear or nonlinear.
The last property explains why finite element equations should be derived from energy functions if
such functions exist. That is not, however, always possible.

6–8
6–9 Exercises

Homework Exercises for Chapter 6

Conservative Systems
Note: the use of a symbolic algebra package, such as Mathematica or MathCad, is recommended for Exercises
6.3 and 6.4 to avoid tedious algebra and generate plots quickly. (There could be a gain from hours to minutes).

EXERCISE 6.1 [A:15] A body of volume V and density ρ is in an uniform gravity field g acting along the
−z axis. The body displaces to another position defined by the small-displacement field u(x, y, z). Find the
expression of the load potential P as an integral over the body if the change in shape of the body is negligible.

EXERCISE 6.2 [A:20] Work out the expression of the internal forces for (6.15). Then extend this relation to the
three-dimensional case in which the ends of the spring move by u x1 , u y1 , u z1 , u x2 , u y2 , u z2 in the x, y, z space.

EXERCISE 6.3 [A+N/C:30] Consider the shallow arch model shown in Figure E6.1. This consists of two
identical linear springs of axial stiffness k pinned to each other and to unmoving pinned supports as shown. The
springs are assumed able to resist both tensile and compressive forces. The distance between the supports is 2L.
The undeformed springs form an angle α with the horizontal axis.
The central pin in loaded by a dead vertical force of magnitude f , positive downwards, which is parametrized
as f = λk L. Only symmetrical deformations of the arch are to be considered for this Exercise. Consequently
the system has just one degree of freedom which we take to be the displacement u under the load, also positive
downwards. The response of this system exhibits the snap-through behavior sketched in Figure E6.2.
(a) Show that the internal energy U and load potential P of the two-spring system are given by
 2
1 1
U = kL 2
− , P = f u, (E6.1)
cos α cos θ

where θ is the angle shown in Figure E6.1, which is linked to u by the relation tan θ + u/L = tan α.

(b) Derive the exact equilibrium equation

∂
= 0,
r (u, λ) = (E6.2)
∂u
in which  = U − P is the total potential energy, and λ = f /(k L) is the dimensionless state parameter.
For convenience rewrite this as
r (µ, λ) = 0, (E6.3)
in terms of the dimensionless state parameter
u
µ= . (E6.4)
L tan α

(c) Derive the exact equation for the limit load parameters

∂λ(µ) 
= 0. (E6.5)
∂µ µ=µ
L ,λ=λ L

(Hint: the exact equation in terms of the angular coordinate θ is cos3 θ L = cos α). Solve this trigonometric
equation1 for the limit-load parameters λ L1 and λ L2 and the dimensionless displacements µ L1 and µ L2 at
those points assuming that α = 30◦ .

1 Equation (E6.5) is equivalent to det K = 0 because for a one-DOF system det K = K = ∂λ/∂µ.

6–9
Chapter 6: CONSERVATIVE SYSTEMS 6–10

u k
k

;; ;;
θ
α

L L

Figure E6.1. Two-spring model of shallow arch.

L1

u or µ

L2

Figure E6.2. Snap-through response of shallow arch (sketch).

(d) If the arch initially is and remains sufficiently “shallow” throughout its snap-through behavior, we may make
the small-angle approximations,

cos α ≈ 1 − 12 α 2 , cos θ ≈ 1 − 12 θ 2 , sin α ≈ tan α ≈ α, sin θ ≈ tan θ ≈ θ. (E6.6)

Recast the energy, equilibrium equations, and limit load equations in terms of these approximations, obtaining
U as a quartic polynomial in θ , r as a cubic polynomial in θ, etc, then replace in terms of µ. As a check,
the residual equation in terms of λ and µ should be given by (4.16). Calculate the limit load parameters λ L1
and λ L2 , and the dimensionless displacements
√ µ L1 and µ L2 at those loads. Verify that these displacements
correspond to the angles θ L = ±α/ 3.
(e) Draw the control-state response curves r (µ, λ) = 0, derived using the exact nonlinear equations and those
from the small-angle approximations on the λ, µ plane (as in the sketch of Figure E6.2, going up to µ ≈ 2.5)
for α = 30◦ .

EXERCISE 6.4 [A+N:15] Derive the current stiffness parameter κ defined in Equation (5.8) for the approximate
(small-angle) model of the two-spring arch of Exercise 6.3. Plot the variation of κ(µ) as µ varies from 0 to µ L1
at the first limit point, with µ along the horizontal axis. Does κ vanish at the limit point?

6–10
 5
.

Critical Points
and Related
Properties

5–1
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–2

TABLE OF CONTENTS

Page
§5.1. Classification of Critical Points 5–3
§5.2. Limit Point Sensors 5–6
§5.3. *Turning Points 5–6
§5.4. *Derivatives of Energy Functions 5–6
§5.5. *Energy Increments 5–8
§5. Exercises . . . . . . . . . . . . . . . . . . . . . . 5–10

5–2
5–3 §5.1 CLASSIFICATION OF CRITICAL POINTS

This Chapter provides additional material on properties of the one-parameter force residual equa-
tions. It begins with a study of critical points, which are classified into limit and bifurcation points.
Limit point “sensors” and turning points are briefly described. The section concludes with some
mathematical derivations for conservative systems, which will be of use in Chapters dealing with
incremental solution methods.

§5.1. Classification of Critical Points

Throughout this Chapter it is assumed that the structural system is conservative1 and consequently
K is symmetric.
Response points at which K becomes singular are of great interest in the applications because of
their intimate connection to structural stability. These are called critical points, and also nonregular
or singular points. At these points the velocity vector v is not uniquely determined by q from (4.8).
Physically this means that the structural behavior cannot be controlled by the parameter λ.
It is convenient to distinguish the following types of critical points:
1. Isolated limit points, at which the tangent (4.17) to the equilibrium path is unique but normal
to the λ axis so v becomes infinitely large.
2. Multiple limit points, at which there the tangent lies in the null space of K and is not unique
but still normal to the λ axis.
3. Isolated bifurcation points, also called branch points or branching points, from which two
equilibrium path branches emanate and so there is no unique tangent. The rank deficiency of
K is one.
4. Multiple bifurcation points, from which more than two equilibrium path branches emanate.
The rank deficiency of K is two or greater.
A critical point that is both a limit and a bifurcation point is classified as a multiple bifurcation
point.
Figures 5.1, 5.2 and 5.3 illustrate isolated limit points (identified as L 1 , L 2 , . . .) and bifurcation
points (identified as B1 , B2 , . . .).
To classify critical points we proceed as follows. Let z be a null right eigenvector of K at a critical
point, that is,
Kz = 0. (5.1)
Since K is assumed symmetric, zT K = 0; that is, z is also a left null eigenvector. The parametric
differential equation of the equilibrium path is Ku̇ = q λ̇, which multiplied through by dt becomes

K du = q dλ. (5.2)

Premultiply both sides of (5.2) by zT and use zT K = 0 to get

zT q dλ = 0. (5.3)

1 A property elaborated upon in Chapter 6. This property is important in that a symmetric K is guaranteed to have a full
set of eigenvectors. Furthermore left and right eigenvectors coalesce.

5–3
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–4

λ L1
B

L2

u1

Figure 5.1. Limit points L 1 and L 2 and bifurcation point (B) for a two degree of freedom system
(u 1 , u 2 ) shown on the u 1 versus λ plane. Limit point (“snap through” behavior) occurs
before bifurcation. Full lines represent physically “preferred” paths.

Two cases may be considered. If

zT q = 0, (5.4)

then dλ must vanish, and we have a limit point. The point is isolated if z is the only null eigenvector
and multiple otherwise.
On the other hand, if
zT q = 0, (5.5)

then we have a bifurcation or branching point. The point is isolated if z is the only null eigenvector
and multiple otherwise. The key physical characteristic of a bifurcation point is an abrupt transition
from one deformation mode to another mode; the latter having been previously “concealed” by virtue
of being orthogonal to the incremental load vector.
Remark 5.1. If K is not symmetric, several changes must be made in the previous assumptions and derivations.
These are explained in the Chapters that deal with nonconservative systems (29-30) and the possible loss of
stability by growing dynamic oscillations (flutter).

Remark 5.2. If λ is an applied load multiplier, a limit point such as L 1 in Figure 5.1 is called a snap-through
point because the structure “snaps” dynamically to another equilibrium position. The term collapse applies to
critical points beyond which the structure becomes useless.

Remark 5.3. As an isolated limit point is approached, v tends to become parallel to z whereas its magnitude
goes to ∞; that is
v
→ z. (5.6)
|v|
Consequently the normalized v may be a good eigenvector estimate if K has been factored near the limit point.
(This is nothing more that a statement of the well known inverse iteration process for finding eigenvectors.)

5–4
5–5 §5.1 CLASSIFICATION OF CRITICAL POINTS

λ L1
B1

B2

L2

u1

Figure 5.2. Similar to Figure 5.1, but here the bifurcation point B1 implying “buckling”
behavior occurs before the limit point L 1 , which is physically unreachable.
A more realistic three-dimensional view of this case is shown in Figure 5.3.

λ L1
B1

B2
L2
u2 u1

Figure 5.3. The equilibrium path of Figure 5.2 shown in the 3D space (u 1 , u 2 , λ).
This is the type of response exhibited by a uniformly pressurized deep
arch, for which u 1 and u 2 are amplitudes of the symmetric and antisymmetric
deformation mode, respectively, and λ is a pressure multiplier.

Remark 5.4. The set of control parameters for which det K = 0 while r(u, λ) = 0 is sometimes called the
bifurcation set in the mathematical literature. The name is misleading, however, in that the set may include
limit points; the name critical set would be more appropriate.

Remark 5.5. Showing bifurcation points on the λ versus u plane as in Figures 5.1 and 5.2 may be misleading, as
it conceals the phenomenon of transition from one mode of deformation to another. A more realistic picture is
provided in Figure 5.3, which shows the classical bifurcation behavior for a symmetrically loaded shallow arch;
here u 1 and u 2 measure amplitude of symmetric and antisymmetric displacement shapes, respectively. At B1
the arch, which had been previously deforming symmetrically, takes off along an antisymmetric deformation
mode; at B2 the latter disappears and the arch rejoin the symmetric path.

5–5
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–6

Remark 5.6. Physically the distinction between the two types of critical points is not so marked, inasmuch as
imperfect structures display limit-point behavior. A bifurcation point may be viewed as the limit of a sequence
of progressively sharper limit points realized as the structure strives towards mathematical perfection.

Remark 5.7. For a readable introduction to elastic structural stability along the lines of classical perturbation
theory, the monographs by Thompson and Hunt [181,183] are still unsurpassed. The treatise by Knops and
Wilkes [100] goes deeper into various mathematical questions, but is computationally useless. Brush and
Almroth [37] give more information on computational methods. The survey and book by Bushnell [40] has
more physics (e.g., temperature, plasticity and creep effects) and a wider selection of practical problems. The
connection between potential-based structural stability and modern catastrophe theory is presented in a highly
readable manner by Poston and Steward [141] and Thompson [182].

§5.2. Limit Point Sensors

Scalar estimates of the overall stiffness of the structure as the control parameter varies are useful
as limit points sensors. The following estimator is based on the Rayleigh quotient approximation
to the fundamental eigenvalue of K:
xT Kx
kx = T , (5.7)
x x
where x is an arbitrary nonnull vector, and K is evaluated at an equilibrium position u(λ). An
“equilibrium-path stiffness” estimator is obtained by taking x to be v = K−1 q, in which case
qT v
k = kv = . (5.8)
vT v
This value of course depends on λ. It is convenient in practice to work with the dimensionless ratio

κ = k(λ)/k(0), (5.9)

which takes the value 1 at the start of an analysis stage, and goes to zero as a limit point is approached.
A stiffness estimator with this behavior (although computed in a different way) was introduced by
Bergan and coworkers under the name current stiffness parameter. It should be noted, however,
that no estimator of this type can reliably predict the occurrence of a bifurcation point. Sensors for
such points are described later in the context of augmented equations.
§5.3. *Turning Points
Turning points are regular points at which the tangent is parallel to the λ axis so that v = 0. The unit tangent
takes the form  
0
tu = (5.10)
±1
Although these points generally do not have physical meaning, they can cause special problems in path-
following solution procedures because of “turnback” effects.
To detect the vicinity of a turning point one can check the two mathematical conditions: v becomes orthogonal
to q and u tends to zero faster than q. For example:

| cos(v, q)| < δ |κ| > κmin , (5.11)

where κ is the current stiffness parameter. Typical values may be δ = 0.01, κmin = 100.

5–6
5–7 §5.4 *DERIVATIVES OF ENERGY FUNCTIONS

§5.4. *Derivatives of Energy Functions

If the residual r(u, λ) is derivable from a total potential energy function (u, λ) as in (3.2), then the stiff-
ness matrix and incremental load vector appear naturally as components of the following matrix of second
derivatives:
 ∂ 2 ∂ 2 
 ∂u∂u ∂u ∂λ   K −q

  2 T  = (5.12)
∂  ∂ 2 −qT a
∂λ∂u ∂λ∂λ
where a = ∂ 2 /∂λ2 has not been introduced previously. Obviously the tangent stiffness matrix K (the Hessian
of ) is now symmetric. Note also that

∂q ∂ 3 ∂ ∂ 2 ∂K
= = = = Kλ , (5.13)
∂u ∂u ∂λ∂u ∂λ ∂u∂u ∂λ

is a symmetric matrix.
The complementary energy function ∗ may be defined from the dual Legendre transformation (see e.g.,
Chapter 2.5 of Sewell’s book [163]) as

∂
 + ∗ = u i = uT r = rT u. (5.14)
∂u i

This gives ∗ (r, λ) = rT u −  with u eliminated from r(u, λ) = 0, so now the residual forces are the active
variables. Obviously
∂∗ ∂∗
u= , or u i = . (5.15)
∂r ∂ri
The matrix of second derivatives of ∗ is
 ∂ 2 ∗ ∂ 2 ∗ 
∂r∂r ∂r∂λ   
 F v
 T 2 ∗ 
= T (5.16)
∂ 2 ∗ ∂  v b
∂λ∂r ∂λ∂λ

These are linked to the quantities that appear in (5.12) by the matrix relations

F = K−1 , v = K−1 q = Fq, b = qT K−1 q − a. (5.17)

The converse relations are

K = F−1 , q = Kv, a = vT Kv − b. (5.18)

The tangent flexibility matrix F = K−1 (the Hessian of K ) is now symmetric. Note also that

∂u ∂ 3 ∗ ∂ ∂ 2 ∗ ∂F
= = = = Fλ , (5.19)
∂r ∂r ∂λ∂r ∂λ ∂r∂r ∂λ

is a symmetric matrix.

5–7
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–8

Remark 5.8. The following matrix appears (as amplification matrix) in the study of the stability of incremental methods:

∂v ∂(Fq) ∂F ∂q ∂F ∂K
A= = = q+F = Kq + F . (5.20)
∂u ∂u ∂u ∂u ∂r ∂λ
Although A is unsymmetric, under some general conditions it has real eigenvalues. To show that we express A as the product
of two symmetric matrices:
∂v ∂v ∂r ∂F
A= = = K = Fλ K, (5.21)
∂u ∂r ∂u ∂λ
where the relation (5.19) has been used. If Fλ is nonsingular, the eigensystem Axi = µi xi can be transformed to the
generalized symmetric eigenproblem
Kxi = µi F−1λ xi . (5.22)
If K is positive definite this system has nonzero real roots µi . If Fλ is singular but K positive definite, consideration of the
alternative eigensystem
Fλ yi = µi K−1 yi = µi Fyi , (5.23)
shows that such a singularity contributes only zero roots.

Remark 5.9. Another quantity that appears in the analysis of incremental methods is the vector

∂v ∂v ∂u
v = = = Av = Fλ Kv = Fλ q. (5.24)
∂λ ∂u ∂λ

Remark 5.10. Two other Legendre transforms may be constructed: X (δ, u) and Y (δ, r), in which δ = ∂/∂λ (a generalized
displacement if λ is a load multiplier) is the active variable and either u or r take the role of passive variables. X and Y
together with  and K form a closed chain of Legendre transformations. The functions X and Y are, however, of limited
interest in the present context.

§5.5. *Energy Increments

In this section we continue to assume that r is derivable from the potential  = U − W . For questions such
as positive path traversal it is interesting to obtain an expression of the energy increment on passing from an
equilibrium position (u, λ) to a neighboring configuration (u + u, λ + λ) on the equilibrium path:

 = (u + u, λ + λ) − (u, λ). (5.25)

First we note that adding an arbitrary function of λ to 

 + F(λ), (5.26)

does not change the equilibrium equations or rate forms. To second order in the increments we get

 = rT u + A λ + 12 uT K u − qT u λ + 12 a(λ)2 , (5.27)

with
∂ ∂ 2
A=, a= , (5.28)
∂λ ∂λ2
evaluated at (u, λ). But we can always adjust F(λ) in (5.25) so that A = a = 0. Furthermore at an equilibrium
position r = 0, and along the equilibrium path u = K−1 q λ = uT λ. Substituting we find for the energy
increment
 = U − W = 12 qT u(λ)2 − qT u(λ)2 = − 12 qT u(λ)2 . (5.29)
This formula displays the important function of the product qT u in the energy increment. By extension we
may call
W = qT u(λ)2 (5.30)

5–8
5–9 §5.5 *ENERGY INCREMENTS

the external work increment even if r does not derive from a potential.
To fix the ideas assume that r derives from a quadratic potential

 = 12 uT Ku − qT uλ + Cλ + D, (5.31)

where C and D are arbitrary constants. Then the increment  from an equilibrium position (u, λ) that
satisfies the linear relation Ku = qT λ, to an arbitrary configuration (u + u, λ + λ) is

 = uT (Ku − qT λ) + λ(qT u − C) = −λ(qT u − C) = −(qT vλ − C)λ. (5.32)

Since C is arbitrary, chose it so that ∂/∂λ = −qT u + C = 0. Then

 = −qT v( 12 λ2 ). (5.33)

5–9
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–10

Homework Exercise for Chapter 5

Critical Points and Related Properties

EXERCISE 5.1 Given the one-parameter, two-degree-of-freedom residual-force system

   
r 6u 1 − 2u 2 − u 21 − 12λ
r(u 1 , u 2 , λ) = 1 = (E5.1)
r2 −2u 1 + 4u 2 − u 22 + 2λ
Consider the point P(u 1 , u 2 , λ) located at

u 1 = 2, u 2 = 1, λ = 12 , (E5.2)

(a) Show that P is on an equilibrium path,

(b) Show that P is a critical point,
(c) Determine whether it is a limit or a bifurcation point. [Compute the null eigenvector z of K at that point].
(d) Verify whether the limit point sensor κ is zero at P.

EXERCISE 5.2 Show that all critical points of (E5.1) satisfy either of the equations

63 − u 1 − 36u 2 = 0, 5 − 2u 1 − 3u 2 + u 1 u 2 = 0 (E5.3)

called critical point surfaces, and that the only intersection of these surfaces and the equilibrium path is at
(E5.2).

EXERCISE 5.3 Show that the critical point surface defined by det(K) = 0 is independent of λ if the residual
force system is separable.

EXERCISE 5.4 Show that qT z is independent of λ if the residual force system is separable and the load is
proportional.

EXERCISE 5.5 (Advanced, requires knowledge of matrix eigensystem theory). If K is not symmetric, the
critical point classification argument based on qT z fails. Explain why.

5–10
 4
.

One-Parameter
Residual Equations

4–1
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–2

TABLE OF CONTENTS

Page
§4.1. Introduction 4–3
§4.2. Rate Forms and Incremental Velocity 4–3
§4.3. Separable Residuals and Proportional Loading 4–3
§4.4. Response Visualization by Incremental Flow 4–4
§4.4.1. Diagrams for One Degree of Freedom . . . . . . . . . 4–4
§4.4.2. Diagrams for Multiple Degrees of Freedom . . . . . . . 4–7
§4.5. Intrinsic Geometry of Incremental Flow 4–8
§4.5.1. Tangent Vector . . . . . . . . . . . . . . . . . 4–8
§4.5.2. Normal Hyperplane and Flow-Orthogonal Envelope . . . . 4–8
§4.5.3. ArcLength Distance . . . . . . . . . . . . . . . 4–8
§4.6. *State Vector Scaling 4–11
§4. Exercises . . . . . . . . . . . . . . . . . . . . . . 4–13

4–2
4–3 §4.3 SEPARABLE RESIDUALS AND PROPORTIONAL LOADING

§4.1. Introduction

This Chapter continues on the topic of residual equations introduced in Chapter 3. The general
residual force equation presented there is specialized, through the concept of staging introduced in
§3.4, to the one-parameter form in which r is a function of u (the state) and λ (the control). Together
these form the control-state space. The separable case in which u and λ can be segregated to both
sides of the residual equations, is described.
Further insight into the structural response may be achieved with the help of constant-residual
incremental flows. Paths and orthogonal hypersurfaces are introduced and interpreted geometrically.
Finally, the concepts of arclength and scaling are discussed.

§4.2. Rate Forms and Incremental Velocity

In this section we study further the one-parameter residual equation (3.17), reproduced below for
convenience:
r(u, λ) = 0. (4.1)
The corresponding residual-derivative equations are

ṙ = Ku̇ − qλ̇, (4.2)

r̈ = Kü + K̇u̇ − qλ̈ − q̇λ̇, (4.3)

∂r ∂r
K= , q=− (4.4)
∂u ∂λ
where K is the tangent stiffness matrix introduced in §3.3, and q is the incremental load vector. The
latter is the specialization of the control matrix Q defined in §3.3, to the one-parameter case. These
equations will be used in the sequel instead of the more general (3.14)–(3.15) unless otherwise
noted.
Rate forms of r(u, λ) = 0 are obtained by equating the above derivatives to zero:

ṙ = 0, or Ku̇ = qλ̇, (4.5)

r̈ = 0, or Kü + K̇u = qλ̈ + q̇λ̇. (4.6)

At regular points of the (u,λ) space the tangent stiffness K is nonsingular. If so, we can solve the
first-order rate form (4.5) for u̇:
∂u
u̇ = K−1 qλ̇ = vλ̇, or = u = v, (4.7)
∂λ
where
v = K−1 q. (4.8)
This vector is called the incremental velocity vector and is an important component of all solution
methods based on continuation.

4–3
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–4

§4.3. Separable Residuals and Proportional Loading

The force-balance equivalent of (3.3) for a one-parameter residual equation is
p(u) = f(u, λ). (4.9)
If the right hand side, which represents the external force vector, does not depend on the state
parameters u, that is
p(u) = f(λ), (4.10)
the system of equations (4.1) or (4.9) is called separable. Furthermore, if f is linear in λ the loading
is said to be proportional. Obviously q = ∂f/∂λ is then a constant vector.
Remark 4.1. The more general system (3.3) containing multiple control parameters is said to be separable if

p(u) = f(Λ). (4.11)

In this case the loading is called proportional if f is linear in all control parameters, thus giving a constant
control matrix Q.

Remark 4.2. If a separable system derives from a total potential energy = U − P, then the external work
potential P must be linear in the state parameters u i . Furthermore for the loading to be proportional, P must
also be linear in λ.

§4.4. Response Visualization by Incremental Flow

§4.4.1. Diagrams for One Degree of Freedom

As discussed in Chapter 2, the solution of the one-parameter residual form
r(u, λ) = 0, (4.12)
is often plotted on the u versus λ plane, where u is a representative component of u.
One such diagram is illustrated in Figure 4.1. If λ is a load amplitude, this is called a load-
displacement response curve or simply a response curve. It is common practice to make the
curve pass through the origin λ = 0, u = 0. More general terms for this geometrization are
equilibrium path or equilibrium trajectory. The path passing through the origin is called the
primary or fundamental path because it usually represents the operation of the structure under
normal service conditions.
A path can, of course, be traversed in two directions. These are identified as positive or + sense,
and negative or − sense. As illustrated in Figure 4.2, we shall use the convention that the positive
sense is associated with increasing values of the pseudo-time t when the path is parametrically
described as u = u(t) and λ = λ(t).
A diagram such as that in Figure 4.1 gives of course only a partial picture of the structural behavior
unless there is only a single degree of freedom. For a better understanding of the way numerical
solution procedures work (or fail to) it is instructive to “look around” the equilibrium path by
considering the perturbed residual equation

r(u, λ) = rc , (4.13)

4–4
4–5 §4.4 RESPONSE VISUALIZATION BY INCREMENTAL FLOW

Response curve r = 0

Figure 4.1. Typical response diagram showing primary equilibrium path.

λ +

− Pseudo-time t increasing

Pseudo-time t decreasing

Figure 4.2. Positive and negative traversal senses on a path.

where rc is a constant vector. This is the general solution of ṙ = 0. Additional information can
be conveyed by drawing the solutions of (4.13) for various values of the right-hand side near zero.
This produces constant-residual paths as illustrated in Figure 4.3. Collectively these paths form the
incremental flow whose differential equation is either ṙ = 0, or, if we take λ ≡ t :
∂r
r = = 0, (4.14)
∂λ
where primes denote derivatives with respect to λ. This can also be presented as
∂u ∂r
r = K + = Ku − q = 0. (4.15)
∂λ ∂λ

4–5
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–6

P(u, λ ) t+

positive tangent at P

λ
r =0

Figure 4.3. The incremental flow field as a family of constant-residual trajectories.

normal "hyperplane" at P
P(u, λ ) t+

positive tangent at P

λ
r =0

Figure 4.4. Incremental flow (full curves) and the flow-orthogonal envelope
(dashed curves). This envelope reduces here to a family of curves
because there is only one degree of freedom u.

4–6
4–7 §4.4 RESPONSE VISUALIZATION BY INCREMENTAL FLOW

If K is nonsingular, solving (4.15) yields u = K−1 q = v. The incremental solution methods

covered later exploit these forms, which explains the qualifier “incremental” applied to the flow.
Figure 4.3 also illustrates the construction of the tangent vector t+ at an arbitrary point P(u, λ).
This procedure is described more precisely in §4.5.
Figure 4.4 depicts a set of curves whose trajectories are orthogonal to the incremental flow. This
set is called the flow-orthogonal envelope. It will be explained later in §4.4 that this set generally
consists of a family of hypersurfaces. For a system with one degree of freedom, however, the
envelope reduces to a family of curves, as in Figure 4.4. This concept will be useful later in
explaining how incremental-iterative solution methods work.

Example 4.1. For simple one-degree of freedom systems it is easy to plot the incremental flow using standard
graphic packages. As an example consider the following residual equation, which is obtained as solution of
one of the Exercises of Chapter 6:

r (µ, λ) = α 3 µ(1 − µ)(2 − µ) − λ. (4.16)

Here µ is a dimensionless state parameter and α an angle in radians characterizing the reference position of
the structure. The following Mathematica program produces the incremental flow plot for α = 30◦ using the
ContourPlot function:

alpha = Pi/6; r = alpha^3*mu*(1-mu)*(2-mu)-lambda;

ContourPlot[r,{mu,0,2},{lambda,-.1,.1},PlotPoints->30];

0.1

0.05

-0.05

-0.1
0 0.5 1 1.5 2

Figure 4.5. Incremental flow plot for the residual (4.16) produced
by Mathematica via its ContourPlot function.

Examination of Figure 4.5 shows that the r = const curves are simply translations of each other along the λ
axis because λ appears simply as −λ. This is typical of proportional loading situations.

§4.4.2. Diagrams for Multiple Degrees of Freedom

If the number of degrees of freedom increases to N > 1 the incremental flow still remains a family
of curves in the N + 1-dimensional control-state space space (u, λ). Visualization, however, is
restricted to N = 2 as illustrated in Figure 4.6. For three or more degrees of freedom, only cross
sections of the control-state space can be displayed, in which one or two representative degrees of
freedom or functions of such are plotted. This “projection” requires some ingenuity and experience.

4–7
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–8

The flow-orthogonal envelope becomes a family of ordinary surfaces if N = 2, as illustrated in

Figure 4.7. For three or more degrees of freedom, the envelope becomes a family of hypersurfaces.
§4.5. Intrinsic Geometry of Incremental Flow

§4.5.1. Tangent Vector

At a generic regular point P of coordinates (u, λ), not necessarily on the equilibrium path, we can
construct an unnormalized tangent vector t defined by
   
u v
t= = , (4.17)
λ 1

where v = K−1 q is the incremental velocity vector (4.8). Tangent vectors are illustrated in Figures
4.3 and 4.8 for one and two degrees of freedom, respectively.
The tangent vector normalized to unit length is
 
v/ f
tu = , (4.18)
1/ f
where f is the scaling factor
 
f = |t| = + ||t||2 = + 1 + vT v. (4.19)
The positive tangent direction and the positive unit tangent are defined as
   
+ v + t+ v/ f
t =± , tu = =± . (4.20)
1 f 1/ f
The positive tangent direction points in the positive sense of path traversal, as defined in §4.2 and
Figure 4.2.
§4.5.2. Normal Hyperplane and Flow-Orthogonal Envelope
The hyperplane N P normal to t at P(u, λ) has the equation
vT u + λ = 0, (4.21)
where u = u − u P and λ = λ − λ P are increments from P. Dividing these increments by t
and passing to the limit one obtains

vT u̇ + λ̇ = 0. (4.22)
For a one degree of freedom u the hyperplane reduce to a line in (u, λ) space, as illustrated in
Figure 4.4. For two degrees of freedom the normal hyperplane is an ordinary plane in the 3D space
(u 1 , u 2 , λ), as illustrated in Figure 4.8.
For one degree of freedom (4.22) is the differential equation of a flow orthogonal to the incremental
flow, as illustrated in Figure 4.3; this flow is the envelope of the normals. For two degrees of
freedom (4.22) represents a family of surfaces, see Figure 4.6. For more degrees of freedom (4.22)
is a family of hypersurfaces. The orthogonality property plays an important role in corrective
solution methods.

4–8
4–9 §4.5 INTRINSIC GEOMETRY OF INCREMENTAL FLOW

r =0

u2

u1

Figure 4.6. An incremental flow response diagram for two degrees of freedom.
The plane paths of Figure 4.3 now become space curves.
Only a few paths are shown to reduce clutter.

r =0

u2

u1

Figure 4.7. A response diagram for two degrees of freedom, showing some
members of the flow-orthogonal envelope. Only the primary
equilibrium path r = 0 is shown to reduce clutter.

4–9
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–10

Normal hyperplane
v T u̇ + λ̇ = 0

P t+
λ
+ sense of increasing t

P̄

+
r =0

t+

u2

u1

Figure 4.8. Illustrating the tangent vector and normal hyperplane in

an incremental flow diagram for two degrees of freedom.
Point P is on the primary equilibrium path but P̄ is generic.

§4.5.3. ArcLength Distance

Theleft hand side of the hyperplane equation (4.21) normalized on dividing through by f
1 T
s = (v u + λ), (4.23)
f
acquires the following geometric meaning: s is the signed distance from the normal hyperplane at
P to a point Q( u, λ). For small increments ( u, λ), s may be considered as an approximation
to the arclength s of the path that passes through P because
1 T
ds = (v du + dλ). (4.24)
f
This important concept is illustrated in Figure 4.9.
Remark 4.3. At isolated limit points studied in Chapter 5, the normalization process (4.23) reduces the unit
tangent to    
z ±z
tu = , t+
u = , (4.25)
0 0
where z is the unit length null eigenvector of K, that is, Kz = 0. The sign ambiguity arises because +z and −z
are both eigenvectors; one of them has to be chosen to satisfy the positive-traversal convention. At bifurcation
points and non isolated limit points t is not unique.

4–10
4–11 §4.6 *STATE VECTOR SCALING

Normal hyperplane at P

s
Q
P
λ
Positive tangent direction

Path

u2

u1

Figure 4.9. The concept of arclength distance s from point P to point Q.

Note that the point order is important: the arclength distance
from Q to P is not generally the same as that from P to Q.

Remark 4.4. From (4.19) and (4.24) we note the formulas

du v dλ 1
= , = . (4.26)
ds f ds f

Remark 4.5. In the mathematical literature the incremental flow projected on the u state space is sometimes
called a Davidenko flow in honor of the father of continuation methods, should λ be interpreted as a continuation
parameter.

Remark 4.6. An alternative to plotting (4.13) for response visualization, is to consider the use of the constant-
residual-norm equation
||r(u, λ)|| = C, (4.27)

where ||r|| denotes a vector norm such as, for instance, the Euclidean norm ||r||2 = rT r, and C is a nonnegative
numeric constant. This relation does not generally represent a family of curves but a family of tube-like
hypersurfaces that for sufficiently small C “wrap around” equilibrium paths, as illustrated in Figure 4.10.
Because of the visual clutter evident in that figure, equation (4.27) is less suitable than (4.13) to study what
happens in the neighboorhood of equilibrium paths.

4–11
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–12

||r|| = C2 > C1

||r|| = C3 > C 2 ||r|| = C 1

r =0

u2

u1

Figure 410. For Remark 4.6: illustrating that the constant residual-norm
equation ||r|| = const generally represents a family of
tube-like surfaces “wrapping around” the equilibrium paths.

§4.6. *State Vector Scaling

In applying nonlinear equation solving techniques to structural mechanics (or, in general, to problems in
engineering and physics) the issue of scaling often arises because of two aspects:
1. The residual r has two types of arguments: u and λ. Translational degrees of freedom collected in the
state vector u have physical dimensions of length (displacement) whereas λ is dimensionless.
2. The degrees of freedom in u may have heterogeneous physical dimensions. For example, in the analysis
of finite element models that account for bending effects u may contain both translations and rotations.
To reduce the sensitivity of solution procedures to these factors, it is often advisable to introduce a scaling of
the state vector u to render it dimensionless and thus placed on an equal footing with λ:

u = Su. (4.28)
Here the scaling matrix S is diagonal, and a superposed tilde identifies a scaled quantity. If all entries of u
have homogeneous dimensions, one may take simply S = (1/u) I, where the scalar u has the dimension of u.
The scaled versions of other quantities defined previously are

u = S u, q = S−1 q,
 
K = S−1 KS−1 , (4.29)
     
 Sv 
v 

v = Sv, t= = , f = 1 + v S v = 1 +
T 2 vT
v, (4.30)
1 1
 

v
tu = (1/ 
 f) , 
s = ( u + λ)/ 
vT  f = (vT S2 u + λ)/ 
f. (4.31)
1

4–12
4–13 Exercises

Homework Exercises for Chapter 4

One-Parameter Residual Equations

EXERCISE 4.1 [A:5+15] Consider the residual force equations

     
r1 u 1 + 3u 22 − 21 0
= = . (E4.1)
r2 u 2 + 6u 1 u 2 − 2 0

(a) Is this system of equations separable in the sense discussed in §4.3?

(b) If so, can f and p be expressed as gradients of scalar functions U and P and what are these?

EXERCISE 4.2 [A:15+15] Suppose that (E4.1) is to be solved in two stages:

Stage 1. Start from 1 = 2 = 0 and go to 1 = 0 and 2 = 5. Parameter λ varies from 0 to 1.

Stage 2. Start from 1 = 0, 2 = 5 and go to 1 = 2 = 10. Again λ varies from 0 to 1.
(a) Express the residual in the one-parameter form (4.1) for each stage.
(b) Find the expression of the incremental load vector q in each stage. Is the loading proportional?

EXERCISE 4.3 [A:20] Suppose the first residual force above is replaced by r1 = u 1 + 3u 22 − 221 .
(a) Is the system still separable?
(b) For the same two stages of the previous exercise, is the loading proportional?

EXERCISE 4.4 [A:25] For stage 1 of Exercise 4.2, write down the analytical expressions of the incremental
velocity, the tangent vectors t and tu , the normal hyperplane equation, and the differential equations of the
flow-orthogonal envelope. Note: explicit inversion of K−1 may be done using the formulas to invert a 2 × 2
matrix.

EXERCISE 4.5 [A:25] Verify the assertion of Remark 4.6 by using the Euclidean norm ||r|| = rT r of the
residual vector.

EXERCISE 4.6 [A:25] Explain whether the unnormalized tangent vector t introduced in §4.5.1 may be defined
as  
u̇
t= , (E4.2)
λ̇
and whether this definition is more general than (4.17).

4–13
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–10

Table 2.1 Engineering Applications of Nonlinear Structural Analysis

Application Explanation

Strength analysis How much load can the structure support before
global failure occurs?
Deflection analysis When deflection control is of primary importance

Stability analysis Finding critical points (limit points or bifurcation

points) closest to operational range
Service configuration analysis Finding the “operational” equilibrium form of certain
slender structures when the fabrication and service
configurations are quite different (e.g. cables, inflat-
able structures, helicoids)
Reserve strength analysis Finding the load carrying capacity beyond critical
points to assess safety under abnormal conditions.
Progressive failure analysis A variant of stability and strength analysis in which
progressive deterioration (e.g. cracking) is consid-
ered.
Envelope analysis A combination of previous analyses in which multiple
parameters are varied and the strength information
thus obtained is condensed into failure envelopes.

three types of nonlinearity, with emphasis on the geometric one.6

The four sources are discussed in more detail in following sections. To remember where the nonlin-
ear terms appear in the governing equations, it is useful to recall the fields that continuum mechanics
deals with, and the relationships among these fields. For linear solid continuum mechanics infor-
mation is presented in Figures 2.7 and 2.8.7
In linear solid mechanics or linear structural mechanics the connecting relationships shown in Figure
2.8 are linear, and so are the governing equations obtained by eliminating all fields but one.
Any of these relations, however, may be nonlinear. Tracing this fact back to physics gives rise to
the types of nonlinearities depicted in Figure 2.9. Relations between body force and stress (the
equilibrium equations) and between strains and displacements (the kinematic equations) are closely

6 the exclusion of constitutive or material nonlinearities does not imply that there are less important than the others. Quite
the contrary. But the topic is covered in separate courses offered in Civil Engineering departments.
7 These are the Tonti diagrams introduced in the IFEM course.

2–10
 2
.

A Tour of
Nonlinear
Analysis

2–1
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–2

TABLE OF CONTENTS

Page
§2.1. Introduction 2–3
§2.2. Equilibrium Path and Response Diagrams 2–3
§2.2.1. Load-deflection response . . . . . . . . . . . . . 2–3
§2.2.2. Terminology . . . . . . . . . . . . . . . . . 2–3
§2.3. Special Equilibrium Points 2–4
§2.3.1. Critical points . . . . . . . . . . . . . . . . . 2–4
§2.3.2. Turning points . . . . . . . . . . . . . . . . 2–5
§2.3.3. Failure points . . . . . . . . . . . . . . . . . 2–5
§2.4. Linear Response 2–5
§2.5. Tangent Stiffness and Stability 2–6
§2.6. Generalized Response 2–7
§2.7. Response Flavors 2–8
§2.8. Engineering Applications 2–9
§2.9. Sources of Nonlinearities 2–9
§2.10. Geometric Nonlinearity 2–11
§2.11. Material Nonlinearity 2–12
§2.12. Force BC Nonlinearity 2–13
§2.13. Displacement BC Nonlinearity 2–14
§2. Exercises . . . . . . . . . . . . . . . . . . . . . . 2–15
§2. Solutions to .Exercises
. . . . . . . . . . . . . . . . . . . . . 2–19

2–2
2–3 §2.2 EQUILIBRIUM PATH AND RESPONSE DIAGRAMS

§2.1. Introduction

This chapter reviews nonlinear structural problems by looking at the manifestation and physical
sources of nonlinear behavior.
We begin by introducing response as a pictorial characterization of nonlinearity of a structural
system. Response is a graphical representation of the fundamental concept of equilibrium path.
This concept permeates the entire course because of both its intrinsic physical value and the fact
that incremental solution methods (mentioned in Chapter 1) are based on it.
Finally, nonlinearities are classified according to their source in the mathematical model of con-
tinuum mechanics and correlated with the physical system. Examples of these nonlinearities in
practical engineering applications are given.

§2.2. Equilibrium Path and Response Diagrams

The concept of equilibrium path plays a central role in explaining the mysteries of nonlinear
structural analysis. This concept lends itself to graphical representation in the form of response
diagrams. The most widely used form of these pictures is the load-deflection response diagram.
Through this representation many key concepts can be illustrated and interpreted in physical,
mathematical or computational terms.

§2.2.1. Load-deflection response

The gross or overall static behavior of many structures can be characterized by a load-deflection
or force-displacement response. The response is usually drawn in two dimensions as a x-y plot
as illustrated in Figure 2.1. In this figure a “representative” force quantity is plotted against a
“representative” displacement quantity. If the response plot is nonlinear, the structure behavior is
nonlinear.

Remark 2.1. We will see below that a response diagram generally depicts the relationship between inputs and
outputs. Or, in more physical terms, between what is applied and what is measured. For structures the most
common inputs are forces and the most common outputs are displacements or deflections1

Remark 2.2. The qualifier “representative” implies a choice among many possible candidates. For relatively
simple structures the choice of load and deflection variables is often clear-cut from considerations such as
the availability of experimental data. For more complex structures the choice may not be obvious, and many
possibilities may exist. The load is not necessarily an applied force but may be an integrated quantity: for
example the weight of traffic on a bridge, or the lift on an airplane wing.

Remark 2.3. This type of response should not be confused with what in structural dynamics is called the
response time history. A response history involves time, which is the independent variable, plotted usually
along the horizontal axis, with either inputs or outputs plotted vertically.

1 A deflection is the magnitude or amplitude of a displacement. Displacements are vector quantities whereas deflections
are scalars.

2–3
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–4

Representative
load

Equilibrium path

Representative
deflection
Reference state

Figure 2.1. A load-deflection response diagram.

§2.2.2. Terminology
A smooth curve shown in a load-deflection diagram is called a path.2 Each point in the path
represents a possible configuration or state of the structure. If the path represents configurations
of static equilibrium it is called an equilibrium path. Each point in an equilibrium path is called an
equilibrium point. An equilibrium point is the graphical representation of an equilibrium state or
equilibrium configuration.
The origin of the response plot (zero load, zero deflection) is called the reference state because it is
the configuration from which loads and deflections are measured. However, the reference state may
be in fact chosen rather arbitrarily, and this freedom is exploited in some nonlinear formulations
and solution methods, as we shall see later.
For problems involving perfect structures3 the reference state is unstressed and undeformed, and is
also an equilibrium state. This means that an equilibrium path passes through the reference state,
as in Figure 2.1.
The path that crosses the reference state is called the fundamental equilibrium path or fundamental
path for short. (Many authors also call this a primary path.) The fundamental path extends from
the reference state up to special states called critical points informally described in §2.3. Any path
that is not a fundamental path but connects with it at a critical point is called a secondary path. See
Figure 2.2.
§2.3. Special Equilibrium Points
Certain points of an equilibrium path have special significance in the applications and thus receive
special names. Of interest to our subject are critical, turning and failure points.

2 The terms branch and trajectory are also used. “Branch” is commonly used in the treatment of bifurcation phenomena,
whereas “trajectory” has temporal or historical connotation.
3 A concept to be explained later in connection with stability analysis. A perfect structure involves some form of idealization
such as perfectly centered loads or perfect fabrication. An imperfect structure is one that deviates from that idealization
in measurable ways.

2–4
2–5 §2.4 LINEAR RESPONSE

Representative
load Critical point

Secondary path
Fundamental or
primary path
Initial linear
response Representative
deflection
Reference state

Figure 2.2. Fundamental (primary) and secondary equilibrium paths.

§2.3.1. Critical points

Critical points are characterized mathematically in later chapters. It is sufficient to note here that
there are two types:

1. Limit points, at which the tangent to the equilibrium path is horizontal, i.e. parallel to the
deflection axis, and

2. Bifurcation points, at which two or more equilibrium paths cross.

At critical points the relation between the given characteristic load and the associated deflection
is not unique. Physically, the structure becomes uncontrollable or marginally controllable there.
This property endows such points with engineering significance.

§2.3.2. Turning points

Points at which the tangent to the equilibrium path is vertical, i.e. parallel to the load axis, are called
turning points. These are not critical points and have less physical significance, but are of interest
for some structures. They have some computational significance, however, because they can affect
the performance of certain solution methods.

§2.3.3. Failure points

Points at which a path suddenly stops or “breaks” because of physical failure are called failure
points. The phenomenon of failure may be local or global in nature. In the first case (e.g, failure of
a noncritical structure component) the structure may regain functional equilibrium after dynamically
“jumping” to another equilibrium path. In the latter case the failure is catastrophic or destructive
and the structure does not regain functional equilibrium.
In the present exposition, bifurcation, limit, turning and failure points are often identified by the
letters B, L, T and F, respectively.
Equilibrium points that are not critical are called regular.

2–5
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–6

Representative
load
goes on forever

Linear fundamental path

Representative
deflection

Figure 2.3. The response diagram for a purely linear structural model.

§2.4. Linear Response

A linear structure is a mathematical model characterized by a linear fundamental equilibrium path

for all possible choices of load and deflection variables. This is shown schematically in Figure 2.3.
The consequences of such behavior are not difficult to foresee:

1. A linear structure can sustain any load whatsoever and undergo any displacement magnitude.
2. There are no critical, turning or failure points.
3. Response to different load systems can be obtained by superposition.
4. Removing all loads returns the structure to the reference position.

The requirements for such a model to be applicable are:

• Perfect linear elasticity for any deformation

• Infinitesimal deformations
• Infinite strength

These assumptions are not only physically unrealistic but mutually contradictory. For example, if
the deformations are to remain infinitesimal for any load, the body must be rigid rather than elastic,
which contradicts the first assumption. Thus, there are necessarily limits placed on the validity of
the linear model.
Despite these obvious limitations, the linear model can be a good approximation of portions of the
nonlinear response. In particular, the fundamental path response in the vicinity the reference state.
See for instance Figure 2.2. Because for many structures this segment represents the operational
or service range, the linear model is widely used in design calculations. The key advantage of
this idealization is that the superposition-of-effects principle applies. Practical implications of the
failure of the superposition principle are further discussed in Chapter 3.

2–6
2–7 §2.6 GENERALIZED RESPONSE

Control parameter λ

Equilibrium path

State parameter µ or u

Figure 2.4. A control-state response diagram.

§2.5. Tangent Stiffness and Stability

The tangent to an equilibrium path may be informally viewed as the limit of the ratio
force increment
displacement increment
This is by definition a stiffness or, more precisely, the tangent stiffness associated with the repre-
sentative force and displacement. The reciprocal ratio is called flexibility or compliance.
The sign of the tangent stiffness is closely associated with the question of stability of an equilibrium
state. A negative stiffness is necessarily associated with unstable equilibrium. A positive stiffness
is necessary but not sufficient for stability.4
If the load and deflection quantities are conjugate in the virtual work sense, the area under a
load-deflection diagram may be interpreted as work performed by the system.
§2.6. Generalized Response
It is often useful to be able to generalize the load-displacement curve of Figure 2.1 in the following
way. A control-state response involves two ingredients:
1. A control parameter, called λ, plotted along the vertical axis versus
2. A state parameter, called u or µ, plotted along the horizontal axis.5
We shall see in following Chapters that λ and u (or µ) characterize in some way the actions applied
to the structure and the state of the structure, respectively.
A diagram such as that shown in Figure 2.4 is called a control-state response. Throughout this
exposition the abbreviated term response is frequently used in this generalized sense. In practice
the control parameter is often a load amplitude or load factorwhereas the state parameter is a
displacement amplitude. Thus the usual load-deflection response is a particular case of the control-
state response.

4 These sign criteria would be sufficient for a one-degree-of-freedom system.

5 We shall use the symbol µ primarily for dimensionless state quantities.

2–7
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–8

(a) (b) (c)

F F L

R R R

Figure 2.5. Basic flavors of nonlinear response: (a) Linear until brittle failure,
(b) Stiffening or hardening, (c) Softening.

Remark 2.4. The interpretation of the tangent-to-the-path as stiffness discussed in §2.5 does not necessarily
carry over to more general control-state diagrams. Similarly, the interpretations of the sign of the tangent and
of the enclosed-area in terms of stability indicator and stored work, respectively, do not necessarily hold. This
is because control and state are not necessarily conjugate in the virtual work sense.

§2.7. Response Flavors

The response diagrams in Figure 2.5 illustrate three “monotonic” types of response: linear, hard-
ening, and softening. Symbols F and L identify failure and limit points, respectively.
A response such as in (a) is characteristic of pure crystals, glassy, and certain high strength composite
materials.
A response such as in (b) is typical of cable, netted and pneumatic (inflatable) structures, which may
be collectively called tensile structures. The stiffening effect comes from geometry “adaptation”
to the applied loads. Some flat-plate assemblies also display this behavior initially.
A response such as in (c) is more common for structure materials than the previous two. A linear
response is followed by a softening regime that may occur slowly or suddenly. More “softening
flavors” are given in Figure 2.6.
The diagrams of Figure 2.6 illustrate a “combination of basic flavors” that can complicate the
response as well as the task of the analyst. Here B and T denote bifurcation and turning points,
respectively.
The snap-through response (d) combines softening with hardening following the second limit point.
The response branch between the two limit points has a negative stiffness and is therefore unstable.
(If the structure is subject to a prescribed constant load, the structure “takes off” dynamically when
the first limit point is reached.) A response of this type is typical of slightly curved structures such
as shallow arches.
The snap-back response (e) is an exaggerated snap-through, in which the response curve “turns
back” in itself with the consequent appearance of turning points. The equilibrium between the
two turning points may be stable and consequently physically realizable. This type of response is
exhibited by trussed-dome, folded and thin-shell structures in which “moving arch” effects occur

2–8
2–9 §2.9 SOURCES OF NONLINEARITIES

(d) (e) (f) (g)

L L B B
T
F B

T T
L F
L F L F

R R R R

Figure 2.6. More complex response patterns: (d) snap-through, (e) snap-back,
(f) bifurcation, (g) bifurcation combined with limit points and snap-back.

following the first limit point; for example cylindrical shells with free edges and supported by end
diaphragms.
In all previous diagrams the response was a unique curve. The presence of bifurcation (popularly
known as “buckling” by structural engineers) points as in (f) and (g) introduces more features. At
such points more than one response path is possible. The structure takes the path that is dynamically
preferred (in the sense of having a lower energy) over the others. Bifurcation points may occur in
any sufficiently thin structure that experiences compressive stresses.
Bifurcation, limit and turning points may occur in many combinations as illustrated in (g). A
striking example of such a complicated response is provided by thin cylindrical shells under axial
compression.

§2.8. Engineering Applications

Nonlinear Structural Analysis is the prediction of the response of nonlinear structures by model-
based simulation. Simulation involves a combination of mathematical modeling, discretization
methods and numerical techniques. As noted in Chapter 1, finite element methods dominate the
discretization step.
Table 2.1 summarizes the most important applications of nonlinear structural analysis.

§2.9. Sources of Nonlinearities

A response diagram characterizes only the gross behavior of a structure, as it might be observed
simply by conducting an experiment on a mechanical testing machine. Further insight into the source
of nonlinearity is required to capture such physical behavior with mathematical and computational
models for computer simulation.
For structural analysis there are four sources of nonlinear behavior. The corresponding nonlinear
effects are identified by the terms material, geometric, force B.C. and displacement B.C., in which
B.C. means “boundary conditions.” In this course we shall be primarily concerned with the last

2–9
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–10

Table 2.1 Engineering Applications of Nonlinear Structural Analysis

Application Explanation

Strength analysis How much load can the structure support before
global failure occurs?
Deflection analysis When deflection control is of primary importance

Stability analysis Finding critical points (limit points or bifurcation

points) closest to operational range
Service configuration analysis Finding the “operational” equilibrium form of certain
slender structures when the fabrication and service
configurations are quite different (e.g. cables, inflat-
able structures, helicoids)
Reserve strength analysis Finding the load carrying capacity beyond critical
points to assess safety under abnormal conditions.
Progressive failure analysis A variant of stability and strength analysis in which
progressive deterioration (e.g. cracking) is consid-
ered.
Envelope analysis A combination of previous analyses in which multiple
parameters are varied and the strength information
thus obtained is condensed into failure envelopes.

three types of nonlinearity, with emphasis on the geometric one.6

The four sources are discussed in more detail in following sections. To remember where the nonlin-
ear terms appear in the governing equations, it is useful to recall the fields that continuum mechanics
deals with, and the relationships among these fields. For linear solid continuum mechanics infor-
mation is presented in Figures 2.7 and 2.8.7
In linear solid mechanics or linear structural mechanics the connecting relationships shown in Figure
2.8 are linear, and so are the governing equations obtained by eliminating all fields but one.
Any of these relations, however, may be nonlinear. Tracing this fact back to physics gives rise to
the types of nonlinearities depicted in Figure 2.9. Relations between body force and stress (the
equilibrium equations) and between strains and displacements (the kinematic equations) are closely

6 the exclusion of constitutive or material nonlinearities does not imply that there are less important than the others. Quite
the contrary. But the topic is covered in separate courses offered in Civil Engineering departments.
7 These are the Tonti diagrams introduced in the IFEM course.

2–10
2–11 §2.10 GEOMETRIC NONLINEARITY

Displacement
Prescribed BCs
displacements Displacements Body forces

Kinematic Equilibrium
equations equations

Force
Constitutive (Traction)
equations Prescribed
Strains Stresses
BCs tractions
or forces

Figure 2.7. Fields in solid continuum mechanics

and connecting relationships.

linked in a “duality” sense, and so the term geometric nonlinearities applies collectively to both sets
of relations. The force BC nonlinearities couple displacements and applied forces (surface tractions
and/or body forces) and thus bring the additional links drawn in Figure 2.8.
In the following sections these sources of nonlinearities are correlated to the physics in more detail.

u=^
u
u^ on Su
u b

T
e=Du D σ+ b = 0
in V in V

σ =Ee σn = ^t
e σ ^t
in V on S t

Figure 2.8. Same as Figure 2.7, with symbols and equations

written down for the linear case.

§2.10. Geometric Nonlinearity

Physical source
Change in geometry as the structure deforms is taken into account in setting up the strain-
displacement and equilibrium equations.

Applications

2–11
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–12

Force B.C.
Displacement B.C. nonlinearities
nonlinearities

^
u u b

Geometric
nonlinearities

Material
nonlinearities

e σ ^t

Figure 2.9. Graphical depiction of sources of nonlinearities

in solid continuum mechanics.

Slender structures in aerospace, civil and mechanical engineering applications. Tensile structures
such as cables and inflatable membranes. Metal and plastic forming. Stability analysis of all types.

Mathematical source
Strain-displacement equations:
e = Du (2.1)
The operator D is nonlinear when finite strains (as opposed to infinitesimal strains) are expressed
in terms of displacements. Internal equilibrium equations:

b = −D∗ σ (2.2)

In the classical linear theory of elasticity, D∗ = DT is the formal adjoint of D, but that is not
necessarily true if geometric nonlinearities are considered.

Remark 2.5. The term geometric nonlinerities models a myriad of physical problems:

Large strain. The strains themselves may be large, say over 5%. Examples: rubber structures (tires, mem-
branes), metal forming. These are frequently associated with material nonlinearities.

Small strains but finite displacements and/or rotations. Slender structures undergoing finite displacements
and rotations although the deformational strains may be treated as infinitesimal. Example: cables, springs,
arches, bars, thin plates.

Linearized prebucking. When both strains and displacements may be treated as infinitesimal before loss of
stability by buckling. These may be viewed as initially stressed members. Example: many civil engineering
structures such as buildings and stiff (non-suspended) bridges.

2–12
2–13 §2.12 FORCE BC NONLINEARITY

§2.11. Material Nonlinearity

Physical source
Material behavior depends on current deformation state and possibly past history of the deformation.
Other constitutive variables (prestress, temperature, time, moisture, electromagnetic fields, etc.)
may be involved.

Applications
Structures undergoing nonlinear elasticity, plasticity, viscoelasticity, creep, or inelastic rate effects.

Mathematical source
The constitutive equations that relate stresses and strains. For a linear elastic material

σ = Ee (2.3)

where the matrix E contains elastic moduli. If the material does not fit the elastic model, general-
izations of this equation are necessary, and a whole branch of continuum mechanics is devoted to
the formulation, study and validation of constitutive equations.

Remark 2.6. The engineering significance of material nonlinearities varies greatly across disciplines. They
seem to occur most often in civil engineering, that deals with inherently nonlinear materials such as concrete,
soils and low-strength steel. In mechanical engineering creep and plasticity are most important, frequently
occurring in combination with strain-rate and thermal effects. In aerospace engineering material nonlinearities
are less important and tend to be local in nature (for example, cracking and “localization” failures of composite
materials).

Remark 2.7. Material nonlinearities may give rise to very complex phenomena such as path dependence,
hysteresis, localization, shakedown, fatigue, progressive failure. The detailed numerical simulation of these
phenomena in three dimensions is still beyond the capabilities of the most powerful computers.

§2.12. Force BC Nonlinearity

Physical Source
Applied forces depend on deformation.

Applications
The most important engineering application concerns pressure loads of fluids. These include
hydrostatic loads on submerged or container structures; aerodynamic and hydrodynamic loads
caused by the motion of aeriform and hydroform fluids (wind loads, wave loads, drag forces). Of
more mathematical interest are gyroscopic and non-conservative follower forces, but these are of
interest only in a limited class of problems, particularly in aerospace engineering.

Mathematical source
The applied forces (prescribed surface tractions
t and/or body forces b) depend on the displacements:

2–13
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–14


t =
t(u), b = b(u), (2.4)
the former being more important in practice.

§2.13. Displacement BC Nonlinearity

Physical source
Displacement boundary conditions depend on the deformation of the structure.

Applications
The most important application is the contact problem,8 in which no-interpenetration conditions
are enforced on flexible bodies while the extent of the contact area is unknown. Non-structural
applications of this problem pertain to the more general class of free boundary problems, for
example: ice melting, phase changes, flow in porous media. The determination of the essential
boundary conditions is a key part of the solution process.

Mathematical source
For the contact problem: prescribed displacements 
d depend on internal displacements u:


d =
d(u) (2.5)

More complicated dependencies can occur in the free-boundary problems mentioned above.

8 Contact-impact in dynamics.

2–14
2–15 Exercises

Homework Exercises for Chapter 2

A Tour of Nonlinear Analysis

EXERCISE 2.1 [D:10 Explain the difference, if any, between a load-deflection response and a control-state
response.

EXERCISE 2.2 [D:20] Can the following occur simultaneously: (a) a limit and a bifurcation point, (b) a
bifurcation and a turning point, (c) a limit and a turning point, (d) two bifurcation points coalescing into one.
If you answer “yes” to an item, sketch a response diagram to justify that reply.

EXERCISE 2.3 [D:25] In §2.10–13, nonlinearities are classified according to physical source into geomet-
ric, material, force boundary conditions, and displacement boundary conditions. For each of the following
mechanical systems indicate the source(s) of nonlinearity that you think are significant; note that there may
be more than one. (If you are not familiar with the underlying concepts, read those sections.)

(a) a long, slender elastic pipe bent under end couples while the pipe material stays elastic. See Figure E2.1.
(b) an inflating balloon. See Figure E2.2.
(c) a cable deflecting under action of wind forces while its material stays elastic. See Figure E2.3.
(d) a forming process in which hot metal is extruded through a rigid die. See Figure E2.4.
(e) a metal anchor is drilled into the soil to serve as a cable support; the hole is then filled with concrete.
See Figures E2.5 and E2.6. The question refers to the soil-drilling process, ignoring dynamics.
(f) a hefty bird — say a condor — sucked into an aircraft jet engine. Ignore dynamics; engine is the structure,
bird the load.

EXERCISE 2.4 [D:15] Can you think of a mechanical component that has the load-deflection response diagram
pictured in Figure E2.7? (Explain why). Hint: Think of a helicoidal spring.

2–15
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–16

Slender tube bent

by end couples

Figure E2.1. Slender elastic pipe bent under end couples for Exercise 2.3(a).

Figure E2.2. Inflating balloon for Exercise 3(b).

2–16
2–17 Exercises

Wind wind load

Cable

Figure E2.3. Cable deflecting under wind forces for Exercise 2.3(c).

Die

Hot metal

Figure E2.4. Hot metal extruded trough a rigid die for Exercise 2.3(d).

2–17
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–18

;;;;;;
;
Figure E2.5. Drill element of a cable anchor, for Exercise 2.3(e).

;;;;;;
;
;;;
;;; ;
(a) (b)

hole

;;;;;;
;
concrete grouting

;;;;;;
;
soil

Figure E2.6. Configuration of cable anchor after drilling in the soil, for Exercise 2.3(e).

Axial force B

II
III

R
Axial deflection (shortening)

Figure E2.7. A “mystery” response diagram for Exercise 2.4.

2–18
 1
.

Overview

1–1
Chapter 1: OVERVIEW 1–2

§1.1 WHERE THE BOOK FITS

This book is an introduction to the analysis of nonlinear elastic structures by the Finite Element
Method (FEM). It embodies five Parts:
I Overview of Nonlinear Problems.
II Formulation of Geometrically Nonlinear Finite Elements.
III Solution Methods.
IV Application to Stability Analysis.
V Nonconservative Problems.
This Chapter presents an overview of where the book fits within the larger scope of Mechanics. It
is assumed that the reader has a good idea of what finite elements are, so this aspect is glossed over.

§1.2 WHERE THIS MATERIAL FITS

The field of Mechanics can be subdivided into three major areas:

 Theoretical
Mechanics Applied (1.1)
Computational

Theoretical mechanics deals with fundamental laws and principles of mechanics studied for their
intrinsic scientific value. Applied mechanics transfers this theoretical knowledge to scientific and
engineering applications, especially as regards the construction of mathematical models of physical
phenomena. Computational mechanics solves specific problems by simulation through numerical
methods implemented on digital computers.

REMARK 1.1
Paraphrasing an old joke about mathematicians, one may define a computational mechanician as a person who
searches for solutions to given problems, an applied mechanician as a person who searches for problems that
fit given solutions, and a theoretical mechanician as a person who can prove the existence of problems and
solutions.

§1.2.1 Computational Mechanics

Several branches of computational mechanics can be distinguished according to the physical scale
of the focus of attention:

 Nanomechanics and micromechanics

  Solids and Structures

Computational Mechanics Continuum mechanics Fluids (1.2)



 Multiphysics
Systems

Nanomechanics deals with phenomena at the molecular and atomic levels of matter. As such it
is closely interrelated with particle physics and chemistry. Micromechanics looks primarily at the

1–2
1–3 §1.2 WHERE THIS MATERIAL FITS

crystallographic and granular levels of matter. Its main technological application is the design and
fabrication of materials and microdevices.
Continuum mechanics studies bodies at the macroscopic level, using continuum models in which
the microstructure is homogenized by phenomenological averages. The two traditional areas of
application are solid and fluid mechanics. The former includes structures which, for obvious
reasons, are fabricated with solids. Computational solid mechanics takes a applied-sciences ap-
proach, whereas computational structural mechanics emphasizes technological applications to the
analysis and design of structures.
Computational fluid mechanics deals with problems that involve the equilibrium and motion of
liquid and gases. Well developed related areas are hydrodynamics, aerodynamics, atmospheric
physics, and combustion.
Multiphysics is a more recent newcomer. This area is meant to include mechanical systems that
transcend the classical boundaries of solid and fluid mechanics, as in interacting fluids and structures.
Phase change problems such as ice melting and metal solidification fit into this category, as do the
interaction of control, mechanical and electromagnetic systems.
Finally, system identifies mechanical objects, whether natural or artificial, that perform a distin-
guishable function. Examples of man-made systems are airplanes, building, bridges, engines,
cars, microchips, radio telescopes, robots, roller skates and garden sprinklers. Biological systems,
such as a whale, amoeba or pine tree are included if studied from the viewpoint of biomechanics.
Ecological, astronomical and cosmological entities also form systems.1
In this progression of (1.2) the system is the most general concept. Systems are studied by de-
composition: its behavior is that of its components plus the interaction between the components.
Components are broken down into subcomponents and so on. As this hierarchical process continues
the individual components become simple enough to be treated by individual disciplines, but their
interactions may get more complex. Thus there is an art in deciding where to stop.2

§1.2.2 Statics vs. Dynamics

Continuum mechanics problems may be subdivided according to whether inertial effects are taken
into account or not:

Statics
Continuum mechanics (1.3)
Dynamics

In dynamics the time dependence is explicitly considered because the calculation of inertial (and/or
damping) forces requires derivatives respect to actual time to be taken.
Problems in statics may also be time dependent but the inertial forces are ignored or neglected.
Static problems may be classified into strictly static and quasi-static. For the former time need not

1 Except that their function may not be clear to us. “The usual approach of science of constructing a mathematical model
cannot answer the questions of why there should be a universe for the model to describe. Why does the universe go to
all the bother of existing?” (Stephen Hawking).
2 Thus in breaking down a car engine, say, the decomposition does not usually proceed beyond the components you can
buy at a machine shop.

1–3
Chapter 1: OVERVIEW 1–4

be considered explicitly; any historical time-like response-ordering parameter (if one is needed)
will do. In quasi-static problems such as foundation settlement, creep deformation, rate-dependent
plasticity or fatigue cycling, a more realistic estimation of time is required but inertial forces are
still neglected.

§1.2.3 Linear vs. Nonlinear

A classification of static problems that is particularly relevant to this book is


Linear
Statics
Nonlinear

Linear static analysis deals with static problems in which the response is linear in the cause-and-
effect sense. For example: if the applied forces are doubled, the displacements and internal stresses
also double. Problems outside this domain are classified as nonlinear.

§1.2.4 Discretization methods

A final classification of CSM static analysis is based on the discretization method by which the
continuum mathematical model is discretized in space, i.e., converted to a discrete model of finite
number of degrees of freedom:



Finite Element Method (FEM)




Boundary Element Method (BEM)
Finite Difference Method (FDM)
Spatial discretization method (1.4)

 Finite Volume Method (FVM)



 Spectral Method
Mesh-Free Method

For linear problems finite element methods currently dominate the scene, with boundary element
methods posting a strong second choice in specific application areas. For nonlinear problems the
dominance of finite element methods is overwhelming.
Classical finite difference methods in solid and structural mechanics have virtually disappeared
from practical use. This statement is not true, however, for fluid mechanics, where finite difference
discretization methods are still dominant. Finite-volume methods, which address finite volume
method conservation laws, are important in highly nonlinear problems of fluid mechanics. Spectral
methods are based on transforms that map space and/or time dimensions to spaces where the problem
is easier to solve.
A recent newcomer to the scene are the mesh-free methods. These are finite different methods on
arbitrary grids constructed by a subset of finite element techniques

§1.2.5 FEM Variants

The term Finite Element Method actually identifies a broad spectrum of techniques that share
common features outlined in introductory FEM textbooks. Two subclassifications that fit well

1–4
1–5 §1.3 THE FEM ANALYSIS PROCESS

VERIFICATION
Mathematical Discretization + solution error
model
IDEALIZATION
REALIZATION
FEM
SOLUTION

Ideal Discrete Discrete

physical model solution
system

IDEALIZATION & VERIFICATION

DISCRETIZATION
solution error

generally irrelevant

Figure 1.1. The Mathematical FEM. The mathematical model (at top) is the source of the process.
Discrete model and solution follow from it. The ideal physical system is inessential.

applications to structural mechanics are


 Displacement 
 Stiffness
Equilibrium
FEM Formulation FEM Solution Flexibility (1.5)

 Mixed Mixed (a.k.a. Combined)
Hybrid

Using the foregoing classification, we can state the topic of this book more precisely: the compu-
tational simulation of nonlinear static structural problems by the Finite Element Method. Of the
variants listed in (1.5), emphasis is placed on the displacement formulation and stiffness solution.
This combination is called the Direct Stiffness Method or DSM.

§1.3 THE FEM ANALYSIS PROCESS

A model-based simulation process using FEM involves doing a sequence of steps. This sequence
takes two canonical configurations depending on the environment in which FEM is used. These
are reviewed next to introduce terminology.

§1.3.1 The Mathematical FEM

The process steps are illustrated in Figure 1.1. The process centerpiece, from which everything
emanates, is the mathematical model. This is often an ordinary or partial differential equation in
space and time. A discrete finite element model is generated from a variational or weak form of
the mathematical model.3 This is the discretization step. The FEM equations are processed by an
equation solver, which delivers a discrete solution (or solutions).

3 The distinction between strong, weak and variational forms is discussed in advanced FEM courses. In the present book
such forms will be stated as recipes.

1–5
Chapter 1: OVERVIEW 1–6

generally
Ideal irrelevant
Mathematical
model

CONTINUIFICATION

SOLUTION

Physical FEM Discrete Discrete

system model solution

IDEALIZATION & VERIFICATION

DISCRETIZATION
solution error

simulation error= modeling + solution error

VALIDATION

Figure 1.2. The Physical FEM. The physical system is the source of
the process. The ideal mathematical model is inessential.

On the left Figure 1.1 shows an ideal physical system. This may be presented as a realization of
the mathematical model; conversely, the mathematical model is said to be an idealization of this
system. For example, if the mathematical model is the Poisson’s equation, realizations may be a
heat conduction or a electrostatic charge distribution problem. This step is inessential and may be
left out. Indeed FEM discretizations may be constructed without any reference to physics.
The concept of error arises when the discrete solution is substituted in the “model” boxes. This
replacement is generically called verification. The solution error is the amount by which the
discrete solution fails to satisfy the discrete equations. This error is relatively unimportant when
using computers, and in particular direct linear equation solvers, for the solution step. More
relevant is the discretization error, which is the amount by which the discrete solution fails to
satisfy the mathematical model.4 Replacing into the ideal physical system would in principle
quantify modeling errors. In the mathematical FEM this is largely irrelevant, however, since the
ideal physical system is merely that: a figment of the imagination.

§1.3.2 The Physical FEM

The second way of using FEM is the process illustrated in Figure 1.3. The centerpiece is now
the physical system to be modeled. Accordingly, this sequence is called the Physical FEM. The
processes of idealization and discretization are carried out concurrently to produce the discrete
model. The solution is computed as before.
Just like Figure 1.1 shows an ideal physical system, Figure 1.1 depicts an ideal mathematical
model. This may be presented as a continuum limit or “continuification” of the discrete model. For
some physical systems, notably those well modeled by continuum fields, this step is useful. For

4 This error can be computed in several ways, the details of which are of little practical importance.

1–6
1–7 §1.3 THE FEM ANALYSIS PROCESS

others, notably complex engineering systems, it makes no sense. Indeed FEM discretizations may
be constructed and adjusted without reference to mathematical models, simply from experimental
measurements.
The concept of error arises in the physical FEM in two ways, known as verification and validation,
respectively. Verification is the same as in the Mathematical FEM: the discrete solution is replaced
into the discrete model to get the solution error. As noted above this error is not generally important.
Substitution in the ideal mathematical model in principle provides the discretization error. This is
rarely useful in complex engineering systems, however, since there is no reason to expect that the
mathematical model exists, and if it does, that it is more relevant than the discrete model. Validation
tries to compare the discrete solution against observation by computing the simulation error, which
combines modeling and solution errors. Since the latter is typically insignificant, the simulation
error in practice can be identified with the modeling error.

§1.3.3 Synergy of Physical and Mathematical FEM

The foregoing physical and mathematical sequences are not exclusive but complementary. This
synergy5 is one of the reasons behind the power and acceptance of the method. Historically the
physical FEM was the first one to be developed to model very complex systems such as aircraft.
The mathematical FEM came later and, among other things, provided the necessary theoretical
underpinnings to extend FEM beyond structural analysis.
A glance at the schematics of a commercial jet aircraft makes obvious the reasons behind the
physical FEM. There is no differential equation that captures, at a continuum mechanics level,6 the
structure, avionics, fuel, propulsion, cargo, and passengers eating dinner.
There is no reason for despair, however. The time honored divide and conquer strategy, coupled
with abstraction, comes to the rescue. First, separate the structure and view the rest as masses
and forces, most of which are time-varying and nondeterministic. Second, consider the aircraft
structure as built of substructures7 wings, fuselage, stabilizers, engines, landing gears, and so on.
Take each substructure, and continue to decompose it into components: rings, ribs, spars, cover,
plates, etc, continuing through as many levels as necessary. Eventually those components become
sufficiently simple in geometry and connectivity that they can be reasonably well described by
continuum mathematical models provided, for instance, by Mechanics of Materials or the Theory
of Elasticity. At that point, stop. The component level discrete equations are obtained from a
FEM library based on the mathematical model. The system model is obtained by going through
the reverse process: from component equations to substructure equations, and from those to the
equations of the complete aircraft. This system assembly process is governed by the classical
principles of Newtonian mechanics expressed in conservation form.
This multilevel decomposition process is diagrammed in Figure 1.3, in which the intermediate

5 This interplay is not exactly a new idea: “The men of experiment are like the ant, they only collect and use; the reasoners
resemble spiders, who make cobwebs out of their own substance. But the bee takes the middle course: it gathers its
material from the flowers of the garden and field, but transforms and digests it by a power of its own.” (Francis Bacon).
6 Of course at the atomic and subatomic level quantum mechanics works for everything, from landing gears to passengers.
But it would be slightly impractical to model the aircraft by 1036 interacting particles.
7 A substructure is a part of a structure devoted to a specific function.

1–7
Chapter 1: OVERVIEW 1–8

ENT
M PON
at ical CO EVEL
hem L
Matmodel
ent
ponns
ary Com atio
Libr en t equ
FEM Com
pon e
TEM
ret
discdel
mo
SYS EL
LEV e
plet
Comution
s o l
em
Systrete
discdel
m o
l
sica
Phy tem
sys

Figure 1.3. Combining physical and mathematical modeling through multilevel FEM.
Only two levels (system and component) are shown for simplicity.

substructure level is omitted for simplicity.

REMARK 1.2
More intermediate decomposition levels are used in some systems, such as offshore and ship structures, which
are characterized by a modular fabrication process. In that case the decomposition mimics the way the system is
actually constructed. The general technique, called superelements, is outlined in the IFEM book. In nonlinear
analysis the analysis procedure is more complex since the static condensation technique of linear analysis no
longer works.

REMARK 1.3
There is no point in practice in going beyond a certain component level while considering the complete model,
since the level of detail can become overwhelming without adding significant information. Further refinement
or particular components is done by the so-called global-local analysis techniques. This technique is an
instance of multiscale analysis.

For sufficiently simple structures, passing to a discrete model is carried out in a single idealization
and discretization step, as illustrated for the truss roof structure shown in Figure 1.4. Multiple
levels are unnecessary here. Of course the truss may be viewed as a substructure of the roof, and
the roof as a a substructure of a building.

§1.4 INTERPRETATIONS OF THE FINITE ELEMENT METHOD

Just like there are two complementary ways of using the FEM, there are two complementary
interpretations for teaching it. One interpretation stresses the physical significance and is aligned
with the Physical FEM. The other focuses on the mathematical context, and is aligned with the
Mathematical FEM.

1–8
1–9 §1.4 INTERPRETATIONS OF THE FINITE ELEMENT METHOD

member

support

joint
Physical System

IDEALIZATION

;;
;;
;;

;;
;;
;;
Figure 1.6. The idealization process for a simple structure. The physical system,
here a roof truss, is directly idealized by the mathematical model: a
pin-jointed bar assembly. This coalesces with the discrete model.

§1.4.1 Physical Interpretation

The physical interpretation is aligned with the view of Figure 1.2. This interpretation has been
shaped by the discovery and extensive use of the method in the field of structural mechanics. This
relationship is reflected in the use of structural terms such as “stiffness matrix”, “force vector” and
“degrees of freedom.” This terminology carries over to non-structural applications.
The basic concept in the physical interpretation is the breakdown (≡ disassembly, tearing, partition,
separation, decomposition) of a complex mechanical system into simpler, disjoint components
called finite elements, or simply elements. The mechanical response of an element is characterized
in terms of a finite number of degrees of freedom. These degrees of freedoms are represented as
the values of the unknown functions as a set of node points. The element response is defined by
algebraic equations constructed from mathematical or experimental arguments. The response of
the original system is considered to be approximated by that of the discrete model constructed by
connecting or assembling the collection of all elements.
The breakdown-assembly concept occurs naturally when an engineer considers many artificial and
natural systems. For example, it is easy and natural to visualize an engine, bridge, aircraft or
skeleton as being fabricated from simpler parts.
The underlying theme is divide and conquer. If the behavior of a system is too complex, the
recipe is to divide it into more manageable subsystems. If these subsystems are still too complex
the subdivision process is continued until the behavior of each subsystem is simple enough to fit a
mathematical model that represents well the knowledge level the analyst is interested in. In the finite
element method such “primitive pieces” are called elements. elements The behavior of the total
system is that of the individual elements plus their interaction. A key factor in the initial acceptance
of the FEM was that the element interaction can be physically interpreted and understood in terms
that were eminently familiar to structural engineers.

1–9
Chapter 1: OVERVIEW 1–10

§1.4.2 Mathematical Interpretation

This interpretation is closely aligned with the configuration of Figure 1.1. The FEM is viewed as
a procedure for obtaining numerical approximations to the solution of boundary value problems
(BVPs) posed over a domain . This domain is replaced by the union ∪ of disjoint subdomains
(e) called finite elements. In general the geometry of  is only approximated by that of ∪(e) .
The unknown function (or functions) is locally approximated over each element by an interpolation
formula expressed in terms of values taken by the function(s), and possibly their derivatives, at a
set of node points generally located on the element boundaries. The states of the assumed unknown
function(s) determined by unit node values are called shape functions. The union of shape functions
“patched” over adjacent elements form a trial function basis for which the node values represent the
generalized coordinates. The trial function space may be inserted into the governing equations and
the unknown node values determined by the Ritz method (if the solution extremizes a variational
principle) or by the Galerkin, least-squares or other weighted-residual minimization methods if the
problem cannot be expressed in a standard variational form.

§1.5 THE SOLUTION MORASS

In nonlinear analysis the two FEM interpretations are not equal in importance. Nonlinear analysis
demands a persistent attention to the underlying physics to avoid getting astray as the “real world”
is covered by layer upon layer of mathematics and numerics.
Why is concern for physics of paramount importance? A key component of finite element nonlinear
analysis is the solution of the nonlinear algebraic systems of equations that arise upon discretization.

FACT

The numerical solution of nonlinear systems in “black box” mode is much

more difficult than in the linear case.

The key difficulty is tied to the essentially obscure nature of general nonlinear systems, about which
very little can be said in advance. And you can be sure that Murphy’s law8 works silently in the
background.
One particularly vexing aspect of dealing with nonlinear systems is the solution morass. A deter-
minate system of 1, 1000, or 1000000 linear equations has, under mild conditions, one and only
one solution. The computer effort to obtain this solution can be estimated fairly accurately if the
sparseness (or denseness) of the coefficient matrix is known. Thus setting up linear equation solvers
as “black-box” stand-alone functions or modules is a perfectly sensible thing to do.
By way of contrast, a system of 1000 cubic equations has 31000 ≈ 10300 solutions in the complex
plane. This is much, much larger than the number of atoms in the Universe, which is merely 1050

8 If something can go wrong, it will go wrong.

1–10
 1–11 §1.6 HISTORICAL BACKGROUND

give or take a few. Suppose just several billions or millions of these are real solutions. Which
solution(s) have physical meaning? And how do you compute those solutions without wasting time
on the others?
This combinatorial difficulty is overcome by the concept of continuation, which engineers also call
incremental analysis. Briefly speaking, we start the analysis from an easily computable solution
— for example, the linear solution — and then try to follow the behavior of the system as actions
applied to it are changed by small steps called increments. The previous solution is used as a starting
point for the iterative solution-search procedure. The underlying prescription: follow the physics.
This technique is interwined with the concept of response explained in Chapter 2.
REMARK 1.4
Not surprisingly, incremental analysis was used by the aerospace engineers that first used the finite element
method for geometrically nonlinear analysis in the late 1950s. Techniques have been considerably refined
since then, but the underlying idea remains the same.

We conclude this overview with a historical perspective on nonlinear finite element methods in
solid and structural mechanics, along with a succint bibliography.

§1.6 HISTORICAL BACKGROUND

In the history of finite element methods the year 1960 stands out. The name “finite element method” appears
for the first time in the open literature in an article by Clough [52]. And Turner, Dill, Martin and Melosh [187]
publish a pioneering paper in nonlinear structural analysis. The then-five-year-old “direct stiffness method”
(what we now call displacement-assumed finite element method) was applied to
“problems involving nonuniform heating and/or large deflections . . . in a series of linearized steps. Stiffness
matrices are revised at the beginning of each step to account for changes in internal loads, temperatures, and
geometric configuration.”
Thirty years and several thousand publications later, computerized nonlinear structural analysis has acquired
full adult rights, but has not developed equally in all areas.
The first fifteen years (1960-1975) were dominated by formulation concerns. For example, not until the
late 1960s were correct finite-deflection incremental forms for displacement models rigorously derived. And
interaction of flow-like constitutive behavior with the spatial discretization (the so called “incompressibility
locking” effects) led to important research into constitutive equations and element formulations.
While the investigators of this period devoted much energy to obtaining correct and implementable nonlinear
finite-element equations, the art of solving such equations in a reliable and efficient manner was understandably
neglected. This helps to explain the dominance of purely incremental methods. Corrective methods of Newton
type did not get much attention until the early 1970s, and then only for geometrically nonlinear problems. At
the time of this writing, progress in numerical solution techniques has been uneven: well developed for certain
problems, largely a black art in others. To understand the difference, it pays to distinguish between smooth
nonlinearities and rough nonlinearities.

§1.6.1 Smooth Nonlinearities

Problem with smooth nonlinearities are characterized by continuous, path-independent nonlinear relations at
the local level. Some examples:

1. Finite deflections (geometric nonlinearities). Nonlinear effects arise from strain-displacement equations,
which are well behaved for all strain measures in practical use.

1–11
Chapter 1: OVERVIEW 1–12

2. Nonlinear elasticity. Stresses are nonlinear but reversible functions of strains.

3. Follower forces (e.g., pressure loading). External forces are smooth nonlinear functions of displacements.

A unifying characteristic of this problem is that nonlinearities are of equality type, i.e., reversible, and these
relations are continuous at each point within the structure. Mathematicians call these smooth mappings.
It is important to point out, however, that the overall structural behavior is not necessarily smooth; as witnessed
by the phenomena of buckling, snapping and flutter. But at the local level everything is smooth: nonlinear
strain-displacement equations, nonlinear elasticity law, follower pressures.
Methods for solving this class of problems are highly developed, and have received a great deal of attention
from the mathematical and numerical analysis community. This research has directly benefitted many areas
of structural analysis.
Let us consider finite deflection problems as prototype. Within the finite element community, these were
originally treated by purely incremental (step-by-step) techniques; but anomalies detected in the mid-1960’s
prompted research into consistent linearizations. A good exposition of this early work is given in the book by
Oden [123]. Once formulation questions were settled, investigators had correct forms of the “residual” out-
of-balance forces and tangent stiffness matrix, and incremental steps began to be augmented with corrective
iterations in the late 1960s. Conventional and modified Newton methods were used in the corrective phase.
These were further extended through restricted step (safeguarded Newton) and, more recently, variants of the
powerful conjugate-gradient and quasi-Newton methods.
But difficulties in detecting and traversing limit and bifurcation points still remained. Pressing engineering
requirements for post-buckling and post-collapse analyses led to the development of displacement control,
alternating load/displacement control, and finally arclength control. The resultant increment-control methods
have no difficulty in passing limit points. The problem of reliably traversing simple bifurcation points without
guessing imperfections remains a research subject, while passing multiple or clustered bifurcation points
remains a frontier subject. A concerted effort is underway, however, to subsume these final challenges.
These reliable solution methods have been implemented into many special-purpose finite element programs,
and incorporation into general-purpose programs is proceeding steadily.

REMARK 1.5
As noted above, incremental methods were the first to be used in nonlinear structural analysis. Among the pre-1970
contributions along this line we may cite Argyris and coworkers [9, 10], Felippa [62], Goldberg and Richard [82], Marcal,
Hibbitt and coworkers [89,108,109], Oden [122], Turner, Martin and coworkers [110,187,188],

REMARK 1.6
The earliest applications of Newton methods to finite element nonlinear analysis are by Oden [122], Mallet and Marcal [107],
and Murray and Wilson [117,118]. During the early 1970s Stricklin, Haisler and coworkers at Texas A&M implemented
and evaluated self-corrective, pseudo-force, energy-search and Newton-type methods and presented extensive comparisons;
see Stricklin et. al. [175–178], Tillerson et. al. [185], and Haisler et. al. [85]. Almroth, Brogan, Bushnell and coworkers
at Lockheed began using true and modified Newton methods in the late 1960s for energy-based finite-difference collapse
analysis of shells; see Brogan and Almroth [34], Almroth and Felippa [5], Brush and Almroth [37], and Bushnell [38–39].
By the late 1970s Newton-like methods enjoyed widespread acceptance for geometrically nonlinear analysis.

REMARK 1.7
Displacement control strategies for finite element post-buckling and collapse analysis were presented by Argyris [11] and
Felippa [62] in 1966, and generalized in different directions by Sharifi and Popov [165,166] (fictitious springs), Bergan et.
al. [25,26], (current stiffness parameter), Powell and Simons [142] and Bergan and Simons [27] (multiple displacement
controls). A modification of Newton’s method to traverse bifurcation points was described by Thurston [184]. Arclength
control schemes for structural problems may be found in the following source papers: Wempner [194], Riks [154], Schmidt

1–12
1–13 §1.6 HISTORICAL BACKGROUND

[160], Crisfield [48,49], Ramm [146], Felippa [65–66], Fried [73], Park [133], Padovan [130,131], Simo et.al. [167], Yang
and McGuire [199], Bathe and Dvorkin [20]. Other articles of particular interest are Bathe and Cimento [18], Batoz and
Dhatt [22], Bushnell [39], Bergan [27], Geradin et al. [79,80]. Meek and Tan [112], Ramm [146,147], Riks [155–157], Sobel
and Thomas [170], Zienkiewicz [200,201,203]. Several conferences have been devoted exclusively to nonlinear problems
in structural mechanics, for example [12,28,17,125, 179,180,198]. Finite element textbooks and monographs dealing rather
extensively with nonlinear problems are by Oden [123], Bathe [19], Bushnell [40], White [196] and Zienkiewicz [202].

REMARK 1.8

In the mathematical literature the concept of continuation (also called imbedding) can be traced back to the 1930s. A survey
of the work up to 1950 is given by Ficken [70]. The use of continuation by parameter differentiation as a numerical method
is attributed to Davidenko [54]. Key papers of this early period are by Freudenstein and Roth [72], Deist and Sefor [58] and
Meyer [114], as well as the survey by Wasserstrom [191]. This early history is covered by Wacker [190].

REMARK 1.9

Arclength continuation methods in the mathematical literature are generally attributed to Haselgrove [87] and Klopfestein
[99] although these papers remained largely unnoticed until the late 1970s. Important contributions to the mathematical
treatment are by Abbott [1], Anselone and Moore [8], Avila [14], Brent [31], Boggs [29], Branin [30], Broyden [35,36],
Cassel [43], Chow et. al. [45], Crandall and Rabinowitz [47], Georg [77,78], Keller and coworkers [44,56,57,93–96],
Matthies and Strang [111], Moore [115,116], Pönish [139,140], Rheinboldt and coworkers [59,113, 149–150], Watson
[192] and Werner and Spence [195]. Of these, key contributions in terms of subsequent influence are [45,94,149]. For
surveys and edited proceedings see Allgower [2,3], Byrne and Hall [42], Küpper [104,105], Rall [145], Wacker [190], and
references therein. Textbooks and monographs dealing with nonlinear equation solving include Chow and Hale [46], Dennis
and Schnabel [61], Kubı́ček and Hlaváček [102], Kubı́ček and Marek [103], Ortega and Rheinboldt [128], Rabinowitz [143],
Rall [144], Rheinboldt [153], and Seydel [164]. Of these, the book by Ortega and Rheinboldt remains a classic and an
invaluable source to essentially all mathematically oriented work done prior to 1970. The book by Seydel contains material
on treatment of conventional and Hopf bifurcations not readily available elsewhere. Nonlinear equation solving is interwined
with the larger subject of optimization and mathematical programming; for the latter the textbooks by Gill, Murray and
Wright [81] and Fletcher [71] are highly recommended.

§1.6.2 Rough Nonlinearities

Rough nonlinearities are characterized by discontinuous field relations, usually involving inequality constraints.
Examples: flow-rule plasticity, contact, friction. The local response is nonsmooth.
Solution techniques for these problems are in a less satisfactory state, and case-by-case consideration is called
for. The local and overall responses are generally path-dependent, an attribute that forces the past response
history to be taken into account.
The key difficulty is that conventional solution procedures based on Taylor expansions or similar differential
forms may fail, because such Taylor expansions need not exist! An encompassing mathematical treatment
is lacking, and consequently problem-dependent handling is presently the rule. For this class of problems
incremental methods, as opposed to incremental-iterative methods, still dominate.

REMARK 1.10

Earliest publications on computational plasticity using finite element methods are by Gallagher et. al. [74], Argyris [9,10],
Marcal [108], Pope [138] and Felippa [62]. By now there is an enormous literature on the numerical treatment of inelastic
processes, especially plasticity and creep. Fortunately the survey by Bushnell [40], although focusing on plastic buckling,
contains over 300 references that collectively embody most of the English-speaking work prior to 1980. Other important
surveys are by Armen [13] and Willam [197]. For contact problems, see Oden [126], Bathe and Chaudhary [19], Kikuchi
and Oden [97,98], Simo et. al. [167] Stein et. al. [173], Nour-Omid and Wriggers [121], and references therein.

1–13
Chapter 1: OVERVIEW 1–14

§1.6.3 Hybrid Approach

What does an analyst do when faced with an unfamiliar nonlinear problem? If the problem falls into the
smooth-nonlinear type, there is no need to panic. Robust and efficient methods are available. Even if the
whizziest methods are not implemented into one’s favorite computer program, there is a wealth of theory and
practice available for trouble-shooting.
But what if the problem include rough nonlinearities? A time-honored general strategy is divide and conquer.
More specifically, two powerful techniques are frequently available: splitting and nesting.
Splitting can be used if the nonlinearities can be separated in an additive form:

Smooth + Rough

This separation is usually done at the force level. Then the smooth-nonlinear term is treated by conventional
techniques whereas the rough-nonlinear term is treated by special techniques. This scheme can be particularly
effective when the rough nonlinearity is localized, for example in contact and impact problems.
Nesting may be used when a simple additive separation is not available. This is best illustrated by an actual
example. In the early 1970s, some authors argued that Newton’s method would be useless for finite-deflection
elastoplasticity, as no unique Jacobian exists in plastic regions on account of loading/unloading switches. The
argument was compelling but turned out to be a false alarm. The problem was eventually solved by “nesting”
geometric nonlinearities within the material nonlinearity, as illustrated in Fig. 1.1.
In the inner equilibrium loop the material law is “frozen”, which makes the highly effective Newton-type meth-
ods applicable. The non-conservative material behavior is treated in an outer loop where material properties
and constitutive variables are updated in an incremental or sub-incremental manner.
Another application of nesting comes in the global function approach (also called Rayleigh-Ritz or reduced-
basis approach), which is presently pursued by several investigators. The key idea is to try to describe the
overall response behavior by a few parameters, which are amplitudes of globally defined functions. The small
nonlinear system for the global parameters is solved in an inner loop, while an external loop involving residual
calculations over the detailed finite element model is executed occasionally.
Despite its inherent implementation complexity, the global function approach appears cost-effective for smooth,
path-independent nonlinear systems. This is especially so when expensive parametric studies are involved, as
in structural optimization under nonlinear stability constraints.

REMARK 1.11
For geometric-material nesting and subincremental techniques see Bushnell [38–41], and references therein. The global-
function approach in its modern form was presented by Almroth, Stern and Brogan [7] and pursued by Noor and coworkers
under the name of reduced-basis technique; see Noor and Peters [119] and Noor [120] as well as the chapter by Noor in this
volume. For perturbation techniques see the survey by Gallagher [75].

§1.6.4 Summary of Present Status

Solution techniques for smooth nonlinearities are in a fairly satisfactory state. Although further refinements
in the area of traversing bifurcation points can be expected, incremental-iterative methods implemented with
general increment control appear to be as reliable as an engineer user may reasonably expect.
For rough nonlinearities, case-by-case handling is still necessary in view of the lack of general theories and
implementation procedures. Separation or nesting of nonlinearities, when applicable, can lead to signifi-
cant gains in efficiency and reliability, but at the cost of programming complexity and problem-dependent
implementations.

1–14
 30
.

Dynamic
Stability
Analysis

30–1
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–2

TABLE OF CONTENTS

Page
§30.1. Introduction 30–3
§30.2. The Linearized Equations of Motion 30–3
§30.3. The Characteristic Problem 30–4
§30.3.1. Connection with the Free-Vibration Eigenproblem . . . . . 30–5
§30.4. Characteristic Exponents and Stability 30–5
§30.4.1. Negative Real Case: Harmonic Oscillations . . . . . . . 30–5
§30.4.2. Positive Real Case: Divergence . . . . . . . . . . . 30–6
§30.4.3. Complex Case: Flutter . . . . . . . . . . . . . . 30–7
§30.4.4. Stable and Unstable Regions in the Complex Plane . . . . . 30–8
§30.5. Graphical Representations 30–8
§30.5.1. Root locus plots . . . . . . . . . . . . . . . . 30–8
§30.5.2. Amplitude Plots . . . . . . . . . . . . . . . . 30–10
§30.6. Regression to Zero Frequency and Static Tests 30–11
§30. Exercises . . . . . . . . . . . . . . . . . . . . . . 30–12
§30. Solutions to .Exercises
. . . . . . . . . . . . . . . . . . . . . 30–16

30–2
30–3 §30.2 THE LINEARIZED EQUATIONS OF MOTION

§30.1. Introduction

If the loading is nonconservative the loss of stability may not show up by the system going into
another equilibrium state but by going into unbounded motion. To encompass this possibility we
must consider the dynamic behavior of the system because stability is essentially a dynamic concept
(recall the definition in §25.1).
The essential steps are as follows. We investigate the motion that occurs after some initial per-
turbation is applied to the equilibrium state being tested, and from the properties of the motion
we can infer or deny stability. It if turns out that the perturbed motion consists of oscillations of
increasing amplitude, or is a rapidly increasing departure from the equilibrium state, the equilibrium
is unstable; otherwise it is stable.
The practicality of this approach depends crucially on the linearization of the equations of motion of
the perturbation. Thus we avoid having to trace the ensemble of time histories for every conceivable
dynamic departure from equilibrium — which for a system with many degrees of freedom would
clearly be a computationally forbidding task.
By linearizing we can express the perturbation motion as the superposition of complex exponential
elementary solutions. The characteristic exponents of these solutions can be determined through
a characteristic value problem or eigenproblem. This problem includes the free-vibration natural
frequency eigenproblem as particular case when the system is conservative and the tangent stiffness
matrix is symmetric. Through the stability criterion discussed in §30.3, the set of characteristic ex-
ponents gives complete information on the linearized stability of the system at the given equilibrium
configuration.
In practical studies the characteristic exponents are functions of the control parameter λ. Assuming
that the system is stable for sufficiently small λ values, say λ = 0, we are primarily concerned with
finding the first occurrence of λ at which the system loses stability. The transition to instability may
occur in two different ways, which receive the names divergence and flutter, respectively.1
The distinction between divergence and flutter instability is important in that the singular-stiffness
test discussed in Chapter 26 remains valid if the stability loss occurs by divergence, although of
course the tangent stiffness is not necessarily symmetric. Therefore it follows that in that case
we may fall back upon the static criterion, which is simpler to apply because it does not involves
information about mass and damping. Such a regression is not possible, however, if the loss of
stability occurs by flutter.

§30.2. The Linearized Equations of Motion

The structure is in static equilibrium under a given value of the control parameter λ. The equilibrium
state is defined by the state vector u. At time2 τ = 0 apply a dynamic input (e.g., an impulse) to this
configuration and examine the subsequent motion of the system. Roughly speaking if the motion
is unbounded (remains bounded) as τ tends to infinity the system is dynamically unstable (stable).

1 These names originated in aeronautical engineering applications, more specifically the investigation of sudden airplane
“blow ups” during the period 1910-1930. In the mathematical literature flutter goes by the name ‘Hopf bifurcation.’
2 The symbol τ denotes real time because t is used throughout the course to denote a pseudo-time parameter. Only real
time is considered in this Chapter.

30–3
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–4

As noted in the Introduction, to simplify the mathematical treatment we consider only the local
stability condition, in which the imparted excitation is so tiny that the subsequent motion can be
viewed as a linearizable perturbation. We are effectively dealing with small perturbations about
the equilibrium position.
Let M be the symmetric mass matrix, which is assumed positive definite, and K the tangent
stiffness matrix, which is real but generally unsymmetric because of load nonconservativeness. The
perturbation motion is denoted as
d(τ ) = u(τ ) − u(0), τ ≥ 0+ (30.1)
The discrete, unforced, undamped governing equations of motion are
Md̈ + Kd = 0, (30.2)
in which a superposed dot — unlike previous Chapters — denotes differentiation with respect to
real time. The ordinary differential equations (30.2) express the linearized dynamic equilibrium
between stiffness and inertial forces. The stiffness forces generally include nonconservative loading
effects.
Remark 30.1. In structural with rotational DOFs, M might be only nonnegative definite because of the presence
of zero rotational masses. If so it is assumed that those DOFs have been eliminated by a static condensation
process.
The assumption of positive definiteness also excludes the presence of Lagrange multipliers in the state vector
u, because the associated masses of such degrees of freedom are zero. Again the stability criteria can be
extended by eliminating the multipliers in the linearized equation of motion.

Remark 30.2. We shall ignore damping effects because of two reasons:

(1) The effect of diagonalizable, light viscous structural damping does not generally affect stability results
(it certainly does not when stability loss is by divergence). See also Remark 30.4.
(2) The effect of more complicated nonlinear damping mechanisms such as dry friction may not be amenable
to linearization.
Thus cases when damping effects are significant lead to mathematics beyond the scope of this course. Readers
interested in pursuing this topic are referred to the vast literature on the subject of dynamic stability.

§30.3. The Characteristic Problem

The linear ODE system (30.2) can be treated by assuming the eigenmodal expansion
 
d(τ ) = di (τ ) = zi e pi τ , (30.3)
i i

where i ranges over the number of degrees of freedom (number of state parameters). The pi are
generally complex numbers called the characteristic exponents whereas the corresponding column
vectors zi are the characteristic modes or characteristic vectors.3
3 In his classical treatise Nonconservative Problems of the Theory of Elastic Stability, (Pergamon, 1963), Bolotin employs
s for what we call here p, and so do many other authors. This notation connects well to the common use of the Laplace
transform to do more complicated systems. However, we have already reserved s for Piola-Kirchhoff stresses as well as
arclength.

30–4
30–5 §30.4 CHARACTERISTIC EXPONENTS AND STABILITY

Replacing d̈i = pi2 d into (30.2) yields

(K + pi2 M) zi = 0, (30.4)

which is the characteristic problem or eigenproblem that governs dynamic stability. This equation
befits the generalized unsymmetric eigenproblem of linear algebra

Axi = µi Bxi (30.5)

in which matrix A ≡ K is real and generally unsymmetric whereas B ≡ M is real symmetric

positive definite. The eigenvalues µi ≡ − pi2 of this eigenproblem may be either real or complex; if
the latter, they occur in conjugate pairs. The square roots of these eigenvalues yield the characteristic
exponents pi of the eigenmodal expansion (30.3).

§30.3.1. Connection with the Free-Vibration Eigenproblem

If the system is conservative and stable, K is symmetric and positive definite. If so all roots pi2 of
(30.4) are negative real and their square roots are purely imaginary numbers:

pi = ± jωi , (30.6)
√
where j = −1, and the nonnegative real numbers ωi are the natural frequencies of free vibration.
Because p 2 = −ωi2 , (30.4) reduces to the usual vibration eigenproblem

(K − ωi2 M) zi = 0. (30.7)

Thus for the conservative case we regress to a well studied problem. In such a case the system will
simply vibrate, that is, perform harmonic oscillations about the equilibrium position because each
root is associated with the solution

e jωi τ = cos ωi τ + j sin ωi τ. (30.8)

The presence of positive damping will of course damp out these oscillations and the system even-
tually returns to the static equilibrium position.

§30.4. Characteristic Exponents and Stability

The characteristic exponents are generally complex numbers:

pi = αi + jωi , (30.9)
√
where αi and ωi are real numbers, and j = −1. The component representation of the square of
pi is
pi2 = (αi2 − ωi2 ) + 2 jαi ωi , (30.10)
The exponential of a complex number has the component representation

e pi τ = e(αi + jωi )τ = eαi τ (cos ωi τ + j sin ωi τ ), (30.11)

On the basis of this representation we can classify the growth behavior of the subsequent motion
and consequently the stability of the system as examined in the next 3 subsections.

30–5
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–6

di

2π/ωi

Figure 30.1. Harmonic oscillatory motion for the case where root pi2 of (30.4) is negative
real. Equivalently, pi = ± jωi where ωi is the circular frequency.

§30.4.1. Negative Real Case: Harmonic Oscillations

If

pi = ± jωi , d(τ ) = di (τ ), di (τ ) = Ai cos ωi τ + Bi sin ωi τ. (30.12)

where Ai and Bi are determined by initial conditions. The motion di associated with ± jωi is
harmonic and bounded, as illustrated in Figure 30.1. The system is dynamically stable for this
individual eigenvalue.
If all eigenvalues are negative real and distinct, the system is dynamically stable because any
superposition of harmonic motions of different periods is also a harmonic motion. If two or more
eigenvalues coalesce the analysis becomes more complicated because of the appearance of secular
terms that grow linearly in time. These effects can be studied in more detail in treatises in mechanical
vibrations.

di
eαi τ

e−αi τ

Figure 30.2. Aperiodic, exponentially growing motion for the real root case pi2 = αi2 ,
pi = ±αi . Transition to this kind of instability is called divergence.

30–6
30–7 §30.4 CHARACTERISTIC EXPONENTS AND STABILITY

di
eαi τ

2π/ωi

−eαi τ

Figure 30.3. Periodic, exponentially growing motion for case pi = +αi ± jωi with
nonzero αi . Transition to this kind of instability is called flutter.

§30.4.2. Positive Real Case: Divergence

If pi2 is positive real,
pi = ±αi . (30.13)
The +αi square root will give rise to an aperiodic, exponentially growing motion. The other root
will give rise to an exponentially decaying motion. When the two solutions are combined the
exponentially growing one will dominate for sufficiently large τ as sketched in Figure 30.2, and the
system is then exponentially unstable.
As noted above pi2 is generally a function of λ. The transition from stability (in which all roots are
negative real) to this type of instability necessarily occurs when a eigenvalue pi2 (λ), moving from
left to right as λ varies, passes through the origin p 2 = 0 of the p 2 complex plane. This type of
instability is called divergence.

§30.4.3. Complex Case: Flutter

If pi2 is complex, solutions of the eigenproblem (30.4) occur in conjugate pairs because both
matrices M and K are real. Consequently, if pi2 = (αi2 − ωi2 ) + j (2αi ωi ) is a complex eigenvalue
so is its conjugate ( pi2 ) = (αi2 − ωi2 ) − j (2αi ωi ). On taking the square root of this pair we find four
characteristic exponents
±αi ± jωi . (30.14)
Two of these square roots will have positive real parts (+α) and for sufficiently large τ they will
eventually dominate the other pair, yielding exponentially growing oscillations; see Figure 30.3.
This is called periodic exponential instability or flutter instability.
If the system is initially stable (i.e., all roots are negative real) then transition to this type of instability
occurs when at a certain value of λ two real roots coalesce on the real axis and “branch out” into
the complex p 2 plane. This loss of stability is called flutter.

30–7
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–8

Remark 30.3. Frequency coalescence is necessary but not sufficient for flutter. It is possible for two frequencies
to pass by other “like ships crossing in the night” without merging. This happens if there is no mechanism by
which the two associated eigenmodes can exchange energy.

Remark 30.4. The fact that all characteristic motions are either harmonic or exponentially growing is a
consequence of the neglect of damping in setting up the stability problem. As noted in Remark 30.2, the
presence of damping or, in general, dissipative forces, introduces additional mathematical complications that
will not be elaborated upon here. Suffices to say that the addition of damping to a conservative system has
always a stabilizing effect (Rayleigh’s theorem). For non-conservative systems, the preceding statement is no
longer true, and indeed several counterexamples involving destabilizing damping have been constructed over
the past 40 years. In spite of this the effect is not often observed in practice.

Remark 30.5. The occurrence of flutter requires the coalescence of two natural frequencies. Consequently,
flutter cannot occur in systems with one degree of freedom (“it takes two to flutter”). The physical interpretation
of the flutter phenomenon is that one vibration mode absorbs energy and feeds it into another; this transference
or “energy resonance” becomes possible when the two modes have the same frequency.

(a) p 2 plane (b) p plane ω

2αω

Stable Unstable

ω2 − α 2 α

Stable Unstable

Figure 30.4. Stable and unstable regions in (a) the complex p 2 plane, (b) the
complex p plane. For the latter the stable region is the left-half
plane α =
( p) ≤ 0. For (a) it is the negative real axis.

§30.4.4. Stable and Unstable Regions in the Complex Plane

From the preceding study it follows that the only stable region in the complex p 2 -plane is the
negative real axis:

( p 2 ) < 0, ( p 2 ) = 0. (30.15)
The rest of the p 2 complex plane is unstable; see Figure 30.4(a).
On the complex p-plane, the stable region is the left-hand plane

α =
( p) ≤ 0. (30.16)

which includes the imaginary axis α = 0 as stability boundary. The right-hand p-plane α > 0 is
unstable. See Figure 30.4(b).

30–8
30–9 §30.5 GRAPHICAL REPRESENTATIONS

p 2 plane p plane

j pi (λ)
root trajectory
pi2 (λ)

Divergence
root trajectory Divergence − j pi (λ)

Figure 30.5. Root locus plots on the complex p 2 and p planes for divergence instability.

p 2 plane p plane
j p1 (λ) trajectory of
interacting roots

p12 (λ) p22 (λ) j p2 (λ) Flutter

− j p2 (λ)
trajectory of
interacting roots Flutter

− j p1 (λ)

Figure 30.6. Root locus plots on the complex p 2 and p planes for flutter instability.

§30.5. Graphical Representations

§30.5.1. Root locus plots

Graphical representations of the “trajectories” of the eigenvalues pi (λ) as λ is varied on the complex
p 2 or p planes are valuable insofar as enhancing the understanding of the differences between
divergence and flutter. These are called root locus plots4 and are illustrated in Figures 30.5 and
30.6.
Figure 30.5 illustrates loss of stability by divergence. As λ is varied, eigenvalue pi2 passes from the
left-hand plane to the right-hand plane through the origin p 2 = 0. Stability loss occurs at the λ for
which pi2 vanishes. The right-hand diagram depicts the same phenomenon on the p plane, for the
root pair ± pi .

4 The word root in root-locus is used as abbreviation for characteristic root or eigenvalue

30–9
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–10

| p|

| p1 | = |ω1 |

Divergence

λcr
Figure 30.7. Root amplitude plot illustrating loss of stability by divergence at λcr .

| p|

| p1,2 |
| p2 | = |ω2 |

| p1 | = |ω1 |
Flutter

λ
λcr
Figure 30.8. Root amplitude plot illustrating loss of stability by flutter at λcr .

Figure 30.6 illustrates loss of stability by flutter. As λ is varied, two interacting eigenvalues, labeled
as p12 and p22 , coalesce on the negative real axis of the p 2 plane and branch out into the unstable
region. The right-hand diagram depicts the same phenomenon on the p plane for the interacting
roots, which appears in complex-conjugate pairs.

§30.5.2. Amplitude Plots

Another commonly used visualization technique is the characteristic root amplitude or simply root
amplitude plots. These plots show the magnitude of pi (λ), that is | pi (λ)| on the vertical axis against
λ on the horizontal axis. If the eigenvalue is real, | pi | is simply its absolute value whereas if it is
complex | pi | is its modulus.
This graphical representation enjoys the following advantages: (a) the critical value of λ is displayed

30–10
30–11 §30.6 REGRESSION TO ZERO FREQUENCY AND STATIC TESTS

more precisely than with a locus or trajectory plot, (b) all related square roots such as ±αi ± ωi
“collapse” into a single value, and (c) the variation of several important roots (for several values of
i) may be shown without cluttering the picture.
Figures 30.7 and 30.8 illustrate typical root-amplitude plots in loss of stability by divergence and
flutter, respectively.

§30.6. Regression to Zero Frequency and Static Tests

The stability loss by divergence occurs when an eigenvalue pi vanishes. Because ωi = 0 if pi = 0,

this is equivalent to a zero-frequency test on the eigenproblem

(−ωi2 M + K) zi = 0. (30.17)

But if ωi = 0 and M is positive definite, which we assume, then K must be singular. Therefore we
can regress to the static criterion or singular tangent stiffness test

det K(λ) = 0, (30.18)

which allows us to discard the mass matrix. This regression may be useful if one is solving a series
of closely related problems, for example during the design of a structure which is known a priori
to become unstable by divergence.
It should be cautioned, however, that the tangent stiffness matrix K for nonconservative systems
is generally unsymmetric (Chapter 29), and that the test for singularity must take account of that
property.

30–11
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–12

Homework Exercises for Chapter 30

Dynamic Stability Analysis

EXERCISE 30.1 (A+C:25) This Exercise studies the stability of the “follower load” nonconservative system
shown in Figure E30.1.

ux
λP
90◦

2
L

;;
;;

Bar (1) uy

1 2 1
k1 90◦
C0
y
Bar (2) k2
L C

;; ;;
x
3 3

Figure E30.1. Structure for Exercise 30.1.

Two elastic bars, (1) and (2), are supported at 1 and 3 and hinged at 2. The bars have length L, axial stiffnesses
k1 and k2 , respectively, and can only move in the x, y plane. Bar (1) is loaded at node 2 by a force λP1 ,
directed upwards, that stays normal to bar (1) as it displaces. Bar (2) is loaded at node 2 by a force λP2 ,
directed leftwards, that stays normal to bar (2) as it displaces.
For the present exercise set P1 = P2 = P. Furthermore the following simplifying assumptions are to be made:

(A1) The displacements from the reference configuration are so small that C ≡ C0 insofar as setting up the
stability eigensystem5

(A2) The contribution of the geometric stiffness is neglected.

(a) Show that under the simplifying assumptions (A1)–(A2), the tangent stiffness at C ≡ C0 in terms of the
two degrees of freedom u x = u x2 and u y = u y2 , is
  λP  
k1 0 0 1
K= + . (E30.1)
0 k2 L −1 0

5 This is similar to LPB (Chapters 24-25), but here a dynamic analysis is involved.

30–12
30–13 Exercises

The first component of K is the material stiffness whereas the second component is the load stiffness.
Hint for the latter: use the results of Remark 30.4
(b) The linearized dynamic eigenproblem (30.4) is

( pi2 M + K)zi = 0, i = 1, 2. (E30.2)

The exponents pi (the square roots of pi2 ) are generally complex numbers:

pi = αi + jωi , (E30.3)

where α and ω are the real and imaginary part of pi , respectively, zi are associated eigenmodes, and M
is the diagonal mass matrix  
M 0
M= , (E30.4)
0 M

where M is the lumped mass at node 2 (half of the sum of the bar masses). By appropriate normalization
show that the eigenproblem can be reduced to the dimensionless form
      
1 0 κ 0 0 1
p̄ 2 + + λ̄ z̄i = 0, (E30.5)
0 1 0 1 −1 0

where κ = k1 /k2 , p̄ and λ̄ are dimensionless.

(c) Show that the critical positive λ̄cr at which the eigenvalues p̄i2 coalesce is given by the relation

|1 − κ|
λ̄cr = . (E30.6)
2

Further show that if λ̄ > λ̄cr the roots p̄i become complex and hence explain whether loss of stability
occurs. Is it divergence or flutter?
(d) For κ = 0.01, 1.0, 4.0 and 100 plot the dependence of | p̄i | (i = 1, 2) (where |.| denotes the modulus of
a complex number) on λ using
√ √
| p̄|/ κ, λ/ κ, (E30.7)
as vertical and horizontal axes, respectively. Go from λ = 0 up to 2λcr or 1.0, whichever is greater, and
use sufficient steps to get reasonable graphical accuracy.

EXERCISE 30.2 (A+C:25) Do the previous exercise removing assumption (A2), that is, considering now the
effect of the geometric stiffness KG but still assuming C ≡ C0 . Is there any difference with the critical load
result (E30.6)?

EXERCISE 30.3 (A+C:30) Beck’s column6 is the simplest follower-load problem involving a cantilevered
beam-column.7 This problem is shown in Figure E30.2.
The beam-column has length L, elastic modulus E and smallest moment of inertia I . It is loaded by a
compressive force λP which after deformation rotates with the end section of the column and remains tangential
to its deformed axis (see Figure above). The mass M (half of the column mass) is lumped at its free end.

6 M. Beck, Die Knicklast des eiseiting eigenspannen, tangential gedrückten Stabes, Z. angew. Math. Phys., 3, No. 3, 1952.
7 It is sometimes used as a very simple model to illustrate stability analysis of rockets against the “pogo” effect.

30–13
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–14

x θ

M λP
1

C0 C
L

;;;
E, A, I constant

2 y

Figure E30.2. Beck’s column: structure for Exercise 30.3.

If this problem is treated by the static criterion (Euler’s method) one erroneously concludes that the beam
column cannot lose stability for any value of the load λP 8 A dynamic stability analysis, first carried out by
Beck (loc.cit.), shows that stability is lost by flutter at the critical load
EI
λPcr = 20.05093 . (E30.8)
L2

(a) Find the critical dynamic load given by the finite element method if one Euler-Bernoulli beam-column
element is used along the length of Beck’s column. Lateral displacements may be considered infinites-
imal; hence sin θ ≈ θ, cos θ ≈ 1, and the axial force is simply λP. The degrees of freedom are u x1 ,
u y1 and θz1 . Use the material and geometric stiffness matrices given in equations (E24.2) and (E24.3),
respectively, to which an unsymmetric load stiffness matrix K L , which couples the θz1 and u y1 degrees
of freedom, should be added.
(b) Repeat the analysis for two and four elements of equal length along the column. For two elements the
three nodes are 1 (top), 2 (middle of column) and 3 (root). Use lumped masses with M X 2 = MY 2 equal
to one half of the total column mass and M X 1 = MY 1 = M X 2 /2 = MY 2 /2. For four elements there are
five nodes, etc. Use of Mathematica or a similar program is recommended.

8 See for example, pp. 7–8 of Bolotin’s book cited in footnote 3.

30–14
30–15 Exercises

Solution of Exercise 30.3(a) for one-element discretization:

The dynamic matrix perturbation equation taking C ≈ C0 is
 E A 
M 0 0   ü x  L
0 0 0 0 0 
0 M 0 ü y +   0 12E I − 6E I 
− P 0 36 −3L
L3 L2 30L
0 0 0 θ̈z 0 −3L 4L 2
0 − 6E2I 4E I
(E30.9)
L
0 0 0   u x   0 
L

+P 0 0 1 uy = 0
0 0 0 θz 0
where for simplicity u x = u x1 , u y = u y1 , θz = θz1 . The first dynamic equation in u x uncouples and has no effect in the
analysis. The last equation is static in nature because the rotational mass is zero. Thus, we can solve for θz in terms of u y :

P
6E I
− +
L 210 N
θz = uy = uy (E30.10)
4E I 4P L D
−
L 30
where N and D denote the numerator and denominator, respectively, of the relation that links θz to u y . The eigenvalue
equation becomes  12E I 6E I N 36P N P N

p2 M + − − + 3L + P u y = 0. (E30.11)
L3 L2 D 30L D 30L D
One of the bending eigenvalues p 2 of (E30.9) is always ∞ because the rotational mass is zero. Flutter occurs when the two
beding eigenvalues coalesce at infinity. The finite p 2 becomes infinite if D = 0 while N = 0. Thus the critical load for
“flutter at infinity” is
30E I
Pcr = (E30.12)
L2

which is about 50% in error with respect to the analytical value 20.05093E I /L 2 quoted in the exercise statement.

EXERCISE 30.4 (A:25) Do the previous exercise for a one-element discretization if the line of action of
the applied end load is forced to pass through the cantilever root (point 2). Does the structure loses stability
dynamically or statically?

30–15
 29
.

Nonconservative
Loading

29–1
Chapter 29: NONCONSERVATIVE LOADING 29–2

TABLE OF CONTENTS

Page
§29.1. Introduction 29–3
§29.2. Potential Force Example: Gravity 29–3
§29.3. Follower Load and Associated Load Stiffness 29–5
§29.4. General Characterization of the Load Stiffness 29–6
§29.5. Forces Produced by Fluid Motion 29–7
§29.6. Load Stiffness For 2D Fluid Motion 29–9
§29. Exercises . . . . . . . . . . . . . . . . . . . . . . 29–11

29–2
29–3 §29.2 POTENTIAL FORCE EXAMPLE: GRAVITY

§29.1. Introduction
In Chapter 5 a mechanical system was defined to be conservative when both external and internal
forces are derivable from a potential. In this course we consider only elastic systems; conse-
quently the internal forces are derivable from an strain (internal) energy potential U . Thus the
conservative/nonconservative character depends on whether the external loads are conservative or
nonconservative.
Conservative applied forces f may be derived from the external loads potential V by differentiating
with respect to the state variables:
∂V
f= . (29.1)
∂u
Nonconservative forces, on the other hand, are not expressable as (29.1). They have to be worked
out directly at the force level.
In the present Chapter we will give examples of both force types in conjunction with the TL-
formulated two-node bar element The main result is that consideration of nonconservative loads
contributes an unsymmetric component, called load stiffness, to the tangent stiffness matrix. Treat-
ing this effect in stability analysis requires a dynamic criterion, which is covered in Chapter 30.
Remark 29.1. The chief sources of nonconservative forces in various branches of engineering are:
1. Aerodynamic forces (aerospace, civil); hydrodynamic forces (mechanical, marine, chemical); aircraft
and rocket propulsion forces (aerospace); frictional forces (mechanical, civil).
2. Gyroscopic forces (aerospace, electrical).
3. Active control systems (aerospace, electrical, mechanical).
In this Chapter we consider only hydrodynamic (fluid motion) forces as prototype example.

§29.2. Potential Force Example: Gravity

Consider the two-node, three-dimensional bar element immersed in a gravity field of constant
strength g acting along the global −Z axis, as illustrated in Figure 29.1. The bar has reference
length L 0 , reference area A0 and mass density ρ. The element coordinate systems are labeled as
follows:
x̄0 , ȳ0 , z̄ 0 in the reference configuration C0
x̄, ȳ, z̄ in the current configuration C
This distinction between local coordinate systems is introduced here as it becomes necessary in
later Sections. Take a differential element of bar of length d x̄0 in C0 . This moves to a corresponsing
position in C, with a vertical displacement of u z with respect to C0 . See Figure 29.2. The work
potential gained by this displacement is
d V = −ρg A0 u z (x̄0 ) d x̄0 (29.2)
The external potential of the bar element is obtained by linearly interpolating u z = (1−ζ )u z1 +ζ u z2 ,
ζ = x̄/L 0 and integrating over the bar length:
 L  1  
u z1
V =− ρg A0 u z d x̄0 = − A0 g [1 − ζ ζ ] L 0 dζ
0 0 u z2 (29.3)
= −ρg A0 L 0 21 (u z1 + u z2 ).

29–3
Chapter 29: NONCONSERVATIVE LOADING 29–4

x̄
g = directed along − z 2
C
Z, z
1
x̄0
C0

Y, y
E, A0 , L 0
X, x
Figure 29.1. TL bar element displacing in a gravity field g.

x̄
2
C
g Z, z
1
vz x̄0
C0

Y, y
ρ A0 d x̄0
X, x
Figure 29.2. Calculation of external potential.

(As usual in the TL kinematic description, all quantities are referred to C0 .) It follows that the
external force vector for the element is
   
∂ V /∂u x1 0
 ∂ V /∂u y1  0
∂V    
 ∂ V /∂u z1  1
fg = =  = − 2 ρ A0 L 0   .
1
(29.4)
∂u  ∂ V /∂u x2  0
   
∂ V /∂u y2 0
∂ V /∂u z2 1
This can also be derived through basic principles of statics. Note that this vector is independent of

29–4
29–5 §29.3 FOLLOWER LOAD AND ASSOCIATED LOAD STIFFNESS

L x̄
ȳ
pd

θ
C
uX1 uY2
Y, y uY1
pd

1 2 x̄0 ≡ X
uX2
C0
L0
X, x
Figure 29.3. 2D bar under constant “follower” pressure pd .

the current configuration. This is a distinguishing feature of external work potentials that depend
linearly on the displacements, such as (29.3).
§29.3. Follower Load and Associated Load Stiffness
To illustrate the concept of load stiffness with a minimum of mathematics, let us consider a two-
dimensional specialization. The bar element originally lies along the x axis in the reference con-
figuration C0 and moves in the (x, y) plane to C, which forms an angle θ with x. The bar is under a
a constant pressure pd that is always normal to the element as it displaces, as shown in Figure 29.3.
This kind of applied force is called a follower load in the literature.1
From statics the external force vector is obviously
 
− sin θ
 cos θ 
 
 0 
f = 2 pd L 
1
 (29.5)
 − sin θ 
 
cos θ
0
From geometry
L 0 + u X 21 u Y 21
cos θ = , sin θ = , with u X 21 = u X 2 − u X 1 , u Y 21 = u Y 2 − u Y 1 , (29.6)
L L
1 Such loads are often applied by fluids at rest or in motion. The latter case is studied in Sections 29.4-5.

29–5
Chapter 29: NONCONSERVATIVE LOADING 29–6

Consequently  
−u Y 21
 L 0 + u X 21 
 
 0 
f = 12 pd  . (29.7)
 −u Y 21 
 
L 0 + u X 21
0

Take now the partial of the negative of this external load vector with respect to u. The result is a
matrix with dimensions of stiffness, denoted by K L :
 
0 −1 0 0 1 0
1 0 0 −1 0 0 
∂f  
0 0 0 0 0 0
KL = − = 2 pd 
1
. (29.8)
∂u  0 −1 0 0 1 0
 
1 0 0 −1 0 0
0 0 0 0 0 0
K L is called a load stiffness matrix. It arises from displacement-dependent loads.2 We can see
from this example that K L is unsymmetric. A consequence of this fact is that (29.2) does not have
a potential V that is a function of the node displacements.3

§29.4. General Characterization of the Load Stiffness

Suppose that we have a one-parameter conservative system with displacement dependent forces.
Then
 = U (u) − V (u, λ), (29.9)
where the external potential V = V (u, λ) depends on the displacements u in a general fashion.
Then
∂ ∂U ∂V
r= = − = p − f, (29.10)
∂u ∂u ∂u
∂r ∂p ∂f
K= = − . (29.11)
∂u ∂u ∂u
The partial ∂p/∂u gives K M + KG , the material plus geometric stiffness, as discussed in previous
Chapters. The last term gives K L , the conservative load stiffness
∂f ∂2V
KL = − =− 2 (29.12)
∂u ∂u
which is called the conservative load stiffness. This matrix is obviously symmetric because it is the
negated Hessian of V (u, λ) with respect to u. Consequently

K = K M + KG + K L . (29.13)

2 This source of nonlinearity was called force B.C. nonlinearity in Chapter 2.

3 If K L were symmetric we could work backwards and integrate (29.5), expressed in terms of the node displacements, to
find the potential function V .

29–6
29–7 §29.5 FORCES PRODUCED BY FLUID MOTION

These three components of K are symmetric, and so is K.

Now consider a more general structural system subject to both conservative and non-conservative
loads:
r = p − fc − fn , (29.14)
Here fc = ∂ V /∂u whereas fn collects external forces not derivable from a potential. Then

∂r
K= = K M + KG + K Lc + K Ln . (29.15)
∂u
The nonconservative load stiffness matrix, K Ln , is unsymmetric.
Remark 29.2. In practice one derives the total force f from statics, as in the example of §29.3, and obtains
K L by taking the partials with respect to the displacements in u. If the resulting stiffness is unsymmetric the
load is nonconservative. The splitting of K L into a symmetric matrix K Lc and unsymmetric part K Ln can be
done in a variety of ways. (If the unsymmetric part is required to be antisymmetric, however, the splitting is
unique.)

§29.5. Forces Produced by Fluid Motion

To study in more detail a frequent source of non-conservative follower loads, suppose that the bar
element is submerged in a moving fluid whose flow is independent of time — i.e., a steady flow.
See Figure 29.4. We neglect “feedback” effects on the flow due to the presence and motion of the
bar. The steady notion can be described by the fluid-particle velocity field4
 
u f X (X, Y, Z )
u f (X, Y, Z ) = u f Y (X, Y, Z ) , (29.16)
u f Z (X, Y, Z )

For simplicity in the formulation below, we further assume that the velocity field is uniform, i.e.,
does not depend upon (X, Y, Z ), and that it is directed along the x axis:
 
ufX
uf = 0 , (29.17)
0

where u f X is independent of position.

By virtue of drag effects the fluid motion exerts a normal drag force pd (force per unit length) upon
the bar in the current configuration C. The drag force is normal to the bar longitudinal axis x̄ and
it is a function of the magnitude of the velocity component normal to that axis. Furthermore if the
bar cross section is circular or annular, the force is coaxial with the normal velocity vector. For
additional simplicity we shall assume that the cross section satisfies such a geometric constraint5

4 The symbol u and its vector counterparts u and u are commonly used in fluid mechanics to denote velocities rather
than displacements as in structural and solid mechanics. In fact displacements are rarely used in fluids. Subscript f is
introduced here to lessen the risk of confusion with structural displacements.
5 For arbitrary cross sections, the fluid motion exerts drag and lift forces, the latter being normal to the bar axis and to the
normal velocity vector. Lift forces are what makes airplanes fly. This more general situation is dealt with in treatises on
aerodynamics, wind forces and hydraulics.

29–7
Chapter 29: NONCONSERVATIVE LOADING 29–8

z̄

ȳ
Y, y fluid velocity vector
v f X in (x̄, X ) plane

X, x
Z, z
x̄
C0

x̄0

Figure 29.4 Bar element in steady fluid flow.

For slow (laminar) flow the drag force is proportional to the magnitude of the normal velocity
component whereas if the motion is fully turbulent it is proportional to the square of that velocity.
We assume here the latter case. Other drag-velocity dependencies can be similarly treated.
Consider the bar in the (x̄, X ) plane as illustrated in Figure 29.5, and let ȳ be defined as the normal
to the element axis x̄ that is located in this plane and forms an acute angle θ with x. The drag force
on the element per unit length is directed along ȳ and has the value

pd = 12 Cd ρ f d u 2f n (29.18)

where Cd is the drag coefficient,6 ρ f the fluid mass density, d the “exposed width” (for a bar
of circular cross-section, its external diameter), and u f n the fluid-normal velocity u f X cos θ (see
Figure). The total force on the element is pd L, where L is the current length, and this force “lumps”
into 12 pd L at each node.
In order to refer these forces to the global X, Y, Z axes, we need to know the direction cosines t21 ,
t22 and t23 of ȳ with respect to x, y, z. Then the hydrodynamic node force vector in the (X, Y, Z )

6 C D is a dimensionless number tabulated in fluid dynamic handbooks

29–8
29–9 §29.6 LOAD STIFFNESS FOR 2D FLUID MOTION

ȳ, n
paper is plane (x̄, X ) ≡ plane (x̄, ȳ)

1 vfn
θ
X, x vfX
C
2

x̄

Figure 29.5. Normal fluid velocity component in the current bar configuration

system is  
t21
 t22 
 
t 
f = 2 pd L  23 
1
(29.19)
 t21 
 
t22
t23
To compute these direction cosines, one proceeds as follows:
(1) Compute the direction z̄ by taking the cross product of x̄ and X .
(2) Compute the direction ȳ by taking the cross product of z̄ and x̄.
If x̄ and X are parallel, step (1) does not define z but then the fluid flow occurs along the element
axis and the pressure pd vanishes.
Remark 29.3. If the fluid flow is uniform with speed u f j along a general direction j ≡ j, the preceding
derivation must be modified by taking z̄ = x̄ × j, ȳ = z̄ × x̄,
 θ = angle( ȳ, j). Observe that it would be
incorrect to decompose u f j onto its components in the X , Y and Z directions and superpose associated forces,
because the drag force is nonlinear in the velocity.

Remark 29.4. If the flow is steady but nonuniform, numerical integration over elements is generally required.
For this simple elkement integration with the flow velocity evaluated at the element center is often sufficient.

§29.6. Load Stiffness For 2D Fluid Motion

To show what kind of load stiffness is produced by fluid drag forces, consider again the case of
Figure 29.4 but now make pd depend on the “tilt” θ as explained in §29.3; see Figure 29.6. Since
a turbulent-motion-induced drag force is proportional to the square of u f n = u f X cos θ, it may be
expressed as
pd = pd0 cos2 θ (29.20)
where pd0 is pd for θ = 0 (bar normal to fluid motion).

29–9
Chapter 29: NONCONSERVATIVE LOADING 29–10

L x̄
2 θ
ȳ c os
pd 0
=
pd
θ
C
uX1 uY2
uY1
Y, y
pd0

1 2 x̄0 ≡ X
uX2
C0
L0
X, x
Figure 29.6. Follower pressure pd on a 2D bar that depends on the “tilt angle” θ.

The external load vector is  

− sin θ cos2 θ
 cos3 θ 
 
 0 
f = 12 pd0 L   (29.21)
 − sin θ cos2 θ 
 
cos3 θ
0
To differentiate this expression under the assumption that pd0 does not depend on the node dis-
placements, and that L is constant, we need partial derivative expressions such as

∂(− sin θ cos2 θ) ∂ cos θ ∂ sin θ

= −2 sin θ cos θ − cos2 θ
∂u X 21 ∂u X 21 ∂u X 21 (29.22)
1 1
= sc(c − 2s ) = sc(1 − 3s ),
2 2 2
L L
etc. The resulting load stiffness K L = −∂f/∂u is more complicated than (29.8), but still can be
obtained in closed form.
If L is let to vary, then one can substitute cos θ = (L 0 + u X 21 )/L and sin θ = u Y 21 /L to put f in
terms of u X 21 and u Y 21 , and the differentiation to get K L becomes straightforward. Thus the exact
expression is in fact easier to work out than the approximate one. The details of the derivation are
worked out in Exercise 29.5.

29–10
29–11 Exercises

Homework Exercise for Chapter 29

Nonconservative Loading

EXERCISE 29.1 (A:20) Work out fd for the case of a uniform flow of speed u f j in a general direction j as
described in Remark 29.2.

EXERCISE 29.2 (A:15) Specialize the result of Exercise 29.1 to the two dimensional case (bar and flow in
the x, y plane). Differentiate to obtain K L , comparing with (29.20).

EXERCISE 29.3 (A:20) In the previous exercise take into account the effect of friction forces exerted on the
bar by the flow. Use the linear model: the tangential friction force pt per unit length of the bar is directed
along x̄ and has the value C f au f t , where C f is a friction coefficient, a is the “exposed perimeter” of the bar
(for a circular cross section, a = 2π d), and u f t = u f j sin θ is the tangential velocity (fluid velocity projected
on the current bar direction, with proper sign).

EXERCISE 29.4 (A:20) Prove the formulas (29.10).

EXERCISE 29.5 (A:20) Complete the derivation of K L in §29.6.

EXERCISE 29.6 (A:30) A simple example of a gyroscopic force is a torsional moment M̄x directed along the
longitudinal axis x̄ of a beam-column element, which keeps pointing in that direction as the element moves
and rotates. Obtain the gyroscopic force vector fn and associated load stiffness K Ln for a three-dimensional
beam column of length L currently directed along the global x axis. The element degrees of freedom are

uT = [ u x1 u y1 u z1 θx1 θ y1 θz1 u x2 u y2 u z2 θx2 θ y2 θz2 ] . (E29.1)

For this “moment tilting” analysis it is sufficient to assume that: (a) node 1 stays fixed, (b) the element remain
straight, and (c) any deviations from the current x direction are infinitesimal.

29–11
 28
.

Imperfections

28–1
Chapter 28: IMPERFECTIONS 28–2

TABLE OF CONTENTS

Page
§28.1. No Body is Perfect 28–3
§28.2. The Imperfect Hinged Cantilever 28–3
§28.2.1. Equilibrium Analysis . . . . . . . . . . . . . . . 28–3
§28.2.2. Critical Point Analysis . . . . . . . . . . . . . . 28–3
§28.2.3. Discussion . . . . . . . . . . . . . . . . . . 28–4
§28.3. The Imperfect Propped Cantilever 28–5
§28.4. Parametrizing Imperfections 28–7
§28.5. Imperfection Sensitivity at Critical Points 28–8
§28.5.1. Limit Point . . . . . . . . . . . . . . . . . 28–9
§28.5.2. Asymmetric Bifurcation . . . . . . . . . . . . . . 28–9
§28.5.3. Stable Symmetric Bifurcation . . . . . . . . . . . 28–10
§28.5.4. Unstable Symmetric Bifurcation . . . . . . . . . . . 28–10
§28.6. Extensions: Multiple Bifurcation, Continuous Systems 28–10
§28. Exercises . . . . . . . . . . . . . . . . . . . . . . 28–12

28–2
28–3 §28.2 THE IMPERFECT HINGED CANTILEVER

§28.1. No Body is Perfect

In the previous four Chapters we have been concerned with the behavior of geometrically perfect
structures. For the geometrically nonlinear analysis of slender structures, such as those used in
aerospace products, we must often take into account the presence of imperfections. It is useful to
distinguish two type of imperfections, one associated with the physical structure, the other with the
computational model.
Physical imperfections. Physical imperfections may be categorized into fabrication and load im-
perfections. Real structures inevitably carry geometric imperfections inherent in their manufacture.
In addition, loads on structural members that carry primarily compressive loads, such as columns
and cylindrical shells, are not necessarily centered. The load-carrying capacity of certain classes
of structures, notably thin shells, may be significantly affected by the presence of physical imper-
fections. We shall see that high sensitivity to the presence of small imperfections is a phenomenon
associated with certain types of critical points. Structures that exhibit high sensitivity are called
imperfection sensitive.
Numerical imperfections. Imperfections may be incorporated in the computational model for vari-
ous reasons. Numerical imperfections may be used to either simulate actual physical imperfections
or to “trigger” the occurrence of certain types of response. One common application of numerical
imperfections is in fact to “nudge” the structure along a post-bifurcation path, as in Exercises 21.2
and 21.3.
We begin the study of the effect of imperfections through a simple yet instructive one-degree-of-
freedom example: the imperfect hinged cantilever.
§28.2. The Imperfect Hinged Cantilever
We take up again the critical-point analysis of the hinged cantilever already studied in §25.5. But we assume
that this system is geometrically imperfect in the sense that the rotational spring is unstrained when the rigid
bar “tilts” by a small angle  with the vertical. By varying  we effectively generate a family of imperfect
systems that degenerate to the perfect system when  → 0.
Denoting again the total rotation from the vertical by θ as shown in Figure 28.1, the strain energy of the
imperfect system can be written
U (θ, ) = 12 k(θ − )2 . (28.1)
The potential energy of the imperfect system is
1 
(θ, λ, ) = U − V = 12 k(θ − )2 − f L(1 − cos θ ) = k 2
(θ − )2 − λ(1 − cos θ ) , (28.2)
in which as before we take λ = f L/k as dimensionless control parameter.

§28.2.1. Equilibrium Analysis

The equilibrium equation in terms of the angle θ as degree of freedom is
∂
r= = k(θ −  − λ sin θ ) = 0. (28.3)
∂θ
Therefore, the equilibrium path equation of an imperfect system is
θ −
λ= . (28.4)
sin θ

28–3
Chapter 28: IMPERFECTIONS 28–4

ε p

L rigid

;;;;
k

Figure 28.1. The imperfect hinged cantilever. The imperfection

parameter is the initial tilt angle .

§28.2.2. Critical Point Analysis

The first-order incremental equation in terms of θ is the same as in Chapter 25:

K θ̇ − q λ̇ = 0, (28.5)

where
∂r ∂r
K = = k(1 − λ cos θ ), q= = k sin θ. (28.6)
∂θ ∂θ
We have stability if K > 0, that is
1 − λ cos θ > 0, (28.7)

and instability if K < 0, that is

1 − λ cos θ < 0. (28.8)

Critical points are characterized by K (λcr ) = 1 − λcr cos θ = 0, or

1
λcr = . (28.9)
cos θ

On equating this value of λ with that given by the equilibrium solution (28.4) we obtain

θ −  = tan θ. (28.10)

This relation characterizes the locus of critical points as  is varied. It is not difficult to show that these critical
points are limit points if  = 0 (imperfect systems) and a bifurcation point if and only if  = 0 (perfect system).

28–4
28–5 §28.3 THE IMPERFECT PROPPED CANTILEVER

>0 1.75 Unstable <0

1.5

1.25

0.75 Stable
<0 >0
0.5

0.25

θ (rad)
-2 -1 0 1 2

Figure 28.2. Equilibrium paths of the imperfect ( = 0)

and perfect ( = 0) hinged cantilever.

§28.2.3. Discussion

The response of this family of imperfect systems is displayed in Figure 28.2.

In this Figure heavy lines represent the response of the peerfect system whereas light lines represent the
responses of imperfect systems for fixed values of . Furthermore continuous lines identify stable equilibrium
path portions whereas broken lines identify unstable portions. We see that systems with a positive  give
equilibrium paths in two opposite quadrants while systems with a negative  give equilibrium paths in the
remaining two quadrants. The equilibrium paths of the imperfect systems collapse onto the equilibrium paths
of the perfect system as  goes to zero. The locus of critical-point equilibrium states given by (28.10) separates
the stable and unstable domains and is shown in Figure 28.2 as curve ss.
We see that a given imperfect system loaded from its unstrained state will give rise to a constantly rising path
so that no instability is encountered; the deflections merely growing more rapidly as the critical load of the
perfect system is passed. In addition to this natural equilibrium path an imperfect system will also have a
complementary path which lies in the opposing quadrant. However, this path (partly stable and partly unstable)
will not be encountered in a natural loading process that starts from λ = 0.
The response shown in Figure 28.2 is well knwon to structural engineers and is exhibited by the familiar Euler
column which is taught in elementary courses of mechanics of materials. In §28.5 it is shown that this behavior
is characteristic of systems that possess a stable-symmetric bifurcation point.

§28.3. The Imperfect Propped Cantilever

The perfect propped cantilever is shown in Figure 28.3. It differs from the hinged cantilever in that it is
supported by an ordinary (rectilinear) spring of stiffness k attached to the top. An imperfect version is shown
in Figure 28.4, where the initial horizontal displacement  L defines the imperfection parameter .

28–5
Chapter 28: IMPERFECTIONS 28–6

u = L sin θ

;; k p
L(1 − cos θ)

; θ
L

;;
rigid column

;;
Figure 28.3. The perfect propped cantilever.

L u = L sin θ

; k p

;
L(1 − cos θ)

θ
L

;;
Figure 28.4. The imperfect propped cantilever. The imperfection
parameter is , where  L is the displacement from the
vertical at which the rectilinear spring is unstrained.

The potential energy of the imperfect structure is


(u, f ) = U − V = 12 k(u −  L)2 − f L (1 − cos θ) = 12 k(u −  L)2 − f L 1 − (u/L)2 (28.11)

where u = L sin θ is the total horizontal displacement from the vertical, and a constant term has been dropped
from V . It is convenient to take the ratio λ = f L/k as dimensionless control parameter and µ = u/L as the

28–6
28–7 §28.4 PARAMETRIZING IMPERFECTIONS

1.4
>0 <0
1.2
Unstable
1

0.8

Stable
0.6

0.4

<0 >0
0.2

µ
-1 -0.5 0 0.5 1

Figure 28.5. Equilibrium paths of the imperfect ( = 0)

and perfect ( = 0) propped cantilever.

dimensionless state variable. Then the potential energy, upon dividing by k L 2 , becomes

(µ, λ) = 12 (µ − )2 − λ 1 − µ2 (28.12)

The residual equation in terms of µ, λ and the imperfection parameter is

µ
r (µ, λ, ) = µ −  − λ  =0 (28.13)
1 − µ2

Carrying out the analysis as in the previous section, one finds the response paths depicted in Figure 28.5.
The important difference is that the bifurcation point is now of unstable-symmetric type. The equilibrium
paths that emerge from the unloaded state of the imperfect systems are no longer rising but exhibit limit-point
maxima that may be viewed as failure loads. These limit points occur at lower loads than the bifurcation load
of the perfect structure.
Therefore, the load-carrying capacity of the propped cantilever is adverseley affected by the imperfection, and
the structure is said to be imperfection-sensitive.

§28.4. Parametrizing Imperfections

The treatment of imperfections in the foregoing two examples illustrates many features that recur
in more complex cases. Imperfect systems are derived as perturbations of the perfect system.
Imperfections in real structures are seldom known precisely. They are usually random quantities
that can be rigorously treated only by stochastic techniques. Such a treatment, however, would be
hopelessly expensive in nonlinear systems. A more practical deterministic approach consists of
looking at a parametrized family of imperfect systems characterized by a dimensionless imperfection
parameter .

28–7
Chapter 28: IMPERFECTIONS 28–8

λ λcr

<0
L

>0 =0
µ 

Figure 28.6. Effect of initial imperfectons at a limit point.

The system corresponding to  = 0 is called the perfect system. Systems corresponding to  = 0

are described as imperfect systems. Parameter  is inserted in the potential energy (u, λ, ). For
each fixed  the analysis proceeds along the usual lines: residual equilibrium equations, first order
incremental equations, finding critical points, and so on.
The physical interpretation of the parameter depends on the type of structure. For example, in the
analysis of thin shells whose thickness is accurately controlled a natural choice would be the ratio of
the expected imperfection amplitude to the thickness. If the thickness itself may vary locally about
its nominal value — as it would happen, for example, in reinforced concrete shells — the thickness
variation may be taken as the imperfection parameter. In general we may say that the choice of
parameter is tied up to the fabrication method whereas the value of the parameter is determined by
fabrication or construction quality control. In the case of mass-produced systems imperfection data
is sometimes available from actual field measurements.

§28.5. Imperfection Sensitivity at Critical Points

The effect of imperfections on the load-carrying capacity of a structure that fails at a critical point
may drastically vary according to the type of critical point. Structures whose failure loads are
substantially reduced by imperfections are called imperfection sensitive. In this Chapter we review
the question of sensitivity for the four types of critical points introduced in Chapter 23, using typical
response plots.
In the following two-dimensional response plots, the dimensionless control parameter λ is plotted
against a representative state parameter. This λ is assumed to be a load multiplier or load factor that
characterizes the strength of the structure. Following the same conventions as in the example of
§25.2 heavy lines represent the equilibrium path of theperfect system while light lines represent the
equilibrium paths of imperfect systems. Furthermore continuous lines represent stable equilibrium
path segments whereas broken lines represent unstable equilibrium segments.

28–8
28–9 §28.5 IMPERFECTION SENSITIVITY AT CRITICAL POINTS

λ λcr
>0

<0
B

<0

>0

µ 

Figure 28.7. Effect of initial imperfectons at a asymmetric bifurcation point.

§28.5.1. Limit Point

Figure 28.6 pertains to a limit point. We see that the response of the imperfect system is not
dissimilar from that of the corresponding perfect system. The peak or failure load measured by
λcr varies quasi-linearly with the imperfection parameter  and this variation λcr () is shown in the
right-hand diagram of Figure 28.6. As can be observed the function λcr () has normally a finite
and nonzero slope and exhibits no singular behavior as  → 0. We may characterize a system that
fails at a limit point typified by the response of Figure 28.6 as being mildly imperfection sensitive.

§28.5.2. Asymmetric Bifurcation

Typical pictures for an asymmetric point of bifurcation are shown in Figure 28.7. Now we see that
imperfections play a far more significant role in changing the critical-point response of the system
than in the previous case.
For a small positive1 value of , the system loses stability at a limit point that corresponds to
a drastically reduced value of λ. On the other hand a system with a small negative value of 
apparently exhibits no instability in the vicinity of the bifurcation point and follows a stable rising
path. We might note, however, that this continuously rising path travels a region of metastability
and consequently it may not be reliable in the presence of small dynamic disturbances.
The variation of the failure load factor λcr with the imperfection parameter , shown in the right-hand
disgram, is now of considerable interest. For positive imperfections the function λcr () is locally
parabolic, having no singularities as  → 0 but an infinite slope as shown. Thus there is an extreme
sensitivity to initial positive imperfections. Since systems with negative imperfections display no
local failure loads their “buckling” is characterized by a more rapid growth of the deflections as the

1 Positive in the sense that it reinforces the bifurcation phenomenon.

28–9
Chapter 28: IMPERFECTIONS 28–10

λ
λcr

>0 <0

 = 0

µ 

Figure 28.8. Effect of initial imperfectons at a stable-symmetric bifurcation point.

critical load level of the perfect system is reached. It follows that there is no local branch of the
λcr () curve for  < 0.

§28.5.3. Stable Symmetric Bifurcation

Typical pictures for an stable symmetric point of bifurcation are shown in Figure 28.8. The general
behavior is similar to that encountered in the hinged-cantilever example of §25.2. We see that
imperfections play here a relatively minor role in changing the response of the system. Small
positive and small negative imperfections have similar effects, each yielding a continuously stable
and rising equilibrium paths as shown. Therefore, imperfect systems of this type display no sharp
failure load, “buckling” being simply characterized by a more rapid growth of the deflections as the
critical load of the perfect system is approached.

§28.5.4. Unstable Symmetric Bifurcation

Finally, typical pictures for an unstable symmetric point of bifurcation are shown in Figure 28.9. We
see that imperfections play here a significant role in modifying the behavior of the system although
the effect is not so drastic as in the asymmetric bifurcation case. Small positive and small negative
imperfections have symmetrical effects, each now inducing failure at a limit point that corresponds
to a considerably reduced value of λcr (). The variation of λcr with  is shown in the right-hand
diagram. For both positive and negative imperfections the function follows locally (that is, near
 = 0) the so-called “two-thirds law:”
λ ∝  2/3 , (28.14)

discovered originally by Koiter in the 1940s. This law yields a sharp “cusp“ at  = 0 as shown. We
can summarize this case as being one of high imperfection sensitivity to both positive and negative
imperfections.

28–10
28–11 §28.6 EXTENSIONS: MULTIPLE BIFURCATION, CONTINUOUS SYSTEMS

λ λcr

>0 <0
B

<0 >0

µ 

Figure 28.9. Effect of initial imperfectons at an unstable-symmetric bifurcation point.

§28.6. Extensions: Multiple Bifurcation, Continuous Systems

The preceding discussion, as well as the treatment of Chapter 24, pertains to limit points and isolated
bifurcation points of discrete structural systems. This rises questions as to what happens at multiple
or compound bifurcation points, or critical points of continuous systems.
At a multiple bifurcation point of order k (order being the rank deficiency of the tangent stiffness
matrix there) a second-order analysis similar to that carried out in §27.4 show that one can expect
in general 2k crossing branches. For an isolated bifurcation point k = 1 and 2k = 2 as previously
found. But if k ≥ 2 much greater complexity of behavior can be expected. At a multiple bifurcation
point the effect of initial imperfections is typically far more severe than at single bifurcation points
if one or more of the emanating branches are unstable, as is usually the case. The systematic
investigation of these effects remains a frontier research subject, which is nonetheless gaining in
importance because of its application to stability-optimized aerospace structures that are designed
to simultaneously fail in several buckling modes.
As for continuous systems, the aforementioned two-thirds power law for the imperfection sensitivity
at an unstable-symmetric bifurcation point carries over to continuous systems if the geometric
imperfection is assumed to have the shape of the buckling mode. If this common assumption is not
made new power laws may emerge from the continuous analysis. As for asymmetric buckling, it
remains poorly understood in the continuous case.

28–11
Chapter 28: IMPERFECTIONS 28–12

Homework Exercises for Chapter 28

Imperfections

EXERCISE 28.1 Work out the details of the analysis of the imperfect propped cantilever described in §28.3.
In particular, verify that the diagram shown in Figure 28.5 is correct. Obtain the equation of the imperfection
sensitivity diagram λcr () where λcr are load limits (limit points) obtained when  = 0. Plot this diagram for
values of  = 0 to 1; note vertical-tangent “cusp” at  = 0! [Qualitatively, the diagram should look like the
one in Figure 28.9.]

EXERCISE 28.2 A highly simplified, one-DOF “beer-can-like” structure has the total potential function
   
(u, f, ) = U − V, U = k L 2 ( 1 + (u/L) − 1 − γ )2 , V = f L 1 + (u/L) − 1 − (u/L)2 (E28.1)

where u is the state variable, f is the applied load, k, L are structure property constants with dimensions of
spring constant and length, respectively, γ is a dimensionless constant, and  is a dimensionless geometric
imperfection parameter. Reduce (E28.1) to a dimensionless form

(µ, λ, ) (E28.2)

by defining the state parameter µ = u/L and the control parameter λ = f /(k L), and dividing the whole
thing through by k L 2 . From then on take γ = 0. Form the equilibrium equations and generate a response
diagram for the imperfect structure similar to those shown in Figures 28.2 and 28.5, varying  in the range
(-1,1), µ in the range (-0.99,0.99), and λ in (0,1.5). From visual inspection conclude whether the structure
experiences asymmetric bifurcation (that is, it has a preferred buckling direction) or a symmetric one. Draw
the imperfection sensitivity diagram of λcr versus . What is the load capacity drop for imperfections of
magnitude 0.01, 0.1 and 1?

Using Graphic Tools to Expedite HW

Use of built-in graphic tools such as those provided in Matlab, Mathematica or Maple can speed up signifi-
cantly the generation of response diagrams for Exercises 28.1 and 28.2. For example, an initial version of the
diagram shown in Figure 28.2 was produced by the following Mathematica script:

lam[theta_,eps_]:=(theta-eps)/Sin[theta];
p1=Plot[{lam[theta,0.01],lam[theta,-0.01],lam[theta,0.1],
lam[theta,-0.1],lam[theta,0.2],lam[theta,-0.2],lam[theta,0.5],lam[theta,-0.5]},
{theta,-Pi/1.2,Pi/1.2},PlotRange->{0,2},DisplayFunction->Identity];
p2=Plot[lam[theta,0],{theta,-Pi/1.2,Pi/1.2},PlotRange->{0,2},DisplayFunction->Identity];
p3=Plot[1/Cos[theta],{theta,-1.5,1.5},PlotRange->{0,2},DisplayFunction->Identity];
Show[Graphics[Thickness[0.002]],p1, Graphics[Thickness[0.004]],p2,
Graphics[Thickness[0.004]],Graphics[AbsoluteDashing[{5,5}]],p3,
PlotRange->{0,2},Axes->True,AxesLabel->{"theta","lambda"},
DisplayFunction->$DisplayFunction];
The plot cell was then converted and saved as an Adobe Illustrator 88 file, picked up by Adobe Illustrator 6.0
and “massaged” for bells and wistles such as Greek labels, dashed lines, shading of unstable region, etc.

28–12
 27
.

Nonlinear
Bifurcation
Analysis

27–1
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–2

TABLE OF CONTENTS

Page
§27.1. Introduction 27–3
§27.2. Levels of Bifurcation Analysis 27–3
§27.3. Recapitulation of Governing Equations 27–3
§27.3.1. Residual and Rate Equations . . . . . . . . . . . . 27–3
§27.3.2. Stiffness and Load Rates . . . . . . . . . . . . . 27–5
§27.3.3. Limitations of λ-Parametrized Forms . . . . . . . . . 27–5
§27.4. A Deeper Look at Bifurcation 27–5
§27.4.1. State Decomposition . . . . . . . . . . . . . . 27–6
§27.4.2. Failure of first-order rate equations at bifurcation . . . . . 27–6
§27.5. Branch Analysis of Simple Bifurcation 27–6
§27.5.1. State Decomposition . . . . . . . . . . . . . . 27–6
§27.5.2. Finding σ . . . . . . . . . . . . . . . . . . 27–8
§27.6. The Hinged Cantilever 27–10
§27.6.1. Finding the Critical Point . . . . . . . . . . . . . 27–11
§27.6.2. Branching Analysis . . . . . . . . . . . . . . . 27–11
§27. Exercises . . . . . . . . . . . . . . . . . . . . . . 27–14

27–2
27–3 §27.3 RECAPITULATION OF GOVERNING EQUATIONS

§27.1. Introduction

We initiated our study of stability of conservative systems in Chapters 24-25 by using the simplified
model of linearized prebuckling (LPB). This was followed in Chapter 26 by a qualitative study of
the more general stability model that led to classifying isolated critical points into four types: limit
point, asymmetric bifurcation, stable-symmetric bifurcation, and unstable-symmetric bifurcation.
This classification is especially helpful in understanding the effect of imperfections on stability.
This Chapter presents a more detailed mathematical analysis of the phenomenon of bifurcation by
studying equilibrium branches in the vicinity of an isolated bifurcation point. The topic is covered
under the name nonlinear bifurcation to emphasize that we are dealing with the general case as
opposed to the LPB model. A simple example involving a one-degree of freedom system is then
worked out in some detail. The next Chapter takes up the subject of how physical or numerical
imperfections affect structural behavior as regards both limit and bifurcation points.

§27.2. Levels of Bifurcation Analysis

Nonlinear bifurcation analysis can be carried out at different levels of detail, as demanded by
application needs.1 Four levels of increasing detail are schematized in Figure 27.1, which assume
the occurrence of an isolated bifurcation point B.
1. Locate: find where B occurs while tracing a response. Can be done by monitoring changes
of sign of the determinant of K or equivalently tracing the sign of factorization pivots (See
Chapter 21).
2. Determine subspace: having located B, determine vectors y (particular solution) and z (null
eigenvector for isolated bifurcation) that together with λ form an intrinsic subspace “where
the action is.” Requires a partial eigensolution; more precisely getting the null eigenvector(s).
3. Branching analysis: having located B, and computed y and z, find the directions u̇1 and u̇2 of
tangents to the equilibrium paths (branches) that pass through B. Requires an analysis of the
second order rate equations r̈ = 0.
4. Branch curvature analysis: having located B and determined y, z, u̇1 and u̇2 , find the curvatures
of the equilibrium paths (branches) passing through B. Requires an analysis of the third order
...
equation r = 0.
The information necessary for level 3 is quite difficult to obtainfrom a general purpose finite element
programs, while that needed for level 4 is truly inaccesible. For this reason most FE programs can
provide only levels 1 and 2 on a routine basis. In the present Chapter we study up to level 3
(branching analysis), but the practical difficulties of implementing that level should be kept in
mind.

§27.3. Recapitulation of Governing Equations

Below we recapitulate discrete governing equations derived in Chapters 3 and 4, and introduce
additional nomenclature required for the branching analysis carried out in §27.4.

1 For example in preliminary design only the location of the first bifurcation point would be of interest.

27–3
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–4

Analysis Type Rate Equation Order

B Locate 1 (Factorization)

y
B Determine
90 ◦
1 (Eigensolution)
active subspace
z

u̇2
Branching
B 2 if isolated
u˙ 1 analysis

Branch
B curvature 3 if isolated
analysis

Figure 27.1. The four levels of information for nonlinear bifurcation analysis.

§27.3.1. Residual and Rate Equations

The one-parameter residual equilibrium equations are

r(u, λ) = 0, (27.1)

where λ is the stage control parameter and u is the state vector. Solutions of this equation may be
conveniently represented in parametric form

u = u(t), λ = λ(t), (27.2)

where t is a dimensionless path parameter. Two important special choices for pseudotime are

t = λ, t = s, (27.3)

which leads to the λ-parametrized and arclength forms, respectively.

27–4
27–5 §27.4 A DEEPER LOOK AT BIFURCATION

Rate equations are systems of ordinary differential equations obtained by successive differentiation
of (27.1) with respect to t. Recall that K and q denote the tangent stiffness matrix and incremental
load vector, respectively, whose entries are given by
∂ri ∂ri
Ki j = , qi = . (27.4)
∂u j ∂λ
Using superposed dots to denote t-differentiation we obtain

ṙ = Ku̇ − qλ̇ = 0, (27.5)

r̈ = Kü + K̇u̇ − q̇λ̇ − qλ̈ = 0, (27.6)

... ... ...
r = K u + K̇ü + K̈u̇ − q λ − q̇λ̈ − q̈λ̇ = 0. (27.7)
Eq. (27.5) is a system of first-order rate equations, also called the incremental stiffness equations
or simply the stiffness equations. Eq. (27.6) is a system of second-order rate equations, also called
the stiffness-rate equations. Eq. (27.7) is a system of third-order rate equations. And so on. For
the branching analysis undertaken here we will go up to the second-rate equations (27.6).

§27.3.2. Stiffness and Load Rates

In the second-order system (27.6), the stiffness matrix rate K̇ and incremental load vector rate q̇
may be expressed as linear combinations of u̇ and λ̇:

K̇ = Lu̇ + Nλ̇, q̇ = −(Nu̇ + aλ̇) (27.8)

The entries of these new matrices and vectors are given by

∂ 2 ri ∂ 2 ri ∂ Ki j ∂ 2 ri ∂qi
L i jk = , Ni j = = , ai = = . (27.9)
∂u j ∂u k ∂u j ∂λ ∂λ ∂λ∂λ ∂λ

Remark 27.1. Note that L is a three-dimensional array which may be called a cubic matrix to distinguish it
from an ordinary square matrix. (Also referred to as a third order tensor.) Postmultiplying a cubic matrix by
a vector yields an ordinary matrix. For example Lu̇ is a matrix.

§27.3.3. Limitations of λ-Parametrized Forms

... chooses t = λ, simplifications take place in systems (27.5)-(27.7) because λ̇ = 1 and

If one
λ̈ = λ = 0. Using primes to denote differentiation with respect to λ, the first two rate forms reduce
to
Ku − q = 0, (27.10)
Ku + K u − q = 0, (27.11)
in which
K = Lu + N, q = −(Nu + a). (27.12)
These forms are unsuitable, however, near a bifurcation point, because there the relationship between
u and λ ceases to be unique, and the more general parametrized forms such as (27.6) must be used.

27–5
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–6

§27.4. A Deeper Look at Bifurcation

At critical points K becomes singular and therefore possesses at least one null eigenvector, which
as usual is called z. This eigenvector will be normalized to unit length. As discussed in Chapter 5,
if the eigenvector is orthogonal (non-orthogonal) to the incremental load vector, the critical point
is a bifurcation (limit) point.

§27.4.1. State Decomposition

The conditions for bifurcation may be summarily stated as

Kz = 0, z = zT z = 1, qT z = 0. (27.13)

Because the structural system is assumed to be conservative, K is symmetric. Consequently z is

also a left eigenvector of K = KT . In structural mechanics, the eigenvector z is called a buckling
mode or buckling shape. This term conveys the idea that the structure jumps from a prebuckling
state into the new shape. Although the name is appropriate in the LPB model, we shall see that it
is not necessarily appropriate in the general case.
At the bifurcation point B the state vector u and the control parameter λ assume values u B and
λ B , respectively. As in Chapter 25 we study small deviations of u and λ in the neighborhood of
B. These deviations are denoted by u = u − u B and λ = λ − λ B , respectively. For small
deviations from the bifurcation point the relation between u and λ may be linearized as

u = (σ z + y) λ, (27.14)

where y is the particular solution introduced in §25.2, and σ is the buckling mode amplitude.
Dividing by t and passing to the limit t → 0 we obtain the rate form of the above equation:

u̇ = (σ z + y)λ̇. (27.15)

This decomposition of u̇ in the y, z plane is depicted in Figure 27.2, a duplicate of Figure 25.1.

§27.4.2. Failure of first-order rate equations at bifurcation

At bifurcation points the first-order rate equations (27.5) yield no information on the buckling
mode amplitude. This is worked out in Exercise 25.3, which shows that σ = 0/0 and is therefore
indeterminate. To get deterministic information in the vicinity of a bifurcation point it is necessary
to use information from higher-order rate equations. This is covered in the following subsection
for a isolated (simple, distinct) bifurcation point. For this case the second-order rate equations
(27.6) are usually sufficient.

27–6
27–7 §27.5 BRANCH ANALYSIS OF SIMPLE BIFURCATION

zT y = 0
||z||2 = 1
σ zλ̇ u̇

B y
yλ̇

Figure 27.2. State decomposition at isolated bifurcation point B, depicted in the (y, z) plane.

§27.5. Branch Analysis of Simple Bifurcation

The subsequent analysis assumes that the rank deficiency of K at bifurcation is only one, and so z
is the only null eigenvector. This is called an isolated, simple or distinct bifurcation point. We shall
see that at such points there can be at most two equilibrium paths that intersect at B. Such paths
are called branches.

§27.5.1. State Decomposition

Assume that we have located a bifurcation point B and computed the buckling mode z. Our next
task is to examine the structural behavior in the neighborhood of B. This analysis is important to
answer questions pertaining to the safety of the structure and its sensitivity to imperfections.
We have seen that the state variation rate u̇ from the bifurcation point can be decomposed into
a homogeneous solution component σ z in the buckling mode direction, and a particular solution
component y, which is orthogonal to z:
u̇ = (y + σ z)λ̇, yT z = 0 (27.16)
See the geometric interpretation in Figure 27.2.
The particular solution vector y solves the system
Ky = q, zT y = 0, (27.17)
which is simply the first-order incremental flow equation augmented by a normality constraint.
Imposing this constraint removes the singularity (rank deficiency) of K at B.
Remark 27.2. The homogeneous solution z lies in the null space of K whereas the particular solution y lies in
the range space of K. In more physical terms we may say that y “responds” to the load (the incremental load
vector q) whereas z, like any homogeneous solution, is dictated by the boundary conditions.

27–7
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–8

Branch 2

2
λ

1
nt

ch
ge

an
n
Ta

Br
y

u˙ 2 λ̇ 1
Tangent
B λ̇
u˙ 1

yλ̇
z

Figure 27.3. Intersection of two equilibrium paths at an isolated bifurcation

point B, depicted in the y, z, λ) subspace.

Remark 27.3. The decomposition (27.16) is analogous in many respects to the decomposition of element
motions into purely-deformational and rigid-body, studied in Chapter 10. Here z take the role of rigid body
mode. The decomposition (27.16) is, however, expressed in terms of rates because it is local: it is restricted
to the vicinity of the bifurcation point.

§27.5.2. Finding σ
The first-level information on the equilibrium branches at B is given by their tangents at B. Because
we can obtain y and z from the first-order rate equations, these tangents are fully determined if in
addition we know σ , u̇ and λ̇ at B.
But as noted previously the first-order rate equations (27.5) do not provide information on the
buckling mode amplitude σ . To get that information it is necessary to go to the second-order
system (27.6), which is repeated here for convenience:
Kü + K̇u̇ − qλ̈ − q̇λ̇ = 0 (27.18)
Premultiplying both sides by zT and taking account of the bifurcation conditions (27.13) we get at
B the scalar equation
zT K̇u̇ − zT q̇λ̇ = 0 (27.19)
Replacement of K̇ and q̇ by the expressions (27.8) gives
zT (Lu̇ + Nλ̇)u̇ + zT (Nu̇ + aλ̇)λ̇ = 0. (27.20)
Finally, substitution of u̇ by its homogeneous-plus-particular decomposition (27.16) yields
   
zT L(y + σ z)λ̇ + Nλ̇ (y + σ z)λ̇ + zT N(y + σ z)λ̇ + aλ̇ λ̇ = 0. (27.21)

27–8
27–9 §27.5 BRANCH ANALYSIS OF SIMPLE BIFURCATION

Removing the common differential factor (λ̇)2 and collecting terms in σ we arrive at the quadratic
equation
aσ 2 + 2bσ + c = 0, (27.22)
in which
a = zT Lzz, b = zT [Lzy + Lyz + 2Nz] , c = zT [Lyy + 2Ny + a] . (27.23)
This quadratic equation generally provides two roots: σ1 and σ2 . In what follows we shall assume
that these two roots are real (see Remark 27.5 below).
Substitution of σ1 and σ2 into (27.16) furnishes the branching directions at the bifurcation point:
u̇1 = (y + σ1 z)λ̇, u̇2 = (y + σ2 z)λ̇. (27.24)
These are sketched in Figure 27.3 in the three-dimensional space (y, z, λ) with origin at B. Figure
27.4 projects this picture onto the (y, z) plane for additional clarity.

y
Ta
ng
ent

1
nt
2

ge
Bra

n
nch

Ta
2
1
σ2 zλ̇ σ1 zλ̇ Branch

u̇2 u˙ 1
yλ̇
B z

Figure 27.4. Same as Figure 27.3 but looking down the λ axis onto the (y, z) plane.

The key result of this subsection is that there are at most two branches emanating from a simple
bifurcation point. The classification of such points into asymmetric and symmetric bifurcation
points according to the values of σ1 and σ2 appears in the Exercises. In the following section an
illustrative example is worked out by hand.

27–9
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–10

P=λq
A
θ

L
rigid

;;;;
k

Figure 27.5. The hinged cantilever.

Remark 27.4. If a = 0 one root, say σ2 , becomes infinite while the other is σ1 = −c/(2b), assuming b = 0.
Then u̇2 becomes aligned with z. Only in this case it is justified to call z the “buckling mode.”

Remark 27.5. Intuitively it appears that the two roots of (27.22) must be real. The argument goes as follows:
one of the two branches is supposed to exist since B has been located by hypothesis on the equilibrium path.
Its tangent at B must therefore correspond to one of the roots of (27.22). Since one of the roots is by hypothesis
real, the other must also be real because a, b and c are real coefficients.
This indirect proof is not intellectually satisfying, especially to a mathematician. It would be preferable to
prove the root reality by direct reasoning. However the writer has not been able to find such a proof in the
literature, and personal efforts (one hour trying) have been so far unrewarding.

Remark 27.6. If a = b = c = 0 the second-order rate form (27.6) does not provide any local information
as regards branches at B. Then one must continue to the third order rate form (27.7). This will give a cubic
equation in σ with four real coefficients. Since such an equation can have one or three real roots, things get far
more complicated. If all four coefficients vanish, one must go to the fourth-order rate form, and so on. (For a
mathematician specialized in this kind of analysis, hell is a place where the first one million rate forms yield
no information.)

§27.6. The Hinged Cantilever

The branch analysis technique is illustrated on the hinged-cantilever problem depicted in Figure 27.5. A rigid
rod of length L supported by a torsional spring of stiffness k is axially loaded by a dead force P = λq,
q = k/L. Note that k has the physical dimension of force × length, i.e. of a moment. Hence the definition
P = λk/L renders λ dimensionless, which is convenient for hand analysis.
The dimensionless stage control parameter is λ = P L/k. As state parameter we chose the tilt angle θ as most
appropriate for hand analysis. The total potential energy is
1 
 = U − V = 12 kθ 2 − Pu = 12 kθ 2 − P L(1 − cos θ ) = k 2
θ 2 − λ(1 − cos θ . (27.25)

27–10
27–11 §27.6 THE HINGED CANTILEVER

The equilibrium equation in terms of θ is

∂
r= = k(θ − λ sin θ ) = 0, (27.26)
∂θ
This has the two solutions
θ
θ = 0, λ= , (27.27)
sin θ
which pertain to the primary (vertical or untilted) and secondary (tilted) equilibrium paths, respectively. The
two paths intersect at λ = 1, which is therefore a bifurcation point.

§27.6.1. Finding the Critical Point

The incremental equation in terms of θ is

K θ̇ − q λ̇ = 0, (27.28)
with
∂r ∂r
K = = k(1 − λ cos θ ), q=− = k sin θ. (27.29)
∂θ ∂λ
On the primary path, θ = 0, the stiffness vanishes at

λ = 1, or P = k/L . (27.30)

On this path the stiffness is positive (negative) if λ < 1 (λ > 1), respectively. On the secondary path,
λ = θ/ sin θ, the stiffness is given by  
θ cos θ
K =k 1− , (27.31)
sin θ
which vanishes at θ = 0 because θ/ sin θ → 1 as θ → 0. If θ = 0, K > 0. The various cases as regards the
sign of K are summarized in Figure 27.6. Because K is a scalar, positive and negative values corresponds to
stable and unstable equilibrium, respectively, with neutral stability at B. Stable (unstable) paths are showns
with full (dashed) lines.
It is seen that θ = 0◦ and λ = 1 is the only point at which K vanishes, and consequently is the only critical
point. Let us verify now that the critical point is a bifurcation point. Since the system has only one degree of
freedom, the normalized null eigenvector is simply the scalar z = 1, and the inner product zT q reduces to

zq = q = k sin θ (27.32)

which vanishes at θ = 0◦ . Consequently (λ = 1, θ = 0◦ ) is a bifurcation point.

§27.6.2. Branching Analysis

In this problem the particular solution y vanishes because there is only one degree of freedom. We may
therefore take
θ̇ = σ zλ̇ = σ λ̇ (27.33)
The second-order rate equation is
λ sin θ θ̇ θ̇ − 2 cos θ θ̇ λ̇ = 0,
which upon substituting θ̇ = σ λ̇ yields the quadratic equation (27.22) with a = λ sin θ, b = −2, c = 0. At
the bifurcation point (λ = 1, θ = 0) we get

0.σ 2 − 2σ = −2σ = 0 (27.34)

27–11
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–12

K <0
Secondary path

B K >0
K>0

K =0

K >0
Primary (fundamental) path

R
θ
Figure 27.6. The sign of the stiffness coefficient K for the hinged cantilever response.

The two roots of (27.34) as a quadratic equation are

σ1 = 0, σ2 = ∞, (27.35)

leading to the solutions

θ̇ = 0, λ̇ = 0. (27.36)

These branches are the tangents to the primary (vertical bar) and secondary (tilted bar), respectively, at the
bifurcation point. See Figure 27.7.

This Figure also sketches the post-buckling response, which for this problem is easily obtained from the
exact equilibrium solutions (27.27). According to the qualitative classification of Chapter 26, the bifurcation
point is of stable-symmetric type. This subclassification of a symmetric bifurcation point into stable and
unstable cannot be discerned, however, from the branch-tangent analysis, because it requires information on
the curvature of the z-directed branch.

27–12
27–13 §27.6 THE HINGED CANTILEVER

Tangent 1: θ̇ = 0
Secondary path

Tangent 2: λ̇ = 0

R
θ

Figure 27.7. The two branch directions at bifurcation point of the hinged cantilever.

27–13
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–14

Homework Assignments for Chapter 27

Nonlinear Bifurcation Analysis

EXERCISE 27.1 [A:15] Consider

∂K ∂K ∂q
L= , N= , a=− . (E27.1)
∂u ∂λ ∂λ
Are these relations true?

EXERCISE 27.2 [A:20] If a → 0 in the quadratic equation (27.22) while b = 0, one of the roots, say σ1 , goes
to ∞ whereas the other one becomes σ2 = −c/2b. This is called a symmetric bifurcation. Show that in such
a case the branch direction corresponding to σ1 coincides with the buckling mode z, and draw a bifurcation
diagram similar to Figure 27.1.

EXERCISE 27.3 [A:40] Algebraically prove that the roots of the quadratic equation (27.5) are real2

EXERCISE 27.4 [A:25] The LPB first order rate equations are ṙ = Ku̇ − qλ̇ = 0, in which K = K0 + λK1
and where K0 , K1 and q are constant. Using Exercise 25.3 (posted solution), show that LPB can only predict
symmetric bifurcation. What wonderful thing happens if K1 y = 0?

EXERCISE 27.5 [A:20] The propped cantilever shown in Figure 28.3 consists of a rigid bar of length L
pinned at A and supported by a linear extensional spring of stiffness k. The spring is assumed to be capable
of resisting both tension and compression and retains its horizontal orientation as the system deflects. The bar
may rotate all the way around the pin. The rigid bar is subjected to a vertical dead load P that remains vertical.
Define dimensionless control and state parameters as
P
λ= , µ = sin θ. (E27.2)
kL
Analyze the stability of the propped cantilever in a manner similar to §27.5. Show that the secondary equi-
librium path is the circle λ2 + µ2 = 1 and sketch the response paths showing the complete circle. From this
diagram, can you tell whether the bifurcation point at λ = 1 is stable-symmetric or unstable-symmetric? How
about the one at λ = −1?

2 A very difficult assignment worth of a paper. I am not aware of anybody that has done for the general case.

27–14
 26
.

Qualitative
Analysis of
Critical Points

26–1
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–2

TABLE OF CONTENTS

Page
§26.1. General Notion of Stability 26–3
§26.2. Stability of a Discrete Conservative System 26–3
§26.3. Stability Transformation at a Limit Point 26–4
§26.4. Stability Exchange at Bifurcation Points 26–7
§26.4.1. Asymmetric bifurcation . . . . . . . . . . . . . . 26–7
§26.4.2. Stable-symmetric bifurcation . . . . . . . . . . . . 26–8
§26.4.3. Unstable-symmetric bifurcation . . . . . . . . . . . 26–8

26–2
 26–3 §26.2 STABILITY OF A DISCRETE CONSERVATIVE SYSTEM

§26.1. General Notion of Stability

In Chapter 24 stability was informally defined as the ability of a physical system to return to
equilibrium when disturbed. If the equilibrium is static in nature, we speak of static stability. For
a more precise definition concerning a mechanical system, let’s hear Dirichlet:1

“The equilibrium [of a mechanical system] is stable if, in displacing the points of the system from
their equilibrium positions by an infinitesimal amount and giving each one a small initial velocity,
the displacements of different points of the system remain, throughout the course of the motion,
contained within small prescribed limits”

Some essential ingredients of this definition are:

(1) Stability is a quality of one solution — an equilibrium solution of the system.

(2) The problem of ascertaining the stability of a solution concerns the “neighborhood”
of the particular solution and is therefore a local one.
(3) The concept of stability is inherently dynamic in nature. But for a conservative
system dynamics can be “factored out” of the problem, and we are left with a static
criterion.

§26.2. Stability of a Discrete Conservative System

As discussed in Chapter 24, the static stability of a conservative mechanical system can be tested
completely using a static criterion. Such criterion, often referred to as the Euler stability test, the
energy test, or the method of adjacent states, relates to the positive definiteness character of the
second variation of the potential energy.
We know that a stationary value of the total potential energy with respect to the state variables is
necessary and sufficient for the equilibrium of the system. Proceeding one step further, a complete
relative minimum of the total potential energy is necessary and sufficient for the stability of an
equilibrium state.
For a discrete system with a finite number of degrees of freedom the criterion can be enunciated in
terms of the definiteness of the tangent stiffness matrix K if all state variables are of displacement
type, which we assume in the sequel (see Remark below). For a conservative system we know that
K is a symmetric matrix.

1 As it appears in his Appendix to the German translation of Lagrange’s Mécanique Analytique (1853).

26–3
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–4

L (u, λ2 ), λ2 > λcr

(u, λcr )
(u, λ1 ), λ1 < λcr
λcr λ

Primary (fundamental) path

u

Figure 26.1. Transformation of potential energy at a limit point.

Remark 26.1. The restriction to systems with displacement state variables aims to exclude those in which
Lagrange multipliers are carried along as degrees of freedom. For such systems the criterion applies upon
eliminating the multipliers, but such elimination is often messy and would complicate the exposition.

The stability criterion for a conservative system is summarized in the following table.

If K evaluated at an The potential energy  Then the equilibrium

equilibrium position is at that position has a position is

positive definite strict minimum stable

positive semidefinite cylindrical or inflexion point neutrally stable
indefinite saddle point unstable

If the eigenvalues of K are easily available a test for stability is immediate.2 If all eigenvalues
are greater than zero, the matrix is positive definite and the equilibrium is stable. If one or more
eigenvalues are zero and the rest positive, the equilibrium is neutrally stable. If one or more
eigenvalues are negative, the equilibrium is unstable.
In practice an eigenvalue test can be recommended only for small matrices, say of order less than 20
or so. For larger matrices the same information can be obtained more economically by decomposing
K using triangular factorization or Gauss elimination, as discussed in Remark 24.2. If all pivots
are positive, the equilibrium is stable. If at least one pivot is negative, the equilibrium is unstable.
The border case of neutral stability is more difficult to detect in the presence of rounding errors.

2 See §24.2 for computational details.

26–4
26–5 §26.3 STABILITY TRANSFORMATION AT A LIMIT POINT

Primary (fundamental) equilibrium path

L1

( u1 , u 2 , λ1 )

u2 L2

u1

Figure 26.2. Typical potential energy surface in snap-through response.

§26.3. Stability Transformation at a Limit Point

As discussed in previous sections, there is a close relationship between equilibrium configurations,

occurrence of critical points, and the stability of the system. We now examine qualitatively, follow-
ing the classical treatise of Thompson and Hunt,3 four types of critical points from the standpoint
of the variation of the total potential energy in the neighborhood of equilibrium states. This is
done with the typical response plots in which the control parameter λ is the vertical axis while a
representative displacement u or deformation mode amplitude is shown along the horizontal axis
as state parameter.
Drawing conventions are as follows: heavy lines represent equilibrium paths, continuous lines
denoting stable paths while broken lines denote unstable paths. Plots of total potential energy
(u, λ) at various fixed values of λ are shown as “shaded profile” energy surfaces. These surfaces
deform as the parameter λ changes. Equilibrium configurations correspond to stationary points of
 with respect to u. Strong minima (maxima) of this surface are associated with stable (unstable)
equilibrium configurations.
We first consider the case of a limit point, which is shown in Figure 26.1. The fundamental
equilibrium path that starts from the origin (the reference configuration u = 0, λ = 0) is initially
stable. Stability is lost when the local maximum at λ = λcr is reached. λ = λcr . A “snap-through”
response of this form is characteristic of shallow arches and domes. At a fixed value λ = λ1 less
than λcr the total potential energy (u, λ1 ) has a minimum with respect to the state parameter
u on the stable rising region of the path and a maximum on the unstable falling region. As the

3 See References in Appendix Z

26–5
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–6

Secondary path

(u, λ2 ), λ2 > λcr

B

(u, λcr )
λcr λ (u, λ1 ), λ1 < λcr

Primary (fundamental) path

u

Figure 26.3. Asymmetric bifurcation point.

prescribed value of λ is increased the maximum and minimum approach each other and coalesce
when λ = λcr . At this critical point the total potential energy (u, λ) has an horizontal point of
inflexion. At a higher value of λ, say λ2 > λcr , there are no local equilibrium states and the total
potential energy (u, λ2 ) has no stationary point. The critical equilibrium state is seen to be itself
unstable, and the absence of local equilibrium states at values of λ greater than λcr implies that a
physical system under slowly increasing λ will eventually snap-through dynamically. Limit points
are generally insensitive to imperfections.
A more general schematic diagram is shown in three dimensions in Figure 26.2 on a plot of λ
against two state parameters, v1 and v2 . This plot includes a remote rising region of the equilibrium
path since this is often encountered with this type of behavior. A total potential energy surface
(v1 , v2 , λ) is drawn for a fixed value of λ < λcr . As λ is slowly increased through its critical
value the system will “snap through” dynamically, eventually stop, and initiate a large amplitude,
nonlinear vibration about the remote stable equilibrium path. In the presence of some damping the
system will eventually rest on that path.
These figures illustrate the physics well but if we are dealing with a system with many degrees of
freedom care must be taken in drawing conclusions from these schematic figures. On an actual plot
of λ against one of the vi ’s, the limit point is normally seen as a smooth maximum, but it must be
realized that for a certain choice of the state parameter vi the point might appear as a sharp cusp.
The smooth maximum of a path in three-dimensional space can for example be seen as a cusp if
the eye is directed along the horizontal tangent to the path.

26–6
26–7 §26.4 STABILITY EXCHANGE AT BIFURCATION POINTS

Secondary path
(u, λ2 ), λ2 > λcr

(u, λcr )
B
(u, λ1 ), λ1 < λcr

λcr
λ
Primary (fundamental) path

Figure 26.4. Stable-symmetric bifurcation point.

§26.4. Stability Exchange at Bifurcation Points

After the limit point we consider bifurcation or branching points. We cover the three most common
types of bifurcation: asymmetric, stable-symmetric, and unstable symmetric.

§26.4.1. Asymmetric bifurcation

Figure 26.3 shows the case of an asymmetric point of bifurcation. The initially stable fundamental
equilibrium path that emanates from the origin loses its stability on intersecting a distinct and
continuous secondary (post-buckling) equilibrium path. The intersection point B is a critical point
of bifurcation type. An asymmetric bifurcation point is characterized by the fact that both paths
have a nonzero slope with respect to λ at B.
With varying λ the paths exhibit a phenomenon called exchange of stability. For λ1 < λcr the total
potential energy (u, λ1 ) has a minimum with respect to u on the stable region of the fundamental
path and a maximum with respect to u on the unstable region of the post-buckling path. As λ is
increased the maximum and minimum finally coalesce so that at λ = λcr the total potential energy
(u, λcr ) has a horizontal point of inflexion at the critical equilibrium state. At λ values over the
critical one the maximum and minimum exchange places. Since an unstable branch emanates from
B, the critical equilibrium state is unstable. In the presence of small disturbances a physical system
under slowly increasing λ would snap dynamically from this critical equilibrium state despite the
existence of stable equilibrium states at higher values of λ.
Critical points of this type are moderately to highly sensitive to the presence of loading or fabrication
imperfections.

26–7
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–8

(u, λ2 ), λ2 > λcr

B

(u , λcr )

(u , λ1 ), λ1 < λcr

λcr
λ Secondary path

Primary (fundamental) path

Figure 26.5. Unstable-symmetric bifurcation point.

§26.4.2. Stable-symmetric bifurcation

Symmetric bifurcation points are characterized by the fact that the intersecting path has zero slope
with respect to the control parameter at B. These points may be categorized into stable and unstable,
depending on whether the intersecting post-buckling) path is “rising” or “falling”.
Figure 26.4 depicts the case of an stable-symmetric point of bifurcation. Here a fundamental
equilibrium path rising monotonically from the reference state is seen to intersect a stable rising
secondary (post-buckling) path that passes smoothly through the critical equilibrium state with zero
slope. The continuation of the fundamental path beyond B is unstable. The total potential energy
(u, λ1 ), where λ1 < λcr , has a single stationary value with respect to u, namely the minimum
on the stable region of the fundamental path, and as the value of λ is increased this minimum is
transformed into two minima and one maximum. The critical equilibrium state is neutrally stable
and the secondary path is stable, so a physical system under slowly increasing λ would exhibit no
dynamic snap but would follow the stable rising post-buckling path, the direction taken depending
on the small disturbances or imperfections which are inevitably present.
Critical points of this type are insensitive to the presence of imperfections.

§26.4.3. Unstable-symmetric bifurcation

The last configuration examined here is the unstable-symmetric point of bifurcation, shown in
Figure 26.5. Here the fundamental path intersects an unstable falling path which as in the previous
case has a zero slope at the critical equilibrium state. At a prescribed value of λ = λ1 < λcr
the total potential energy (u, λ1 ) has now three stationary values with respect to u, namely two

26–8
26–9 §26.4 STABILITY EXCHANGE AT BIFURCATION POINTS

maxima on the unstable post-buckling or secondary path, and a minimum on the stable region of
the fundamental path. As the figure shows, these three stationary points transform into a single
maximum with increasing λ. The critical equilibrium state is seen to be unstable, so a physical
system would snap dynamically from the critical equilibrium state, the direction taken depending
on the postulated small disturbances or imperfections.
Critical points of this type are highly sensitive to the presence of structural or loading imperfections.
Sometimes the sensitivity is extreme, as in the classical case of the axially compressed cylindrical
shell discussed in §25.4.

26–9
 25
.

Bifurcation:
Linearized
Prebuckling II

25–1
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–2

TABLE OF CONTENTS

Page
§25.1. Introduction 25–3
§25.2. State Decomposition at Bifurcation Point 25–3
§25.3. LPB Assumptions 25–3
§25.4. Limitations of LPB 25–5
§25.4.1. When LPB Works . . . . . . . . . . . . . . . . 25–6
§25.4.2. And When It Doesn’t . . . . . . . . . . . . . . 25–7
§25.4.3. How to Extend the Applicability of LPB . . . . . . . . 25–8
§25. Exercises . . . . . . . . . . . . . . . . . . . . . . 25–9

25–2
25–3 §25.3 LPB ASSUMPTIONS

§25.1. Introduction

This Chapter continues with the subject of linearized prebuckling (LPB) bifurcation analysis. It
goes deeper than Chapter 24 in that it probes the assumptions (so far stated without proof) behind
LPB, and the practical modeling implications that emanate from these assumptions.
To present some of the derivations in mathematical terms it is necessary to introduce the concept
of state decomposition at the bifurcation point and to define homogeneous and particular solutions.
This is done briefly in §24.2 primarily as a means of introducing notation for S24.3 and following.
The detailed mathematical analysis of this decomposition is relegated to Chapter 26.

§25.2. State Decomposition at Bifurcation Point

Recall from previous Chapters that an isolated bifurcation point at λcr is characterized by a singular
tangent stiffness at the equilibrium configuration,

K(ucr , λcr ) z = 0, (25.1)

and by the normalized null eigenvector (buckling mode) z = 0, z = 1, being orthogonal to the
incremental load vector:
qT z = zT q = 0. (25.2)
Assume that we have located a bifurcation point B and computed the buckling mode z. Our next
task is to examine the structural behavior in the neighborhood of B. We shall be content with looking
at the so-called branching direction information. This information characterizes the tangents to the
equilibrium branches that cross at B.
To carry out this task we borrow from algebraic ODE theory. Consider the variation in the state
vector u measured from its value u B at buckling:

u = u − u B (25.3)

Divide this increment by t, t being the timelike parameter introduced in Chapter 3, and pass to
the limit:
u
u̇ = lim . (25.4)
t→0 t

This variation rate u̇ from the bifurcation point can be decomposed into a homogeneous solution
component σ z in the buckling mode direction, and a particular solution component y, which is
orthogonal to z:
u̇ = (y + σ z)λ̇, yT z = zT y = 0, (25.5)
The particular solution solves the system

Ky = q, yT z = 0, (25.6)

which is simply the first-order incremental equation Ku̇ = qλ̇ augmented by a normality con-
straint. Imposing this constraint removes the singularity (rank deficiency) of K. The geometric
interpretation of this decomposition on the y, z plane is shown in Figure 24.1.

25–3
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–4

zT y = 0
||z||2 = 1
σ zλ̇

B y
yλ̇

Figure 25.1. State decomposition at bifurcation point B.

§25.3. LPB Assumptions

With the notation introduced in §25.2 we may now state the key assumptions invoked in linearized
prebuckling (LPB). (Those collected in item (II) have already been formally stated and used in
Chapter 21.)
(I) The loading is conservative and proportional:

p = q0 + λq. (25.7)

and the structure is linearly elastic. Inother words, the residual equations are derivable from
a potential energy function.
(II) The displacements and displacement gradients prior to the critical state are negligible in the
sense that (a) the material stiffness matrix can be evaluated at the reference configuration,
and (b) the geometric stiffness is proportional to the control parameter λ:

K M ≡ K0 , KG ≡ λK1 , (25.8)

in which K0 is the material matrix evaluated at the reference configuration, also called the
linear stiffness, and K1 is the reference geometric stiffness. As discussed in the previous
Chapter the singular stiffness criterion det K = 0 leads to the eigenproblem

(K0 + λK1 ) z = 0. (25.9)

(III) The particular solution y defined in §25.2 is obtained by solving

(K0 + λcr K1 ) y = q (25.10)

25–4
25–5 §25.4 LIMITATIONS OF LPB

under the constraint yT z = 0. Observe that from assumption (I) q is constant.

We now prove that if these assumptions hold, all critical points determined from the LPB eigenprob-
lem are bifurcation points, that is, zT q vanishes. To show that, premultiply both sides of (25.10)
by zT :
zT q = zT (K0 + λK1 )y = yT (K0 + λK1 )z = yT (Kz) = 0 (25.11)
Note that the transformation zT Ky = yT Kz holds because K0 and K1 are symmetric on account of
the conservativeness assumption (I).
Remark 25.1. Bifurcation points are classified in later sections into various types: unsymmetric, stable-
symmetric, stable-unsymmetric, and so on. It will be shown later that, under most common assumptions, LPB
bifurcation points are generally of symmetric type. The LPB model does not provide, however, information
as to the post-bifurcation stability, so we cannot say whether the bifurcation point is stable-symmetric or
unstable-symmetric.

§25.4. Limitations of LPB

Linearized prebuckling (LPB) is used extensively in engineering design. Standard books in struc-
tural stability1 concentrate upon it. In its finite element version LPB is a feature available in many
finite element programs. Exercising this feature has the advantages of avoiding a full nonlinear
analysis, which can be expensive and time-consuming. Given its practical importance, structure de-
signers (and most especially aerospace designers) should be familiar with the range of applicability
of LPB. The limitations are discussed next.
1. Conservative loading. LPB is a restricted form of the static criterion also known as Euler’s
test (see §24.2). If the loads are not conservative, the dynamic criterion should be used, at
least to check out whether a flutter condition may occur. If the dynamic criterion shows that
stability is lost by divergence, one may regress to the singular-stiffness test criterion.
2. Loss of stability must be by symmetric bifurcation. If the first critical point is a limit point or
asymmetric bifurcation,2 LPB is not strictly applicable although in some cases it may provide
a sufficiently good approximation. Lacking experimental confirmation or a priori knowledge,
the only practical way to check whether the first critical point is symmetric bifurcation is to
go through a full nonlinear analysis.
3. Prebuckling deformations must be small. This assumption fits well many engineering struc-
tures because of the nature of construction materials. The structures that best fit these assump-
tions are straight columns, frameworks and flat plates, as illustrated in Figure 25.2. Care must
be exercised for arches, shells, very thin members, and for imperfection-sensitive structures
in general.
4. Elastic material behavior. If the material is inelastic the structure is not internally conservative.
Then the tangent stiffness depends on the prior deformation history, and the LPB eigenproblem

1 For example, Timoshenko and Gere’s Theory of Elastic Stability.

2 Symmetric bifurcation occurs when bucking in the z and −z directions is equally likely. Asymmetric bifurcation occurs
when one of the directions is physically more likely; for example axially compressed cylinders buckle inwards. This
classification of critical points is covered in more detail in Chapter 11 and following.

25–5
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–6

;;;; ;; Figure 25.2. Structures that are adequately modeled by LPB assumptions.

negligible deformation
prior to buckling

R v

Figure 25.3. Type of response expected under LPB assumptions.

(Branch intersection at B not shown for clarity)

loses meaning. The topic of inelastic buckling (in particular creep and plastic buckling) is an
enormous subject that falls outside the scope of this course.

5. Applied loads should not depend nonlinearly on the displacements. Such a dependence usually
introduces nonconservative effects, thus voiding the conservative-loading assumptions. Even
is the loads remain conservative, the reference geometric stiffness would depend on the load
level, thus leading to a nonlinear eigenproblem.

6. The effect of imperfections is negligible. Some structures are highly imperfection sensitive in
that the first critical load is strongly affected by the presence of imperfections. In such cases
obviously LPB is of limited value or outright irrelevant.

§25.4.1. When LPB Works

25–6
25–7 §25.4 LIMITATIONS OF LPB

(a) (b)

λ L λ B

LPB

LPB Actual
B L Actual

R v R v

Figure 25.4. Two structures that fit the LPB assumptions poorly.

The systems that best fit the LPB model are symmetrically loaded structures such as straight columns
and in-plane-loaded plates (laminas) which are not excessively thin. See Figure 25.2. The lateral
buckling of such structures occurs following very small deformations, as typified by the response
sketch in Figure 25.3.

§25.4.2. And When It Doesn’t

Two examples of structures that are not properly treated by the LPB model are shown in Figure
25.4. The LPB predictions are way off in both cases, but for different reasons.
Case (a) is an axially compressed cylindrical shell made up of almost flat panels joined by curved
panels, forming like a “curved triangle” cross section seen in some combat helicopters and the
Space Shuttle fuselage. There is a substantial redistribution of stresses due to changes on geometry.
The structure eventually collapses at a limit point substantially over the predicted LPB load. The
latter is therefore overly safe.

25–7
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–8

On the other hand, the axially compressed circular cylinder of case (b) is highly imperfection-
sensitive structure that fails at a substantially lower load than that predicted by LPB. Consequently
the LPB prediction is highly unsafe.

§25.4.3. How to Extend the Applicability of LPB

One way to broaden the application of the LPB model is to update the reference configuration3 so
that the prebuckling deformations are reduced. If this is done the control parameter λ is of course
measured from the latest reference configuration and consequently becomes a true stage control
parameter. Limitations on the conservativeness of applied loads and types of critical point, however,
cannot be readily circumvented by this “staging” technique.

3 As naturally done in the CR description, in which the deformational displacements are measured from a continuously
varying configuration, and also in the Updated Lagrangian description.

25–8
25–9 Exercises

Homework Exercises for Chapter 25

Bifurcation: Linearized Prebuckling II

EXERCISE 25.1 [A:15] Find the particular solution y at the lowest bifurcation load of the two-bar example
of Chapter 24.

EXERCISE 25.2 [A:15] Find the particular solution y at the symmetric and antisymmetric bifurcation loads
of the one-element Euler column example of Exercise 24.4.

EXERCISE 25.3 [A:25] The first order residual rate equations is ṙ = 0, where ṙ is given by

ṙ = Ku̇ − qλ̇ = 0, (E25.1)

(E25.1) holds at a bifurcation point where K and q are the tangent stiffness matrix and incremental load vector,
respectively, at bifurcation. Decompose u̇ = (y + σ z)λ̇, where y is the particular solution and z = 0 the
buckling mode normalized to length one. Show that the first-order differential equation system (E25.1) cannot
give information on the “buckling mode amplitude” σ because one gets σ = 0/0. (Hint: premultiply that
equation by an appropriate vector.)

25–9
 24
.

Bifurcation:
Linearized
Prebuckling I

24–1
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–2

TABLE OF CONTENTS

Page
§24.1. Introduction 24–3
§24.2. Loss of Stability Criteria 24–4
§24.2.1. Static criterion . . . . . . . . . . . . . . . . . 24–4
§24.2.2. Dynamic criterion . . . . . . . . . . . . . . . 24–4
§24.3. The Tangent Stiffness Test 24–5
§24.4. Linearized Prebuckling 24–6
§24.5. The LPB Eigensystem 24–7
§24.6. Solving the Stability Eigenproblem 24–7
§24.7. LPB Analysis Example 24–8
§24.8. Summary of LPB Steps 24–11
§24. Exercises . . . . . . . . . . . . . . . . . . . . . . 24–12

24–2
24–3 §24.1 INTRODUCTION

§24.1. Introduction

This Chapter starts a systematic study of the stability of elastic structures. We shall postpone the
more rigorously mathematical definition of stability (or lack thereof) until later because the concept
is essentially dynamic in nature. For the moment the following physically intuitive concept should
suffice:

“A structure is stable at an equilibrium position if it returns to that position

upon being disturbed by an extraneous action”

Note that this informal definition is dynamic in nature, because the words “returns” and “upon”
convey a sense of history. But it does not imply that the inertial and damping effects of true dynamics
are involved. So real time is not involved in the static case.
A structure that is initially stable may lose stability as it moves to another equilibrium position
when the control parameter(s) change. Under certain conditions, that transition is associated with
the occurrence of a critical point. These have been classified into limit points and bifurcation points
in Chapter 5.
For the slender structures that occur in aerospace, civil and mechanical engineering, bifurcation
points are more practically important than limit points. Consequently, attention will be initially
directed to the phenomena of bifurcation or branching of equilibrium states, a set of phenomena
also informally known as buckling. The analysis of what happens to the structure after it crosses a
bifurcation point is called post-buckling analysis.
The study of bifurcation and post-buckling while carrying out a full nonlinear analysis is a math-
ematically demanding subject. But in important cases the loss of stability of a geometrically
nonlinear structure by bifurcation can be assessed by solving linear algebraic eigenvalue problems
or “eigenproblems” for short. This eigenanalysis provides the magnitude of the loads (or, more
generally, of the control parameters) at which buckling is expected to occur. The analysis yields no
information on post-buckling behavior. Information on the buckling load levels is often sufficient,
however, for design purposes.
The present Chapter covers the source of such eigenproblems for conservatively loaded elastic
structures. Chapters 26 through 28 discuss stability in the context of full nonlinear analysis. The
two final Chapters (29–30) extend these concepts to structures under nonconservative loading.
Following a brief review of the stability assessment criteria the singular-stiffness test is described.
Attention is then focused on the particular form of this test that is most used in engineering practice:
the linearized prebuckling (LPB) analysis. The associated buckling eigenproblem is formulated.
The application of LPB on a simple problem is worked out using the bar element developed in
the previous three sections. The assumptions underlying LPB and its range of applicability are
discussed in the next Chapter.

24–3
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–4

§24.2. Loss of Stability Criteria

For elastic, geometrically-nonlinear structures under static loading we can distinguish the following
techniques for stability assessment.
 
 Static criterion (Euler method): singular stiffness

 Conservative

 Dynamic criterion: zero frequency
Loading  

 zero frequency (divergence)


 Nonconservative Dynamic criterion
frequency coalescence (flutter)

§24.2.1. Static criterion

The static criterion is also known as Euler’s method, since Euler introduced it in his famous
investigations of the elastica published in 1744. Other names for it are energy method and method
of adjacent states. To apply this criterion we look at admissible static perturbations of an equilibrium
position1 . These perturbations generate adjacent states or configurations, which are not generally
in equilibrium.
Stability is assessed by comparing the potential energy of these adjacent configurations with that
of the equilibrium position. If all adjacent states have a higher potential energy, the equilibrium is
stable. If at least one state has a lower (equal) potential energy the equilibrium is unstable (neutrally
stable). This comparison can be expressed in terms of the second variation of the potential energy
and hence can be reduced to the assessment of the positive definite character of the tangent stiffness
matrix.
Although stability is a dynamic phenomenon, no true-dynamics concepts such as mass or damping
are involved in the application of the static criterion, which is a key reason for its popularity. But
the reasoning behind it makes it strictly applicable only to conservatively loaded systems, because
a load potential function is assumed to exist.

§24.2.2. Dynamic criterion

The dynamic criterion looks at dynamic perturbations of the static equilibrium position. In informal
terms, “give the structure a (little) kick and see how it moves.” More precisely, we consider small
oscillations about the equilibrium position, and pose an eigenproblem that determines characteristic
exponents and associated eigenmodes. The characteristic exponents are generally complex num-
bers. If all characteristic exponents have no positive real components the equilibrium is dynamically
stable, and unstable otherwise.
These exponents change as the control parameter λ is varied. For sufficiently small values the
structure is stable. Loss of stability occurs when a characteristic exponent enters the right-hand
complex plane. If that happens, the associated mode viewed as a displacement pattern will amplify
exponentially in the course of time. A deeper study of the stable-to-unstable transition mechanism

1 “Admissible” in the sense of the Principle of Virtual Work: variations of the state parameters that are consistent with the
essential boundary conditions (kinematic constraints)

24–4
24–5 §24.3 THE TANGENT STIFFNESS TEST

reveals two types of instability phenomena, which are associated with the physically-oriented terms
terms divergence and flutter.
Divergence occurs when the characteristic exponent enters the right-hand plane through the origin,
and it can therefore be correlated with the zero frequency test and the singular stiffness test.
The dynamic criterion is applicable to both conservative and nonconservative systems. This wider
range of application is counterbalanced by the need of incorporating additional information (mass
and possibly damping) into the problem. Furthermore, unsymmetric eigenproblems arise in the
nonconservative case, and these are the source of many computational difficulties.
§24.3. The Tangent Stiffness Test
The stability of conservative systems can be assessed by looking at the spectrum2 of the tangent
stiffness matrix K. Let µi denote the i th eigenvalue of K. The set of µi ’s are the solution of the
algebraic eigenproblem
Kzi = µi zi . (24.1)
Since K is real symmetric3 all of its eigenvalues are real. Thus we can administer the following
test:

(I) If all µi > 0 the equilibrium position is strongly stable

(II) If all µi ≥ 0 the equilibrium position is neutrally stable
(III) If some µi < 0 the equilibrium position is unstable

In engineering applications one is especially interested in the behavior of the structure as the stage
control parameter λ is varied, and so
K = K(λ). (24.2)
Given this dependence, a key information is the transition from stability to instability at the value
of λ closest to stage start, which is usually λ = 0. This is called the critical value of λ, which we
shall denote as λcr .
If the entries of K depend continuously on λ the eigenvalues of K also depend continuously4 on
λ, although the dependence is not necessarily continuously differentiable. It follows that transition
from strong stability — case (I) — to instability — case (III) — has to go through case (II), i.e. a
zero eigenvalue. Thus a necessary condition is that K be singular, that is
det K(λcr ) = 0, (24.3)

or, equivalently,
K(ucr , λcr )z = 0, (24.4)
where z = 0 is the buckling mode introduced in Chapters 4–5, where it was called a null eigenvector.
Equation (24.3) or (24.4) is the expression of the static test for finding a stability boundary.

2 The spectrum of a matrix is the set of its eigenvalues.

3 Because K = ∂ 2 /∂u∂u is the Hessian of the total potential energy .
4 Continuous dependence of eigenvalues on the entries is guaranteed by the perturbation theory for symmetric and Hermitian
matrices. This continuous dependence does not hold, however, for eigenvectors.

24–5
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–6

Remark 24.1. Equation (24.3) is a nonlinear eigenvalue problem because: (a) K has to be evaluated at an
equilibrium position, and (b) K is a nonlinear function of u, which in turn is a nonlinear function of λ as defined
by the equilibrium path. It follows that in general a complete response analysis has to be conducted to solve
(24.3). Such techniques were called “indirect methods” in the context of critical point location methods in
Chapter 23. This involves evaluating K at each computed equilibrium position, and then finding the spectrum
of K. An analysis of this nature is obviously computationally expensive. One way of reducing part of the cost
is noted in the following remark.

Remark 24.2. If K is known at a given λ, an explicit solution of the eigenproblem (24.1) is not necessary for
assessing stability. It is sufficient to factor K as

K = LDLT (24.5)

where L is unit lower triangular and D is diagonal. The number of negative eigenvalues of K is equal to
the number of negative diagonal elements (“pivots”) of D. Matrix factorization is considerably cheaper than
carrying out a complete eigenanalysis because sparseness can be exploited more effectively.

Remark 24.3. The condition (24.3) is not sufficient for concluding that a system that is stable for λ < λcr will go
unstable as λ exceeds λcr . A counterexample is provided by the stable-symmetric bifurcation point discussed
in later Chapters. The Euler column furnishes a classical example. At such points (24.3) holds implying
neutral stability but the system does not lose stability as the bifurcation state is traversed. Nonetheless the
displacements may become so large that the structure is practically rendered useless.

§24.4. Linearized Prebuckling

We investigate now the first critical state of an elastic system if the change in geometry prior to it
can be neglected. We shall see that in this case the nonlinear equilibrium equations can be partly
linearized, a process that leads to the classical stability eigenproblem or buckling eigenproblem.
The eigenstability analysis procedure that neglects prebuckling displacements is known as lin-
earized prebuckling (LPB). The modeling assumptions that are tacitly or explicitly made in LPB
are discussed in some detail in the next Chapter, as well as the practical limitations that emanate
from these assumptions. In the present Chapter we discuss the formulation of the LPB eigenproblem
and illustrate these techniques on a simple problem using the bar elements developed in previous
Chapters.

24–6
24–7 §24.6 SOLVING THE STABILITY EIGENPROBLEM

§24.5. The LPB Eigensystem

The two key results from the LPB assumptions (which are studied in the next Chapter) can be sum-
marized as follows. Recall from Chapters 8–10 that the tangent stiffness matrix can be decomposed
as the sum of material and geometric stiffness matrices:

K = K M + KG . (24.6)

Then the LPB leads to the following simplifications:

(1) The material stiffness is the stiffness evaluated at the reference configuration:

K M = K0 . (24.7)

(2) The geometric stiffness is linearly dependent on the control parameter λ:

KG = λK1 . (24.8)

where K1 is constant and also evaluated at the reference configuration.

Now the stability test (24.3) requires that K be singular, which leads to the stability eigenproblem

Kz = (K0 + λK1 ) z = 0. (24.9)

In the following Chapter we shall prove that under certain restrictions the critical states deter-
mined from this eigenproblem are bifurcation points and not limit points. That is, they satisfy the
orthogonality test
zT q = 0. (24.10)

The eigenproblem (24.9) befits the generalized symmetric algebraic eigenproblem

Ax = λBx, (24.11)

where both matrices A ≡ K0 and B ≡ −K1 are real symmetric, and x ≡ z are the buckling mode
eigenvectors. If (as usual) the material stiffness K0 is positive definite, eigensystem theory says
that all eigenvalues of (24.11) are real. We cannot in general make statements, however, about the
sign of these eigenvalues. That will depend on the physics of the problem as well as on the sign
conventions chosen for the control parameter(s).

24–7
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–8

p = λq

Y, y k (1)
2 (2)

L (2)
0
;; k (2) 3

L (1
0
)
(1)

;;
X, x k (2) << k (1)

;;
1

Figure 24.1. LPB example involving two bar elements displacing on the X ≡ x, Y ≡ y plane.

§24.6. Solving the Stability Eigenproblem

In production FEM codes, the stability eigenproblem (24.9) is generally treated with special solution
techniques that take full advantage of the sparsity of both K0 and K1 , such as subspace iteration or
Lanczos methods.
For small systems an expedient solution method consists of reducing it to canonical form by
premultiplying both sides by the inverse of K0 . This is possible if K0 is nonsingular, which means
that the λ = 0 configuration is not a critical one. Calling A = K−10 K1 and µ = −1/λ one gets

Azi = µi zi (24.12)
This is a standard algebraic eiegnproblem, which can be solved by standard library routines for the
eigenvalues µi and eigenvectors zi . For example, EigenSystem in Mathematica or Eig in Matlab.
The µi farthest away from zero gives the λi = −1/µi closest to zero.
One disadvantage of this reduction is that A is unsymmetric even if K0 and K1 are. There are
more complicated reduction methods that preserve symmetry. These may be studied in standard
numerical analysis textbooks covering linear algebra; for example Golub and Van Loan.
§24.7. LPB Analysis Example
To illustrate the application of LPB to a very simple example, the 2-bar assembly shown in Figure 24.1 is chosen.
The bars can only displace on the x, y plane, thus the problem is two dimensional. The equivalent-spring
stiffness of the bars is denoted by

(1) E A(1) (2) E A(2)

k = 0
, k = 0
, (24.13)
L (1)
0 L (2)
0

in which A(e) (e) th

0 and L 0 denote the cross sectional areas and lengths, respectively, of the e bar in the reference
configuration, and E is the elastic modulus common to both bars.

24–8
24–9 §24.7 LPB ANALYSIS EXAMPLE

The figure shows the reference configuration C0 for the two bars. That configuration is taken when the applied
load is zero, that is, λ = 0. We shall assume that the stiffness of bar 1 is much greater than that of bar 2, i.e.,
k (1) >> k (2) and is such that the vertical displacement u Y 2 of node 2 under the load is very small compared to
the dimensions of the structure.
Now let the load p = λq be gradually applied by increasing λ. The structure assumes a deformed current
configuration in equilibrium, that is, a target configuration C. According to the LPB basic assumption, the
displacements prior to the buckling load level characterized by λcr are negligible. Therefore C ≡ C0 as long
as |λ| < |λcr |.
The linear finite element equations for the example problem are as follows. For element (1):
    
0 0 0 0 u X1 0
 0 1 0 −1   u Y 1   0 
k (1)  = . (24.14)
0 0 0 0   u X2   0 
0 −1 0 1 uY 2 −λq

For element (2):     

1 0 −1 0 u X2 0
 0 0 0 0   uY 2   0 
k (2)  = . (24.15)
−1 0 1 0   u X3   0 
0 0 0 0 uY 3 0

Assembling and applying the boundary conditions u X 1 = u Y 1 = u X 3 = u Y 3 = 0 we get

  
k (2) 0 u X2 0
= . (24.16)
0 k (1) uY 2 −λq

The linear solution is

λq λq L (1)
u X 2 = 0, uY 2 = − = − 0
(1)
. (24.17)
k (1) E A0

The axial linear strain and Cauchy (true) stress developed in element (1) are
uY 2 q q
(1) = =λ , σ (1) = E (1) = −λ . (24.18)
L (1)
0 E A(1)
0 A(1)
0

According to the assumptions stated above the change in geometry prior to buckling is neglected. Consequently

e(1) ≈ (1) s (1) ≈ σ (1)

The axial strain and stress of element (2) are zero.

The simplified nonlinear finite element equations are, for element (1)
        
 0 0 0 0 1 0 −1 0  u X1 0
 
(1)  0 1 0 −1  N (1)  0 1 0 −1   u Y 1   0 
k  +  =  0 . (24.19)

 0 0 0 0  L (1)  −1 0 1 0  
 u X2
0
0 −1 0 1 0 −1 0 1 uY 2 −λq

where N (1) = A(1)

0 s
(1)
= − p = −λq denotes the axial force in bar element (1).

24–9
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–10

For element (2) we have the same linear matrix equations as before because its geometric stiffness vanishes.
Assembling and applying displacement boundary conditions we get the equations
 λq 
k (2) − 0  
 L (2)  u X2 = 0 .
0
λq (24.20)
0 k (1) − (1) uY 2 −λq
L0

One now regards K in (24.20) as unaffected by the displacements u X 2 and u Y 2 , which is consistent with the
assumption that the change of geometry prior to buckling is neglected. This having being done, setting the
determinant of K to zero yields the buckling eigenproblem:
 λq 
k (2) − 0
L (1)
det  0
λq
 = 0. (24.21)
0 k (1) −
L (1)
0

This matrix is singular if either diagonal element vanishes, which yields the two eigenvalues

λcr 1 = k (1) L (1)

0 /q, λcr 2 = k (2) L (1)
0 /q, (24.22)

as critical values of the load parameter. Since k (2) << k (1) the lowest critical load will be

pcr = λcr 2 q = k (2) L (1)

0 . (24.23)

This is the buckling load obtained under the LPB assumptions.

24–10
24–11 §24.8 SUMMARY OF LPB STEPS

§24.8. Summary of LPB Steps

The foregoing example illustrates the key steps of LPB analysis. These are summarized below for
completeness.

1. Assemble the linear stiffness K0 and solve the linear static problem

K0 u = q0 λ, (24.24)

for λ = 1 and obtain the internal force (stress) distribution. Note: In statically
determinate structures, such as Exercise 24.4, the internal forces and stresses may
be obtained directly from equilibrium. However K0 is still necessary for step 3.
2. Form the reference geometric stiffness K1 for that internal force distribution. The
geometric stiffness is KG = λK1 .
3. Solve the stability eigenproblem

(K0 + λK1 ) zi = 0, or K0 zi = −λi K1 zi , (24.25)

The eigenvalue λi closest to zero is the critical load multiplier, and the associated
eigenvector zi gives the corresponding buckling mode.

24–11
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–12

Homework Exercises for Chapter 24

Bifurcation: Linearized Prebuckling I

EXERCISE 24.1 [A:15] Find the buckling mode (null eigenvector) z (normalized to unit length) for the
example problem of §24.6, and verify the orthogonality condition zT q = 0.

EXERCISE 24.2 [A:15] Find the buckling load for the two-bar problem if bar (2) forms an angle 0 ≤ ϕ < 90◦
with the x axis. Assume still that the system displaces only on the x − y plane and that k (2) << k (1) so that
the stress in bar (2) can be neglected in forming the geometric stiffness. Determine z and verify orthogonality.

EXERCISE 24.3 [A:20] Suppose the load of the two-bar example problem of §24.6 depends on the vertical
displacement of point 2 as p = −λcu 2Y 2 , where c is a constant with dimensions of stress. Show that even if
prebuckling deformations are neglected, the singular stiffness test leads to a nonlinear eigenvalue problem.

EXERCISE 24.4 [A:25] The “Euler column” shown in Figure E24.1 is modelled by one 2-node Euler-Bernoulli
beam column element along its length:

Y, y

;
E, I constant

; ;;
1 2
z X, x λP

Figure E24.1. One-element model of Euler column

The state parameters are the nodal displacements degrees of freedom arranged as

u 
X1
 uY 1 
 θz1 
u= 
 u X2  (E24.1)
 
uY 2
θz2

where θz1 and θz2 are (to first order) the end rotations, positive counterclockwise about z.
The linear material matrix in the reference (undeformed) configuration is

24–12
24–13 Exercises

 EA 
L 0 0 − ELA 0 0
 12E I 6E I − 12E3 I 6E I 
 0 
 L3 L2 L L2 
 4E I − 6E2I 2E I 
 L 0 L 
K0 =  L  (E24.2)
 EA 0 0 
 L 
 12E I 
 − 2I
6E

L3 L
symm 4E I
L
in which E is the elastic modulus, L the element length, A the cross section area, and I the moment of inertia
of the cross section about the z neutral axis.
The exactly-integrated geometric stiffness at the reference configuration is1
 0 0 0 0 0 0

 36 3L 0 −36 3L 
P  4L 2 0 −3L −L 2 
KG = λK1 , K1 =   (E24.3)
30L  0 0 0 
 
36 −3L
symm 4L 2
where N is the axial force in the element (here obviously equal to the applied force λP because the structure
is statically determinate.)
For this problem:

(a) Check that K0 and K1 satisfy translational infinitesimal rigid body motion conditions u X ≡ 1 and u y ≡ 1
if the six degrees of freedom are left unconstrained. (Convert those modes to node displacements, then
premultiply by the stiffness matrices.)

buckling mode

θz1
1 2
θz2 = −θz1

Figure E24.2. Symmetric buckling of one-FE model of Euler column.

(b) Set up the linearized prebuckling eigenproblem

(K0 + λK1 )z = 0 (E24.4)

Apply the support end conditions to remove u X 1 , u Y 1 and u Y 2 as degrees of freedom.

1 See e.g. Przemieniecki’s Theory of Matrix Structural Analysis, loc. cit.

24–13
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–14

(c) Justify that freedom u X 2 can be isolated from the eigenproblem, and proceed to drop it to reduce the
eigenproblem to 2 × 2.

(d) Reduce the 2 × 2 eigenproblem to a scalar one for the symmetric buckling mode sketched in Figure
E24.2. by setting θz1 = −θz2 (see Figure), and get the first critical load parameter λ1 .
(e) Repeat (e) for the antisymmetric buckling mode sketched in Figure E24.3 by setting θz1 = θz2 (see
Figure) and obtain the second critical load parameter λ2 .

θz1
θz2 = θz1
1 2

buckling mode

Figure E24.3. Antisymmetric buckling of one-FE model of Euler column.

(f) Compare the results of (d)–(e) to the exact critical load values
EI EI
P1E = −π 2 , P2E = −4π 2 , (E24.5)
L2 L2
The first one was determined by Euler in 1744 and therefore is called the Euler critical load. For P1E the
FEM result should be within 25%, which is good for one element.

(g) Repeat the calculations of λ1 and λ2 with the following reduced-integration geometric stiffness matrices6
 0 0 0 0 0 0

 9 3L 0 −9 3L 
P  L2 0 −3L −L 2 
K∗1 =   (E24.6)
8L  0 0 0 
 
9 −3L
symm L2
 0 0 0 0 0 0

 24 0 0 −24 0 
P  2L 2 −2L 2 
K∗∗ =  0 0  (E24.7)
1
24L  0 0 0 
 
24 0
symm 2L 2
and comment on the relative accuracy obtained against the exact values.

(h) Repeat the calculations of steps (b) through (e) for a two equal-element discretization. Verify that
the symmetric-mode buckling load is now −10E I /L 2 , which is (surprisingly) close to Euler’s value
P1E = −π 2 E I /L 2 .

6 These matrices are obtained by one-point and two-point Gauss integration, respectively, whereas the K1 of (E24.3) is
obtained by either 3-point Gauss or analytical integration.

24–14
24–15 Exercises

EXERCISE 24.5 [A:25] This is identical in all respects to Exercise 24.4, except that the 2-node Timoshenko
beam model is used, with Mac Neals’s RBF correction for the material stiffness matrix. The net result is
that K0 is identical to (E24.2) but K1 is different. In fact λK1 is given by Equation (9.45) where V = 0
and N = λP. Note: dont be surprised if the one-element results are poor when compared to the analytical
buckling load.

EXERCISE 24.6 [A/C: 20] The column shown in Figure E24.4 consists of 3 rigid bars of equal length L
connected by hinges and stabilized by two lateral springs of linear stiffness k. The applied axial load is
λP, where λ > 0 means compression. Compute the two buckling loads λ1 P and λ2 P in terms of k and L,
assuming the LPB model of infinitesimal displacements from the initial state. Show that one load corresponds

;;;;
to a symmetric buckling mode and the other to an antisymmetric buckling mode, and find which one is critical.

λP

;; ;;
A

L All 3 bars are

k considered rigid
1

;; k
2
L

;;
B

Figure E24.4. Buckling of a segmented-hinged column propped by two springs.

Hint. The two degrees of freedom are the small lateral displacements u 1 and u 2 of hinges 1 and 2, where
u 1 << L, u 2 << L. Write the total potential energy as
λP  2 
 = U − W, U = 12 ku 21 + 12 ku 22 , W = u 1 + (u 1 − u 2 )2 + u 22 . (E24.8)
2L
Explain where the expression of W comes from. Once  is in hand, it is smooth sailing.

24–15
 23
.

Detecting
and Traversing
Critical Points

23–1
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–2

TABLE OF CONTENTS

Page
§23.1. Introduction 23–3
§23.2. Direct Methods and Test Functions 23–3
§23.2.1. Determinant as CPTF . . . . . . . . . . . . . . . 23–4
§23.2.2. Lowest Eigenvalue as CPTF . . . . . . . . . . . . 23–4
§23.2.3. Solving the Minimum-Eigenvalue CPTF System . . . . . . 23–5
§23.3. Indirect Methods 23–5
§23.3.1. Refined Determination of a Critical Point . . . . . . . . 23–6
§23.4. Traversing Critical Points: Difficulties 23–7
§23.5. Newton-Raphson near Critical Points 23–9
§23.5.1. Injecting Perturbations . . . . . . . . . . . . . . 23–9
§23.5.2. Penalty Spring Stabilization . . . . . . . . . . . . 23–9
§23.5.3. Thurston’s Equivalence Transformation . . . . . . . . . 23–10
§23.6. Crivelli’s Procedure 23–11
§23.7. Predictor Stabilization 23–13
§23.8. Critical-Point Test Functions 23–13
§23.9. Branch switching 23–15
§23.9.1. Tangent Predictors . . . . . . . . . . . . . . . 23–15
§23.9.2. Buckling Mode Injection . . . . . . . . . . . . . 23–17
§23.10.Correctors 23–18
§23.11.Treating Bifurcation by Perturbation 23–20
§23. Exercises . . . . . . . . . . . . . . . . . . . . . . 23–21

23–2
23–3 §23.2 DIRECT METHODS AND TEST FUNCTIONS

§23.1. Introduction

Critical points were defined and mathematically characterized in Chapter 5. Chapter 18 explains
how solution procedures of purely incremental type can be adjusted to traverse limit points, and
difficulties in traversing bifurcation points are noted. Chapters 24 and follwing discuss critical
points in the context of stability analysis.
The present Chapter deals with the determination or detection of critical points in statically nonlinear
analysis and with procedures for traversal of such points with incremental/corrective techniques.
Emphasis is given to handling isolated bifurcation points, which offer moderate degree of difficulty:
tougher than limit points, easier than multiple bifurcation. Much of the material that follows is
taken from Chapter 5 of Crivelli’s thesis.1
There are two basic approaches to the problem of detecting critical points. We may attempt to find
critical points by solving an algebraic equation which has those point as a root, or we can try to
detect critical points as we are marching along an equilibrium path.
The first approach embodies what are collectively called direct methods. As the name implies,
these methods look for critical points without being concerned, at least theoretically, with tracing
equilibrium paths up to those points.
The second approach embodies the so-called indirect methods. In these methods the detection of a
critical point is intimately related to the continuation procedure.

Remark 23.1. The simplest example of a direct method is the Linearized PreBuckling (LPB) analysis discussed
in Chapters 24 and 25. This procedure, when applicable, sets up directly an eigenproblem in the reference
configuration and thus avoids tracing the nonlinear response.

§23.2. Direct Methods and Test Functions

A direct method for calculating bifurcation points consists of formulating a suitable set of equations
which has the critical points as solutions. In the context of general nonlinear analysis (that is, when
no a priori simplifying assumptions are made as in the LPB analysis) this set of equations should
include the equilibrium equation because only critical points on the equilibrium path are of interest.
It follows that direct formulations are achieved by augmenting the residual force equilibrium equa-
tion with a set of constraints that characterize the critical point. The main characteristic of this
procedure is that it does not require, at least theoretically, equilibrium path tracing. Thus, ideally it
should be possible to obtain all the critical points first and then join them with equilibrium paths.
Except for certain specialized situations this is not possible, however, because the solution of the
nonlinear equations resulting from this approach requires a good initial guess. Since these equations
are nonlinear, they cannot be solved directly, and linearization coupled with an iterative procedure
is required. This linearization accounts for most of the cost of direct methods.
As mentioned above, the procedure consists of augmenting the original residual force equation by
a set of constraints that define a critical point. This characterization of critical points is intimately

1 L. A. Crivelli, A Total-Lagrangian Beam Element for the Analysis of Nonlinear Space Structures, Ph. D. Dissertation,
Department of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, April 1991.

23–3
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–4

related to the idea of constructing a critical point test function or CPTF. By definition, a CPTF is a
function whose zeros are the critical points of the response.2

§23.2.1. Determinant as CPTF

Since the tangent stiffness matrix is singular at critical points, the most obvious test function is the
determinant of K. To be able to use this function, however, in the context of a Newton-Raphson
iterative procedure we need expressions for the partial derivatives of the determinant with respect
to the state parameters. This approach is generally impractical for several reasons:
1. Analytical expressions for the determinant are hopelessly complicated for any but trivial prob-
lems.
2. The estimation of these derivatives by one-sided finite differences would be enormously ex-
pensive, requiring the assembly and factorization of N + 1 tangent stiffness factorizations for
N degrees of freedom3
3. The detrminant is a notoriously ill-behaved function.

§23.2.2. Lowest Eigenvalue as CPTF

A critical point can also be characterized by checking for a null eigenvalue of K. Thus a prototype
of direct methods, as described in Seydel4 may be expressed generally as
 
r(u, λ)
K(u, λ)z = 0. (23.1)
φ(z)
The last equation, φ(z) = 0, has been added as a constraint on z to rule out the trivial solution
z = 0 and to make the system determinate. There are several possible choices for the normalization
equation, the simplest being to require z to have unit length: zT z = 1.
To solve this equation iteratively, define the augmented incremental vector
 
d
y= η (23.2)
h
where d and η are the corrective changes in u and λ, respectively, defined in Chapter 17, and and
h = zk+1 − zk . is the corrective change in z. Assuming that K does not depend on λ and that the
derivative of K with respect to u, denoted by Kv , is commutative with respect to its two outermost
indices, the resulting Newton system is
   
K q 0 r
Kv z 0 K y = − rλ (23.3)
0 0 aT φ

2 The name “branching test function” is sometimes restricted to test functions that catch only bifurcation points. Here we
shall use the generic term “critical test function” to embody all critical points.
3 That of the tangent stiffness K, plus N tangent stiffnesses corresponding to the perturbation of each state vector component.
4 R. Seydel, From Equilibrium to Chaos, Elsevier, New York, 1988.

23–4
23–5 §23.3 INDIRECT METHODS

where r is the residual from the equilibrium equations, rλ is the residual from the eigenvalue problem
and φ measures the violation of the normalization condition.
This approach has been explored by Wriggers and Simo5 in the context of nonlinear structural
analysis. As can be seen from the previous equation the Jacobian of this augmented system is
neither sparse nor symmetric, and has large storage requirements. In addition, the coefficient
matrix becomes singular at bifurcation points, and hence ill-conditioned in their neighborhood.

§23.2.3. Solving the Minimum-Eigenvalue CPTF System

It is desirable to solve (21.3) while taking advantage of the symmetry of K. This can be done by
using the concept of auxiliary systems. Assuming K is nonsingular, consider the five auxiliary
systems:

Kdr = −r, Kdq = q, Kdh = −rλ , (23.4)

Kdhr = Kv zdr , Kdhq = Kv zdq . (23.5)

Then the solution to the system can be written as:

φ − aT dhr
η=− , d = dr + ηdq , h = dh − dhr + ηdhq . (23.6)
aT dhq

The solution of these five systems at each corrective iteration is expensive. However, the main
disadvantage of this approach is having to compute the u-derivative of the stiffness matrix. Since an
exact derivation of Kv is only practically feasible for simple nonlinear problems, some investigators
(e.g. Seydel and Wriggers-Simo, loc. cit.) propose using numerical differentiation to compute Kv .
In this context it should be observed that only the product Kv z (a square matrix) is required.
Numerical experiments reported in Seydel’s book show that the convergence rate of this approach
depends on whether the critical point is a limit or a bifurcation point. At limit points, the Jacobian of
the augmented equation is nonsingular, which guarantees fast local convergence. At a bifurcation
point, the vector [ zT 0T 0 ]T is a left null eigenvector, which shows that the Jacobian becomes
singular. Thus, slower convergence rates may be expected at bifurcation points. However, experi-
ence shows that for some problems the convergence rate is good for the λ component when a not
very stringent accuracy is required, as can be seen from the examples presented in Wriggers and
Simo.

Remark 23.2. If the critical point is known to be a limit point, the penalty spring augmentation method
described below is much simpler and inexpensive. And even the simplistic random perturbation scheme
described there often works. Thus it appears that the direct method is advantageous in special instances, for
example when direct determination of the CP locations as stability envelope, is desirable, and a good estimation
of that location is available. This occurs in some applications such as optimization with stability constraints.

5 P. Wriggers and J.C. Simo, A General Procedure for the Direct Computation of Turning and Bifurcation Points, Int. J. Nu-
mer. Meth. Engrg., 30, 155–176, 1990.

23–5
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–6

§23.3. Indirect Methods

The underlying idea behind indirect methods is to recognize the occurrence of a critical point while
following an equilibrium path by continuation. The key to success is having a good CPTF τ . We
defer the discussion on how to construct such a CPTF to §21.4, and for the moment assume that
such a function is available.
These methods benefit from data collected from the continuation procedure. A critical point is
characterized by being a zero of the CPTF. Because hitting an exact zero is highly unlikely, we
check for a critical point by invoking the condition

τ (uk+1 , λk+1 ) τ (uk , λk ) < 0 (23.7)

This is called a bracketing or straddling condition. It means that a critical point occurs in the
interval or bracket [λk ≤ λ ≤ λk+1 ]. The next task is to “zero in” this point by reducing the bracket
size.

§23.3.1. Refined Determination of a Critical Point

Assume a critical point has been located by verifying bracketing condition (21.7) on the CPTF. Further, assume
that the test function is a continuous function of λ and has only one isolated zero in the interval. This condition
requires using a sensible incremental step — which is not known a priori. However, the problem of determining
an appropriate step may be solved by a combination of intuition and an adaptive incremental control strategy,
which monitors the convergence behavior of the solution.
Once the algorithm finds that a critical point is bracketed or straddled by two equilibrium solutions, we may
obtain a first approximation to such point by using linear interpolation:

λk+1 − λk k
λc  λk + τ (23.8)
τ k − τ k+1

We present below an adaptive method developed in Crivelli’s thesis (loc. cit. to improve the initial approxi-
mation (21.8). It makes use of iterative inverse interpolation and divided differences.6 We have computed a
set of m values of the CPTF at m unequally spaced values of the control parameter λ. We seek a polynomial
interpolation for τ of the form

τ (λ) = c0 + c1 (λ − λ0 ) + · · · + cm (λ − λ0 ) . . . (λ − λm−1 ) (23.9)

On defining the sequence of divided differences

τ (λ) − τ (λ0 )
τ [λ0 ; λ] =
λ − λ0
.. ..
. . (23.10)
τ [λ0 , . . . , λk−2 ; λ] − τ [λ0 , . . . , λk−2 ; λk−1 ]
τ [λ0 , . . . , λk−1 ; λ] =
λ − λk−1

6 For background in such techniques, see for example G. Dahlquist and A. Björck, Numerical Methods, Prentice-Hall,
Englewood Cliffs, 1974.

23–6
23–7 §23.4 TRAVERSING CRITICAL POINTS: DIFFICULTIES

the coefficients in equation (8) can be computed as

c0 = τ (λ0 )
c1 = τ [λ0 ; λ1 ]
.. .. (23.11)
. .
ck = τ [λ0 , . . . , λk−1 ; λk ]

Observe that this formula is recursive and a new point can be added without great difficulty. Now the problem
of locating a critical point is reduced to finding a zero of (21.9).

Remark 23.3. Observe that we may as well use equation to interpolate λ as a function of τ and obtain λc just
by setting τ = 0. However, both procedures are fundamentally different and may give different results. For
instance, for very sharp limit points and for bifurcation points the test function could stay almost constant for
most of the continuation and change to a steep slope while very close to the limit point. This type of behavior
can be regarded as an exponential relation between τ and λ. Such an exponential can be well approximated
by a polynomial, whereas its inverse —a logarithmic function— may pose severe difficulties. Thus we may
expect faster convergence from the former procedure.

Based on the foregoing remark we choose inverse interpolation for finding a zero of (21.9). For
this we recast that equation into the form

λc = ψ(λc ) (23.12)

or, more explicitly

1
λc = λ0 − [c0 + · · · + cm (λc − λ0 ) . . . (λc − λm−1 )] (23.13)
c1

The equation is then solved by iteration, with λ(0)

c equal to the approximated λc obtained form the
last interpolation, starting with the linear interpolation given by (21.8). Convergence is usually
fast because the error is inversely proportional to the first CPTF derivative, τ = ∂τ/dλ, which is
expected to be large near isolated critical points. Furthermore, should τ be small, we can expect a
good prediction from the linear interpolation formula.
The main advantages of the foregoing procedure is that it is self adaptive and easily programmable.
We can increment the number of points used in the interpolation equation with little extra effort.
We can keep the number of points under a given maximum by discarding the outermost points when
we add a new one. Thus the procedure is flexible and opens up several implementation alternatives.
Furthermore, the storage requirements are kept modest and we can expect good convergence to
reasonable accuracy.
After a new approximation λc has been established, we have to compute the corresponding solution
uc . However, some care has to be exercised since the Jacobian matrix can be very ill-conditioned in
the critical point neighborhood. There are several alternatives to circumventing this ill-conditioning
problem, which merge with the general topic of traversing a critical point. This subject is taken up
in the following subsections.

23–7
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–8

§23.4. Traversing Critical Points: Difficulties

In the remainder of this section we consider the problem of traversing critical points using incremen-
tal/corrective methods. Corrective solution methods, of which Newton-Raphson is the prototype,
may run into difficulties at or near critical points. Essentially those are caused by the fact that the
tangent stiffness matrix becomes singular at such points and consequently ill-conditioned in their
neighborhood. This ill-conditioning can introduce noise which may render the solution procedure
unstable.
Sometimes the problems associated with noise are easily circumvented, whereas in some cases
the problems may become computationally overwhelming and may require either specialized tech-
niques or intensive human intervention to proceed. Generally speaking, the degree of difficulty is
influenced by the following factors:
1. Bifurcation points are more difficult to handle than limit points.
2. Isolated critical points (at which the rank deficiency is one) are easier to handle than multiple
or compound critical points.
3. Dynamic critical points of flutter type are more difficult to handle than static ones. Ultimate
in nastiness (sort of like the triple jump in ice skating) are problems of fracture or localiza-
tion where an equililibrium path suddenly terminates, and dynamic methods are required for
proceeding further.
In what follows we focus on the characteristics of Newton’s method near an isolated critical point of
static type. Even under those relatively benign conditions the topic is still an area of active research.

23–8
23–9 §23.5 NEWTON-RAPHSON NEAR CRITICAL POINTS

§23.5. Newton-Raphson near Critical Points

The behavior of Newton-Raphson methods at critical points have received much attention from the
numerical analysis community.7 Convergence rates have been established and accelerators have
been proposed in a general framework. Under certain conditions, it can be established that the
convergence of the solution component in the null space is linear while that in the range space
converges superlinearly. This property can be exploited in the limit point case.
The purpose of a Newton-Raphson critical point accelerator, or NRCPA, is to recover quadratic
rate of convergence. To achieve this goal on a general setting NRCPAs need information provided
by second order (and ocassionally higher) rate equations. This information is very expensive to
obtain because the rate K̇ of the tangent stiffness matrix, which is a “cubic array” discussed further
in Chapter 24, makes its appearance. Because of such expense, NRCPAs based on higher order
rate equations are rarely used in nonlinear structural mechanics.
The procedures described below represent more practical approaches to the traversal problem. They
attempt to live with the first-order rate information, such as provided by K and q, which is normally
available in finite element programs.

§23.5.1. Injecting Perturbations

By far the simplest technique to attempt critical point traversal is the equivalent of the “Hail Mary”
football play. If at a certain step of an incremental/iterative process the equation solver returns a
“singular stiffness” diagnostic, perturb the state u by a minute random amount and try again. This
may be repeated up to a certain number of “downs” (usually 2 or 3).
This simplistic method sometimes works surprisingly well for crossing “nice” limit points in con-
junction with the arclength increment control method. But at sharp limit points, bifurcation points
or regions where the tangent stiffness has high rank deficiency this technique often fails.

§23.5.2. Penalty Spring Stabilization

As emphasized several times before, the essential difficulty lies in the ill-conditioning of K near
critical points. To stabilize this matrix, Felippa8 has proposed to combine the solution of three
linear systems. The coefficient matrix is rendered nonsingular by adding a fictitious penalty spring
stiffness s to the i th equation:

(K + sEi ) drs = −r, (K + sEi ) dqs = q, (K + sEi ) dse = sei (23.14)

7 See, for example, the publications:

D. W. Decker, H. B. Keller and C. T. Kelly, Convergence Rates for Newton’s Method at Singular Points, Siam J. Nu-
mer. Anal., 20, 296–314, 1983.
C. T. Kelly and R. Suresh, A New Acceleration Method for Newton’s Method at Singular Points, Siam J. Numer. Anal.,
20, 1001–1009, 1983.
G. Reddien, On Newton’s Method for Singular Problems, Siam J. Numer. Anal., 15, 993–996, 1978.
W. Rheinboldt, Numerical Analysis of Continuation Methods for Nonlinear Structural Problems, Computer & Struc-
tures, 15, 1–11, 1978.
8 C. A. Felippa, Traversing critical points by penalty springs, Contrib. C2/1, Proc. NUMETA’87 Conf., M. Nijhoff Pubs.,
Dordrecht, 1987.

23–9
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–10

Here ei is the elementary vector of order N all of whose entries are zero except the i th entry which
is one, and Ei = ei eiT . Some possible choices for the index i are discussed in that article (handed
out in class). One advantage to this approach is that the symmetry and sparseness of K is not
affected by this diagonal correction. The solution of the original Newton-Raphson system (17.13),
reproduced here for convenience:
    
K −q d r
=− , (23.15)
a T
g η c

can be expressed as a linear combination of the solutions of the three auxiliary problems defined in
(21.14):
d = drs + σq dqs + σe dse ,
  . (23.16)
η = − c + aT d /g,
Here the coefficients σq and σe are obtained by requiring that d and η solve the original problem.
The following 2 × 2 unsymmetric system results:
    
g + aT dqs aT dse σq −c − aT drs
= (23.17)
−eiT dqs 1 − eiT dse σe eiT drs

The computation of η from this equation breaks down if g = 0. In such a case, it is recommended
that η be recovered from the equivalent equation:

σe di − di2
η = σq + s (23.18)
dT q
The value of s in equation is irrelevant as long as it is sufficiently large to stabilize K.
An advantage of this procedure is that it avoids the need for partitioning the stiffness matrix proposed
by Rheinboldt (loc. cit.), which requires special treatment of the elements in the i th row and column
of K. Furthermore, none of the right-hand sides of the auxiliary systems in requires access to
elements of K.
Disadvantages of this procedure are as follows:
1. It breaks down at bifurcation points, where the coefficient matrix of the 2 × 2 system becomes
singular.
2. Some overhead is necessary to trace the rate of change of the components of the state vectors,
which is used as a criterion in the selection of the index i.
3. The equation solver should be able to undo part of the reduction, from the current value of the
row index to the selected value of i.

§23.5.3. Thurston’s Equivalence Transformation

A different approach has been proposed by Thurston, Brogan and Stehlin. 9 Here an equivalence
transformation is effected on the stiffness matrix to obtain a partitioning such that the singularity

9 G. A. Thurston, F. A. Brogan and P. Stehlin, Postbuckling Analysis Using a General Purpose Code, AIAA Journal, 24,
1013–1020, 1986.

23–10
23–11 §23.6 CRIVELLI’S PROCEDURE

is confined to a small block-submatrix. To achieve this transformation it is necessary to compute

the eigenvectors corresponding to the smallest eigenvalues of K. These eigenvectors enter the first
columns of the transformation matrix, which is then completed with a permutation matrix in such
a way that renders the full matrix non-singular. The solution vector is then split accordingly and
the components corresponding to the permutation part of the transformation matrix are condensed
and solved in terms of the modal variables. Although this procedure works adequately for almost
singular matrices, it fails for truly singular matrices where the modal variables remain undefined.
§23.6. Crivelli’s Procedure
We present here an alternative procedure developed in Crivelli’s thesis (cited in footnote 1) which
does not involve partitioning but requires access to elements of K. Furthermore, it avoids the need to
estimate penalty coefficients and the choice of the index i is straightforward. It also requires minor
modifications to the solver routine. It can be implemented without even modifying the solver, by
forcing the solver to return when it finds a diagonal term smaller than a given threshold, performing
the required transformations outside the solver and then having the solver continue from the point of
interruption. In addition, the proposed procedure furnishes an approximation to the null eigenvector
and provides a measure of the distance to the critical point.
Define the permutation matrix Pi :
i
 ↓ 
1 ... 0 0 ... 0 0
. .. .. .. .. .. .. 
 .. . . . . . .
 
0 ... ... 0
 1 0 0 
Pi = i → 
0 ... 0 0 ... 0 0  (23.19)
0 ... ... 0
 0 1 0 
 .. .. .. .. .. .. .. 
. . . . . . .
0 ... 0 0 ... 1 0
Observe that  
Ii−1
Pi PiT = 0 = In − ei eiT , (23.20)
In−i
where N is the dimension of Pi and Ik is the k × k identity matrix. Define
K̂ = PiT KP + en enT , ĥ = PiT Kei = PiT ki , (23.21)
and denote ẑ to be the solution of
K̂ẑ = ĥ, z = Pẑ − ei . (23.22)
The vector z so defined satisfies
Kz = kiT zei , ki = Pi K̂PiT z + K ii ei . (23.23)
We show below that if K is singular with rank deficiency of one, and the index i is chosen properly,
then z is the null eigenvector of K. If K is singular, there is a subset of the columns of K that are

23–11
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–12

linearly dependent. Denote by S = {i 1 , . . . , i k } a set of indices such that S ⊂ {1, . . . , n}; this set S
is a subset of the column indices of K. This set is chosen so that there is a linear combination of
columns of K satisfying 
βik kik = 0 (23.24)
i k ⊂S

where ki is the i th column of K. Such a set exists because K is singular. Choose an index i from
such that βi = 0. Column ki can be then expressed as a linear combination of the other columns
of K. Apply now the procedure with the index i chosen as described. It is immediately seen that
kiT z = 0; therefore z is the null eigenvector of K.
If K is almost singular, then z is an approximation to the null eigenvector. It can be shown that if K
is continuously differentiable and K̂ is nonsingular, then z approaches z as (u, λ) approach (uc , λc ).
To solve the original problem, define the following auxiliary systems:

K̂d̂r = PiT r, dr = Pi d̂r , K̂d̂q = PiT q, dq = Pi d̂q (23.25)

Observe that drT ei = dqT ei = 0. Define

d = dr + σq dq + σz z (23.26)

To obtain the coefficients σq and σe we use a similar procedure as before. We have

Kdr = Pi PiT KPi d̂r + kiT dr ei = Pi K̂d̂r + kiT dr ei

(23.27)
= −(I − ei eiT )r + kiT dr ei = −r + (kiT dr + ri )ei

and
Kdq = q + (kiT dq − qi )ei (23.28)
Also observe that
T
kiT dr = drT (Pi K̂ẑ + K ii ei ) = d̂r K̂ẑ
(23.29)
= −rT Pi ẑ = −rT (z + ei ) = −rT z − ri
Thus
kiT dr + ri = −rT z (23.30)
Similarly
kiT dq − qi = −qT z (23.31)
From previous equations it is immediately seen that
 
c + aT d  
Kd − qη = −r + σq + q + kiT zσz + qT zσq − rT z ei (23.32)
g
To recover the original system, we require the terms inside squared brackets to vanish, which leads
to the auxiliary system
    
g + aT dq aT z σq −c − aT dr
= (23.33)
qT z kiT z σz rT z

23–12
23–13 §23.8 CRITICAL-POINT TEST FUNCTIONS

Finally, we have to recover η. Instead of using We can obtain η directly by requiring

η = σq (23.34)

This approach has some advantages over the one pertaining to the penalty spring method. First, if
σz = 0 it reduces to the one used by classical continuation. Second, it enforces prior equations.
Closer inspection of tells us that σe = di if exact arithmetic is used. Thus, the second term vanishes.
When finite precision arithmetic is used, rounding errors may prevent this term from vanishing.
If in addition, a sufficiently large value for the penalty spring s is chosen, could produce some
artificially large value of η with the consequence that the corrector step may move the solution
away from convergence.
Several remarks may be made in regard to the preceding derivation. At a limit point, neither qT z
nor aT z vanish and the equations are well behaved even though kiT z is zero. Thus this procedure can
satisfactory traverse limit points. At a bifurcation point the last row of equation is identically null,
thus the value of σz is not defined. This is consistent with the definition of bifurcation point, since at
these points more than one solution is possible and we cannot continue tracing any branch without
introducing additional information. Thus when a bifurcation point is detected, this procedure breaks
down and we have to resort to a branch switching algorithm as described in the following sections.

§23.7. Predictor Stabilization

A similar procedure can be employed to stabilize the predictor, a subject that has not received much
attention in the literature. If K is almost singular, the predictor may be poor. Using d̂q defined in
equation (11) we define the predictor step u(0) as

u(0) = d̂q λ(0) + σz z (23.35)

Thus  
Ku(0) − qλ(0) = λ(0) (kiT q − qi ) + σz kiT z ei (23.36)
Requiring the term in brackets to vanish and using equations (17) the load increment is obtained as:

(0) kiT z
λ = T σz (23.37)
q z
Assuming we use the global hyperelliptic constraint we can obtain σq as

σz =     (23.38)

 2  T 2 2
 a  ki z k T
z
± 2 dqT Sdq + zT Sz + i
b2
v qT z qT z

since by construction dqT z = 0.

Observe that at a limit point, kiT z = 0, qT z = 0, equation (38) gives σz = , λ(0) = 0, allowing
the solution to move away from the limit point. At a bifurcation point qT z = 0, and this equation
does not give any solution for σz , as expected, since any solution with a nonzero value of σz will lie
on the emanating branch. In this case, we simply set σz = 0 and use prior relations compute λ(0) .

23–13
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–14

§23.8. Critical-Point Test Functions

In this section we investigate several candidates for CPTF. A CPTF is a function that allows
monitoring the continuation procedure for occurrence of critical points. By definition, critical
points should be the only zeros of a CPTF. As mentioned before, critical points are characterized
by the tangent stiffness K being singular. Thus, test functions essentially monitor K to obtain an
estimate of its proximity to a critical point.
A highly accurate estimator of how singular is K is its smallest eigenvalue. Since K is real and
symmetric, its eigenvalues are all real; in addition, if K varies continuously, its eigenvalues also
vary continuously and singularity is simply detected by checking for a null eigenvalue. Thus, the
smallest eigenvalue of K makes a very attractive choice for a CPTF. However, an eigenvalue analysis
at each increment is prohibitively expensive.
Another possible test function is the determinant of K. As noticed by Abbott10 the order of mag-
nitude of the determinant may be computationally inconvenient, especially for large stiff systems,
when the solution is computed far away from bifurcation. Scaling K by 10 changes the determinant
by 10 N ! This difficulty can be overcome by computing the determinant as a pair of numbers, a
characteristic normalized between −1 and 1 and a mantissa. This decomposition is available in
several direct solvers. A drawback of this approach occurs, however, if the null eigenvalue has an
even multiplicity, since in this case the determinant will not change sign and the tester will fail to
detect a critical point.
Two additional choices are considered in the following. Using previous equations the following
relation results:
zT Kz kiT z
= (23.39)
zT z zT z
If we regard the left-hand side of this equation as a Rayleigh quotient, and taking into account that
z is a close approximation to the smallest eigenvector of K when the solution is sufficiently close to
the limit point, we may use the right-hand side of equation to monitor the smallest eigenvalue of K.
Thus, ki z is a candidate for a CPTF. A similar test function is proposed by Seydel (loc.cit.). This
approach is expensive if it is to be carried out at every increment, because the decomposition is only
computed near critical points. However, if by some other means we have detected the proximity
of a critical point, this computation may be obtained as a by-product of the stabilization procedure
described in the previous Chapter. In such case, it becomes an attractive CPTF. We therefore look
for a cheaper alternative that can provide a good estimate of the singularity of K.
Since we use full Newton we generally have a decomposition of K of the form:

K = LT DL (23.40)

where L is an upper triangular matrix and D is diagonal. It is well known that

τ = min D(i) (23.41)

i

10 J. P. Abbott, An Efficient Algorithm for the Determination of Certain Bifurcation Points, J. Comput. Appl. Math., 4,
19–27, 1978.

23–14
23–15 §23.9 BRANCH SWITCHING

also monitors the smallest eigenvalue of K. Moreover, τ = 0 when K is singular and τ < 0 if
K is indefinite. Thus, τ defined is readily available without any overhead and is the test function
adopted here. Once it has been determined that there is a critical point lying between two solutions
we may switch to the test function.

§23.9. Branch switching

In this Chapter we consider the computational treatment of bifurcation points. Such points  T are
characterized by the fact that the augmented stiffness becomes singular there. The vector z 0
is a left singular vector of the augmented
  stiffness, as can be seen by premultiplying the augmented
stiffness matrix by the vector zT 0 and taking into account that q is orthogonal to z. However,
this relation shows that the problem is consistent and as such, a solution exits; more precisely, an
infinite set of solutions exits with any two of them differing by a vector lying in the null space of
K.
The coefficient σz cannot be obtained because the second row of this equation vanishes identically.
Thus, we cannot use here the same procedure used to traverse limit points. As previously mentioned,
at a bifurcation point more than one branch coexist. If σz is kept equal to zero, we may continue
along the current branch, while if another value of σz is chosen appropriately, we may switch to a
different branch. It is seen that the problem is indeterminate and we need an additional condition
to specify on which branch of the solution the continuation will proceed. While continuing on the
known branch can be achieved fairly simply, switching to a crossing branch is not so obvious and
requires a more elaborate procedure.
We will assume that K has a rank deficiency of one and the bifurcation point is isolated or simple.
Under these conditions, there are only two intersecting branches at the bifurcation point. The
process of branch switching requires two steps. First we have to find some point sufficiently close
to the crossing branch. Second, we have to stay on that branch; this requires to avoid any iteration
sending us back to the known branch. This Chapter will be concerned with the first step, calculating
one solution point on the emanating branch.
Suppose we have detected a critical point, and that we have classified it as a bifurcation point. The
scenario is as follows: we have some solutions u(λ) such that r(u, λ) = 0. If the bifurcation point
has been located by using the indirect method, we may assume that we have one solution, say (u, λ),
that approximates the bifurcation point (uc , λc ), and in addition we have a good estimate of the
buckling mode vector z. Denote a solution on the crossing branch by (w, λ), i.e. r(w, λ) = 0. Our
main goal in this section is to find one such solution. This first solution is then used as a starting
point to trace the entire branch.
In general, any method for switching branches consists of two parts, first an approximation to a
point on the crossing branch is guessed by means of a predictor; then an iteration converging to
that branch must be established. In the following subsections we will describe different types of
predictors.

§23.9.1. Tangent Predictors

Possibly the most accurate predictor is the one based on the tangent to the crossing branch. However,
this procedure requires a rather accurate calculation of the bifurcation point (uc , λc ) and of the second
derivatives of the residual vector, which is given in terms of the rate matrices K̇ and q̇. Since we

23–15
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–16

Unacceptable
iteration
Predicted P
λ solution

Acceptable
Unknown branch
C B iteration

known solution

Known branch

Figure 23.1. Schematic representation of a branch switching procedure.

have assumed that the bifurcation point is simple, there can be at most two intersecting branches.
Since q is in the range of Kc , the system Kw = q is consistent at a bifurcation point. Hence, there
is a unique solution w lying in the range of K(uc , λc ) that is, satisfying the relation wT z = 0. Now
the state variation rate u̇ at a bifurcation point can be decomposed into a homogeneous solution
component σ z in the buckling mode direction, and a particular solution component, orthogonal to
z, which is furnished by w:
u̇ = (w + σ z) λ̇ (23.42)
The next step is to find σ . It was pointed out that the first-order rate equations do not provided enough
information for computing σ . To obtain the required information we must resort to the second-
order system. Premultiplying equation by zT and taking into account the bifurcation condition 3 in
Definition 2 we get the scalar equation

zT K̇u̇ − zT q̇λ̇ = 0 (23.43)

The rate matrices K̇ and q̇ may be written as linear combinations of u̇ and λ̇ as:
 
K̇ = Lu̇ + Nλ̇, q̇ = − Nu̇ + jλ̇ (23.44)

Then    
zT Lu̇ + Nλ̇ u̇ + zT Nu̇ + jλ̇ λ̇ = 0 (23.45)
The replacement of u̇ by its decomposition gives
   
zT L (w + σ z) λ̇ + Nλ̇ (w + σ z) λ̇ + zT N (w + σ z) λ̇ + jλ̇ λ̇ = 0 (23.46)

Insertion of the relations

a = zT Lzz, b = zT (Lzw + Nz) , c = zT (Lww + 2Nw + j) (23.47)

23–16
 23–17 §23.9 BRANCH SWITCHING

yields the quadratic equation

aσ 2 + 2bσ + c = 0 (23.48)
If at least one coefficient is nonzero, this quadratic equation provides two roots: σ1 and σ2 —root
σ = ∞ is acceptable. Since by hypothesis a branch reaches the bifurcation point, one of the
roots must be real. Consequently, the other root must also be real. The condition that at least one
coefficient be nonzero serves as a criterion for a simple bifurcation point because it guarantees a
transversal intersection of two branches. It is worth noting that a = 0 characterizes a pitchfork or
symmetric bifurcation while a = 0 characterizes a transcritical or asymmetric bifurcation.
Although this method is systematic and reliable, it suffers from the disadvantage of requiring second
derivatives of the residual in addition to requiring an accurate approximation of the bifurcation
point. Both procedures are rather expensive, particularly the former, since it is not always possible
to obtain an analytical expression for the rate of the stiffness matrix in a reasonably simple way. This
is obviously the case when the formulation used corresponds to elaborate mathematical models.
These problems motivated the search for other alternatives. In general, these alternatives are based
on buckling mode injection.

§23.9.2. Buckling Mode Injection

The purpose of these predictors is to obtain a sufficiently close approximation to the emanating
branches without resorting to the computation of the exact tangents to such branches. The underlying
idea is to obtain an increment vector d representing the difference between one solution in the known
branch and a solution on the unknown branch for a given value of the load parameter λ,

d = w − u, λ=λ (23.49)

Observe that the two tangents at the bifurcation point obtained by replacing the two roots span a
plane. In a sufficiently small neighborhood of the bifurcation point (uc , λc ) both branches lie on
that plane. If λ is sufficiently close to λc it would be possible to approximate the vector d by a vector
almost parallel to this tangent plane. If d could be obtained exactly it would be possible to switch
branches in a straightforward manner. However, in most practical problems only an approximation
to d is possible. Such approximation is generally split into two steps
1. Find a direction pointing to the other branch
2. Find a step along this direction so that the distance between the two solutions is minimized in a
given norm.
The problem is then reduced to finding d and δ such that

|u + δd − w| = 0, r(u, λ) = r(w, λ) = 0 (23.50)

The first condition can be viewed in the more general context of a constraint relation between the
solutions on both branches, thus generalizing constraints.
The first subproblem can be solved by taking d proportional to z since z lies on the plane defined
by the two tangents. This can easily be seen just by taking the difference between the two tangents
and observing that this difference is a multiple of z. This procedure can be viewed as an injection
of the buckling mode into the known solution.

23–17
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–18

The solution to the second subproblem is more difficult. It does not seem to be possible to find a δ
that satisfies the previous equation for a given value of λ without resorting to information provided
by second or higher-order rate equations. We cope with this problem by shifting it to the corrector,
which will adjust λ instead of δ in such a way that a constraint condition between the solutions on
both branches is satisfied. The problem is so restated because the corrector is better equipped to
deal with a problem posed in terms of changes in the state vector and control parameters, because
an incremental equation for δ is not available. Thus, at the predictor level, we will be content with
an estimation of δ. Some values dictated by experience are given in Seydel (loc.cit.).
Note that two predictors are in fact available:

w1,2 = u ± δz (23.51)

For perfect structures undergoing symmetric —pitchfork— bifurcation, the sign choice is irrelevant,
but for asymmetric —transcritical— bifurcation, one branch will be energy preferred. This difficulty
can be overcome by calculating solutions on both half-branchs emanating from the bifurcation point
and choosing the one with least energy.
The success of the whole procedure strongly depends on an appropriate corrector. Correctors should
be carefully designed in order to display selective properties. This is meant as the capability of the
corrector to avoid converging towards the known branch. The design of such correctors is dealt
with in the next section.

§23.10. Correctors

One alternative is to enrich the displacement increment with the buckling mode. Felippa (loc.cit.)
modifies equations so that
d = drs + σq dqs + σe dse ± σz z (23.52)
To provide sufficient equations a length constraint in injected, zT d = δ. Again, the unknown
coefficients are obtained by requiring that d verifies the incremental equation which leads to the
following 3 × 3 unsymmetric auxiliary system:
    
g + aq ae az σq −c − ar
−dqi 1 − dei l σe = dri (23.53)
sq se 1 ±σz δ − sr
in which
az = zT a, dz = zT d, sr = zT drs , sq = zT dqs , se = zdse (23.54)
The selective properties of this corrector are not clearly established.
To obtain a corrector with enhanced selective properties, Skeie and Felippa11 propose using a
constraint that has greater affinity for the emanating branch than for the known branch. For this
purpose, they look at the limit behavior of the local hyperelliptic constraint. They show that this
constraint reduces to a narrow cylinder around the known branch when a/  b/. To obtain a
robust corrector it is necessary to inject information from both branches into the constraint equation.

11 G. Skeie and C.A. Felippa, Detecting and Traversing Bifurcation Points in Nonlinear Structural Analysis, Report CU-
CSSC-89-23, Center for Space Structures and Controls, University of Colorado, Boulder, 1989.

23–18
23–19 §23.10 CORRECTORS

This can easily be achieved if the constraint is expressed in terms of the difference between the
solutions on two different branches, as shown below.
In the correction step, can be generalized to
c(w − u, λ − λ) = 0 (23.55)
Increments to both w and u and to λ are computed so that this constraint is enforced. What we
are trying to avoid is an iteration converging towards the known branch. Experience shows that
this frequently happens when λ = λc is kept fixed while trying to obtain a state vector u satisfying
the residual equation. For example, consider the case of symmetric —pitchfork— bifurcation. In
this case, the crossing branch exists for values of λ either greater or smaller than λc . Since we
do not know beforehand on which side of the bifurcation point the crossing branch lies, there is
a 50% chance that no emanating branch exist for the chosen value of λ. Besides, it is possible
that the emanating branch could not be parametrized by λ close to the bifurcation point and a local
parametrization may be required.
Once again we can expand the solution space. A point in this expanded space is defined by (u, w, λ)
and a parametrization for the solution in this new space is sought. In the case of prior equations the
parameter was the arc-length ; here δ plays the same role. The only difference is that we replace
u, which represents an increment of the solution on the known branch, by w − u which measures
the difference between two solutions lying on different branches. For example, displacement control
results in
wk − u k = δ (23.56)
In summary, we look for two state corrections dw and dv and a load correction η that simultaneously
satisfy
r(uk + dkv , λk + ηk ) = r(wk + dkw , λk + ηk ) = 0 (23.57)
and the constraint condition
c(uk + dkv , wk + dkw ; λk + ηk ) = 0 (23.58)
These corrections are obtained from
    
Kv 0 −q dv rv
0 Kw −q dw = − rw (23.59)
a T
−aT g η c
where Kv = K(uk , λk ) and Kw = K(wk , λk ).
Again it is possible to solve the previous equation while preserving the symmetry and sparseness
of K. If K is nonsingular we may write
dr u = −K−1
v rv , dr w = −K−1
w rw , dqu = K−1
v q, dqw = K−1
w q (23.60)
Thus, the two incremental displacements and the load increment may obtained as:
dv = dr u + ηdqu
dw = dr w + ηdqw
(23.61)
c + aT (dr u − dr w )
η=−
g + aT (dqu − dqw )

23–19
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–20

By using the procedure described in previous equations we require much less storage than when
using the original system, because Kv and Kw can overwrite each other. The extra storage re-
quirement is for four N -dimensional auxiliary vectors. However, it is still demanding in terms of
computer time, because two linear systems of equations with different coefficient matrices must be
solved at each iteration.
As an alternative, we may keep the solution on the known branch fixed and search for another
equilibrium solution that also verifies the constraint equation. That is, solve
 
r(w, λ)
=0 (23.62)
c(u; dw , λ)

This equation resembles the original continuation equation; the main difference being a slight
modification of the constraint condition. However, there is no guarantee that a solution of this
equation will lie on the emanating branch; observe that u and w correspond to two different values
of the load parameter λ and it is possible that two solutions on the same branch satisfy the constraint
equation.
In the examples presented in Crivelli’s thesis the first approach is used, since we are willing to
pay the additional price to guarantee the success of the branch switching. Furthermore this cost is
modest compared to that of the entire continuation process.
A remark is in order here. The stiffness matrices Kv and Kw may become ill-conditioned and a
procedure similar to that leading to equations may be developed. However, note that we are not
looking for a highly accurate solution on the known branch, since we are using it just to repel an
iteration from the crossing branch back to the known branch. Such a process is not required to
provide an highly accurate solution on the crossing branch either, because that solution is just used
to obtain a starting point on the new branch. We can thus apply the transformation to K to avoid
numerical instability, but regarding dr u and dr w as sufficiently good approximations to dr u and dr w ,
respectively. Once a point on the new branch is obtained, the solution can be further improved by
switching back to the continuation procedure used for regular points.

§23.11. Treating Bifurcation by Perturbation

The last approach to be mentioned here is treating bifurcation by perturbation. The idea is to perturb
the residual equation r(u, λ) in such a way that the underlying regularity intrinsic to a bifurcation
point is destroyed. This is accomplished by introducing physical or numerical imperfections. In the
mathematical literature, this approach is referred to as unfolding. The perturbed system displays
limit point behavior rather than bifurcation point behavior. The key difficulty with this approach is
to find a perturbation parameter  and an imperfection distribution that is critical to the underlying
regularity. For simple structures, this parameter  may be guessed from physical considerations.
Once an appropriate value is found, the branch can be traced by the standard continuation procedure
and the problem is reduced to limit point traversal.

23–20
23–21 Exercises

Homework Exercises for Chapter 23

Detecting and Traversing Critical Points

EXERCISE 23.1
[C:25] For the steep 2-bar arch with S = 2 and arbitrary H and subject to a vertical load −λq, find the locus
of all critical points directly by setting up and solving the r = 0 system augmented by det(K) = 0 as CPTF.
Hint: you may try the effect of

ClearAll[Em,A0,S,vX,vY,H,lambda];
Em=1; A0=1; S=2;
Do [
rX = (4*A0*Em*vX*(S^2 + 2*vX^2 + 4*H*vY + 2*vY^2))/(4*H^2 + S^2)^(3/2);
rY = (8*A0*Em*(H + vY)*(vX^2 + 2*H*vY + vY^2))/(4*H^2 + S^2)^(3/2)-lambda;
K={{(4*A0*Em*(S^2 + 6*vX^2 + 4*H*vY + 2*vY^2))/(4*H^2 + S^2)^(3/2),
(16*A0*Em*vX*(H + vY))/(4*H^2 + S^2)^(3/2)},
{(16*A0*Em*vX*(H + vY))/(4*H^2 + S^2)^(3/2),
(8*A0*Em*(2*H^2 + vX^2 + 6*H*vY + 3*vY^2))/(4*H^2 + S^2)^(3/2)}};
sol=NSolve[{rX==0,rY==0,Det[K]==0},{lambda,vX,vY}];
Print["solution for H=",H," is ",({lambda,vX,vY}/.sol)//InputForm],
{H,1/2,3,1/2} ];

EXERCISE 23.2 [A+C:25] The residual equation

r (θ, λ) = θ − λ sin θ = 0, (E23.1)

where v ≡ θ is the only degree of freedom, provides what is perhaps the simplest example of bifurcation.
This has two solution paths, θ = 0 (the fundamental path) and λ = θ/ sin θ (the post-buckling path), which
intersect at B (λ = 1, θ = 0.
Solve this equation using the purely incremental arclength control method programmed in Chapter 18 and
devise a procedure by which the program, starting from λ = θ = 0 and moving up the fundamental path, detects
B and (without cheating) follows one of the post-buckling paths. Probably the simplest one to implement is
perturbation.

EXERCISE 23.3 [C:25] Solve the steep arch H = 2S under a vertical load using the program of Exercise
20.3 and observe whether the program senses the bifurcation point or just marches along the fundamental path.
Experiment with perturbing the problem by inserting a small horizontal load to see whether you can coax it
into the secondary path.

23–21
 22
.

Accelerators
and Line Search

22–1
Section 22: ACCELERATORS AND LINE SEARCH 22–2

TABLE OF CONTENTS

Page
§22.1. Introduction 22–3
§22.2. General Comments 22–3
§22.3. Accelerator Models 22–3
§22.3.1. Tangent Accelerators for Fixed λ . . . . . . . . . . . 22–3
§22.4. References 22–10

22–2
22–3 §22.3 ACCELERATOR MODELS

§22.1. Introduction

Newton-like methods were initially used in geometrically nonlinear analysis with a unit h. This is
optimal for conventional Newton (CNM) near a solution, but not necessarily so otherwise. Since
a key reason for letting h vary is to accelerate convergence to equilibrium, schemes that utilize
nonunitary stepsizes are called accelerators. Another application is diagnosing and overcoming
divergence in “difficult” regions of the response. The latter application is tied up with the line search
technique that comes from the field of optimization. In the present Chapter we study acceleration
formulas and line search schemes based on the relaxation equations.

§22.2. General Comments

The device of letting h be a free parameter has proven useful in the following situations.
1. Oscillatory Response. Successive iterates “bounce” around so that the residuals at successive
steps are almost equal and opposite. This behavior is quite common when the structure
“stiffens” globally. The cure is to go half-way, which means h ≈ 0.5. Taking h < 1 to kill
oscillations is sometimes called underrelaxation.
2. Slow Monotonic Convergence. Successive residuals “point” the same way, but hardly decrease.
This behavior is quite common when the structure “softens” globally, as in extensive creep or
plasticity. The remedy is to “overrelax” with h > 1.
3. Erratic Behavior. The response of many nonlinear structures exhibits “difficult regions”
where sucessive residuals do not show significant correlation, and divergence is likely. If the
nonlinearities are smooth, this behavior is observed as one approaches bifurcation points or
sharp limit points. It is especially pronounced when several such points are clustered, as often
happens in optimized structures. In hard nonlinear problems, such as contact and cracking,
this behavior emanates from local effects. The recommended cure is cautious line search. If
the search delivers a very small h, say h < 0.1, the program should recompute the stiffness
matrix.
The determination of h at each iteration step can be used for monitoring these situations. For
example, if h ≈ 0.5 for two or three consecutive steps, chances are that the iterates are oscillating.
Note that it does not necessarily follow that such h is in fact applied. For example, the author has
sucessfully used the rule that if the predicted h is in the range 0.7 to 1.7, a value of h = 1 is used.
The reason for this strategy is that the application of a nonunitary h generally demands additional
residual calculations.
From the preceding considerations it follows that a “best” value of h need not be know precisely;
e.g., there is little practical difference between h = 1.8 and h = 2.0. This low accuracy requirement
allows wide latitude in simplifying optimal-stepsize formulas so they are practical to use.

§22.3. Accelerator Models

Two models may be use to derive optimal stepsize formulas. At an equilibrium solution, r = 0 and
(if not a critical point) u̇ = 0. Accelerators based on these two models are studied in the following
subsections.

22–3
Section 22: ACCELERATORS AND LINE SEARCH 22–4

§22.3.1. Tangent Accelerators for Fixed λ

We begin by studying accelerators based on the zero-velocity model. If the nonlinearities are
smooth, and an estimate of K̇ or of K̇u̇ can be easily procured, a technique due to Park [3] may be
used for estimating h. Take the fixed-λ second-order relaxation equation in which the term Ẅr is
neglected:
Kü + (K̇ + WK)u̇ = 0. (22.1)
Treat this equation with the forward Euler integrator on u̇:
u̇(k+1) − u̇(k) = h ü(k) , (22.2)
and require that h be such that
u̇(k+1) ≈ 0, (22.3)
whence
(k)
(K̇ + W(k) K(k) ) u̇(k) h ≈ K(k) u̇(k) . (22.4)
For fixed λ we have ṙ = Ku̇, which can be rewritten as
(k)
(K̇ u̇(k) + W(k) ṙ(k) ) h ≈ ṙ(k) . (22.5)
This is a highly undetermined system for h. Three main avenues of research open up here, leading
to a scalar stepsize, a full-matrix stepsize and a diagonal-matrix stepsize, respectively. Only the
first case is treated in some detail; the others are relegated to Remarks below.
Scalar Stepsize. The straightforward way to solve (22.5) is to premultiply both sides by a “weight-
ing” vector w. Park [3] proposes to use w = u̇, a choice that makes h a Rayleigh quotient with
respect to the direction u̇; the underlying idea being to anihilate the dominant error term. The result
is
T
u̇(k) ṙ(k)
h= T (k) T
(22.6)
u̇(k) K̇ u̇(k) + u̇(k) W(k) ṙ(k)
This formula can now be specialized to various Newton-like methods by inserting the appropriate
W.
Remark 22.1. Other weighting vectors may be chosen. For example, u, ṙ or r. The latter choice yields
accelerators of “minimal residual form”, which are studied later as part of the r ≈ 0 model.

Conventional Newton. If W = I, (22.6) specializes to

T (k)
1 u̇(k) K̇ u̇(k)
h CNM = , where ν= . (22.7)
1+ν T
u̇(k) ṙ(k)
If K is constant (linear system), then ν = 0 and h = 1, which of course gives the correct answer.
If the structure is stiffening (softening) in the u̇ direction, then h is less (greater) than 1, which is
the right trend. Numerical experiments with this accelerator suggest that the “averaged” stepsize
1
h H alley = 12 (1 + ). (22.8)
1+ν
gives marginally better performance; for a scalar equation the use of in fact yields Halley’s third
order iteration function; see e.g. Traub [18, p. 232].

22–4
22–5 §22.3 ACCELERATOR MODELS

Remark 22.2. For the quadratic scalar equation r = u 2 = 0, u (0) = 0, we have ṙ = 2u u̇, K̇ = 2u̇, ν = −0.5
and h CNM = 2, which yields the exact solution u = 0 in one step. Similarly, for the equation r = u m = 0 we
get h CNM = m and again we arrive at the solution in one step.

Remark 22.3. At noncritical points the performance of h CNM is not so dramatic as above; in fact it tends to
“overshoot” the solution by the same amount as h = 1 tends to undershoot it. This is the rationale behind the
Halley accelerator.

Modified Newton . Insertion of W = KF into yields

T
u̇(k) ṙ(k)
h MNM = T (k) T
(22.9)
u̇(k) K̇ u̇(k) + ṙ(k) F ṙ(k)

If the term involving K̇ is neglected,

T
∗ u̇(k) ṙ(k)
h MNM = T
(22.10)
ṙ(k) F ṙ(k)

which is more easily calculable given the information available in a MNM iteration. (A similar
simplification in (80) would be too drastic, for it would reduce it to h CNM = 1; but K̇u̇ is more
easily calculable in CNM.) Accelerators for the damped Newton method (73) can be constructed
in a similar manner.
Formulas such as (79) and (82) are called tangent accelerators as they depend only on the state at
pseudotime tk and t-derivatives there. But the presence of rates such as K̇ and ṙ causes computational
difficulties because the rates can be readily calculated only in simple problems. The problem can
be overcome by using secant information as discussed later.

Remark 22.4. Stepsize Matrix. A second approach to solving (78) is to transmute h into a “stepsize matrix”
H given by

H = (K̇ + WK)−1 K. (22.11)

This may be inserted to produce the iteration family

(k)
u(k+1) = u(k) − (K̇ + W(k) K(k) )−1 W(k) r(k) , (22.12)

into which we may replace K̇ by the rank-two estimate derived from Quasi-Newton formulas in Felippa [19].
This approach is presently unexplored.

22–5
Section 22: ACCELERATORS AND LINE SEARCH 22–6

Remark 22.5. Diagonalised Stepsize Matrix. Yet a third approach that merits attention is to transmute h into
a diagonal matrix, H D . This effectively assigns stepsizes component by component, which is in line with the
empirical “alpha method” of Nayak and Basu [20]. To get the stepsize h i associated with the i th component
of u, premultiply both sides by eiT , which is a row vector with unit i th component and zero otherwise. The
result is
T
ei(k) ṙ(k)
hi = T (k) T
(22.13)
ei(k) K̇ u̇(k) + ei(k) W(k) ṙ(k)
Appropiate W may now be inserted and rates replaced by secant relations.

Tangent Accelerators for Variable λ

Let now λ vary during the corrective process. Take the following second-order relaxation equation
which results from neglecting the terms Ẇr and q̇λ̇:

Kü + (K̇ + WK)u = qλ̈ + Wqλ̇. (22.14)

Treat this equation by the integrators

u̇(k+1) − u̇(k) = h ü(k) , λ̇(k+1) − λ̇(k) = h λ̈(k) , (22.15)

and again determine h from

u̇(k+1) ≈ 0, λ̇(k+1) ≈ 0. (22.16)
Upon replacing Ku̇ − qλ̇ by ṙ in the resulting equation, we obtain again the previous result.
Therefore, all previous formulas remain valid for the variable-λ case provided that the assumptions
made above hold. Note that the result is independent of the constraint condition assumed for
increment control.

Secant Accelerators
The presence of rates such as K̇ and ṙ in the tangent accelerators gives rise to computational
difficulties in complex nonlinear problems. More practical is to use secant-type (finite difference)
information to obtain secant accelerators. The idea is to formally replace

K̇ → K(k) = K(k+1) − K(k) ,

u̇ → d(k) = u(k+1) − u(k) = u(k+1) − u(k) , (22.17)
(k) (k+1) (k)
ṙ → g =r −r ,

in the tangent accelerators. This formal procedure, however, has the disadvantage that K is not only
unwieldy to calculate in CNM but unavailable in MNM. Fortunately in the tangent accelerators K̇
appears always in the combination K̇u̇; consequently K is always postmultiplied by d in its secant
counterparts. A convenient expression for K d can be obtained from the Quasi-Newton formulas,
which are multidimensional generalizations of the scalar secant method. Pertinent expressions for
the Broyden family [21] are worked out in Fletcher [19], from where one gets

K̇u̇ → K(k) d(k) = g(k) − K(k) d(k) . (22.18)

22–6
22–7 §22.3 ACCELERATOR MODELS

Substituting the secant expressions yields

g(k) d(k)
T
h̄ = T
(22.19)
g(k) T W(k) d(k) + d(k) (g(k) − K(k) d(k) )

On replacing the appropriate W, secant counterparts of h CNM , h MNM and h ∗MNM are obtained as
T
1 d(k) K(k) d(k)
h̄ CNM = , ν= , (22.20)
1+ν g(k) T d(k)

g(k) d(k)
T
h̄ MNM = T
. (22.21)
g(k) T Fg(k) + d(k) (g(k) − K(k) d(k) )

g(k) d(k)
T
∗
h̄ MNM = , (22.22)
g(k) T Fg(k)

Remark 22.6. The important equation (22.18) is not sensitive to the choice of Quasi-Newton formula. In fact
one always has
K(k+1) d(k) = (K(k) + K(k) ) d(k) = g(k) , (22.23)
which is called the Quasi-Newton condition and serves as a point of departure for defining such formulas.

A Secant Accelerator Algorithm

For the computer implementation it is important to note that secant accelerators involve u(k+1) and
r(k+1) . Since these are not known until the stepsize h has been selected, the process is inherently
iterative. The basic structure of a model algorithm follows.
1. Select an initial h̄ equal to the final stepsize used in the previous iteration step; if the first step,
set h̄ = 1.

2. Compute u(k+1) and r(k+1) based on this h̄.

3. Calculate h̄. If this value differs substantially from the assumed one (say by more than ±50%)
repeat steps 2 and 3 but no more than a fixed number of times (typically 1).

4. If a h̄ has been accepted, use to calculate u(k+1) . Otherwise branch to a line search procedure.
Many stylistic variations are possible. For example, one might decide never to repeat steps 2–3.
This is the same as extrapolating h̄ from the previous-step information, which is economical but
potentially dangerous.

Minimum Residual Accelerator

Accelerators may also be derived from the r ≈ 0 model, in which the residual is viewed as a
function of the stepsize:
r(h) = r(u(k) + c(k+1) + hs(k) ) (22.24)

22–7
Section 22: ACCELERATORS AND LINE SEARCH 22–8

where s(k) is defined by (65). Note that for variable λ vector c(k+1) = v̄ λ(k+1) depends indirectly
on h. The stepsize h is to be chosen so that a norm of r is minimized. Most commonly used is the
scaled Euclidean norm:
r = 12 rT Dr, (22.25)
where D is a positive-definite scaling matrix (usually diagonal) A non-identity diagonal D is useful
when the residual components have different physical units. The choice of D ≡ F = K−1 (if
positive definite) is also of some interest because the residual norm becomes an energy norm.
The general approach to minimizing r (h) is through a line-search method as described in the next
subsection. But if the dependence of r on h is mildly nonlinear, an approximate minimizer can be
obtained by assuming a linear dependence of r on h:

r(h) ≈ h r(0) + (1 − h)r(1), (22.26)

where r(0) and r(1) denote the residual vectors evaluated at h = 0 and h = 1, respectively; if k > 0
then r(0) should be available from the previous iteration. The minimizing condition ∂r/∂h = 0
yields
r(0)T Dr(0) − r(1)T Dr(0)
ĥ = (22.27)
r(0)T Dr(0) − 2r(1)T Dr(0) + r(1)T Dr(1)
Inasmuch as the assumption (22.26) is only strictly valid for linear response behavior, it is recom-
mended to check whether
r (ĥ) < r (1) and r (ĥ) < r (0) (22.28)

holds before accepting ĥ. If this condition is not verified, a line search may be called for.

Remark 22.7. Unlike the accelerators based on the zero-velocity model, the formula (22.27) depends on the
corrective method only indirectly through r(1). This gives it an implementation edge over secant accelerators
in the sense that the stepsize calculations are less method dependent. On the other hand, its performance is
not usually as good as that of secant accelerators.

Remark
 22.8. If D is the identity matrix, ξ is the residual magnitude ratio ξ = r(1)/r(0) =
r(1) r(1)/r(0)T r(0) and ϕ is the angle subtended by r(1) and r(0), the above formula can be presented as
T

1 − ξ cos ϕ
ĥ = (22.29)
1 − 2ξ cos ϕ + ξ 2
which admits of a geometric interpretation studied in Felippa [22].

Remark 22.9. Is it worthwhile accepting a non-unitary ĥ? If so, an extra residual evaluation at ĥ would
be required. A detailed analysis by Felippa [22] indicates that ĥ should be accepted only if it is near 0.5
(underrelaxation) or near 2.0 (overrelaxation).

A Minimum Residual Accelerator Algorithm

The following model algorithm has some points in common with that described for a secant ac-
celerator, but is more cautious. It is offered as an illustration of the kind of strategies that may be
implemented in a general-purpose nonlinear analysis program.

22–8
22–9 §22.3 ACCELERATOR MODELS

1. Select initial stepsize h = 1. If the first iteration, calculate r(0).

2. Evaluate the residual r(1) and calculate ĥ from (100). If the first iteration, accept h = 1 but
record ĥ; else drop through.
3. Branch according to ĥ: (a) if ĥ is outside a “reasonable range” (for example, 0.4 to 2.5) for
two consecutive steps, branch to a line search procedure. (b) if outside the range for this step,
accept h = 1 if ĥ > 2.5, else set h = 0.5, and proceed to next iteration. (c) if 0.75 ≤ ĥ ≤ 1.7,
accept h = 1 and proceed to next iteration. Else drop through.
4. Set h = ĥ and evaluate r (h). If condition (101) is verified, accept this h and go to the next
iteration. Else branch to a line search procedure.

Line Search
A line search is a systematic procedure to find a stepsize h that approximately minimizes a residual-
magnitude measure, which for definiteness shall be assumed to be (98). Vector s(k) is called the
search direction.
Line search procedures are highly developed in optimization work, where they are essential com-
ponents of general-purpose function minimization programs: a line search is carried out at each
iteration. On the other hand, their role in nonlinear equation solving is secondary. Because line
search is fairly expensive in terms of residual evaluations, it is invoked only when a solution
procedure runs into severe difficulties.
The acceptance tests stated below involve the directional derivative g = ∂r (h)/∂h along s. If D is
constant,
∂r ∂r ∂λ
g(h) = = rT D = rT DK(s(k) + v̄ ) (22.30)
∂h ∂h ∂h
The last expression is computationally cumbersome because it involves the tangent stiffness matrix.
More practical is to use finite differences to estimate the factor ∂r/∂h.
Line search algorithms “backtrack” h in a systematic fashion until an acceptance condition is
verified. A widely used acceptance condition in optimization programs of the the early 1970s is
the Goldstein-Armijo (GA) rule which may be stated as

0 < −η1 hg(h) ≤ r (0) − r (h) ≤ −η2 hg(h), (22.31)

where η1 and η2 are scalars that satisfy 0 < η1 ≤ η2 < 1; typical values being η1 ≈ 0.1 and
η2 ≈ 0.9. The upper and lower bounds in (104) insure that h is neither “too large” nor “too small”.
In more modern optimization work [23], the GA rule is replaced by slope conditions of the form
 g(h) − g(h ) 
 1 
|g(h)| ≈   ≤ −ηg(0), (22.32)
h − h1
r (0) − r (h) ≥ −ξ hg(0). (22.33)
where 0 ≤ η < 1 and 0 < ξ ≤ 0.5, ξ ≤ η, and 0 ≤ h 1 < h. A value of η ≈ 0.8, which gives a not
very restrictive line search, can be recommended.

A Line Search Algorithm

22–9
Section 22: ACCELERATORS AND LINE SEARCH 22–10

There are many variations on this theme and readers interested in additional details are referred to
the abundant literature on practical optimization methods. The following algorithm is meant only
to illustrate the basic procedural steps.
1. Estimate g(0) by finite differences, for example

(r(0.1) − r(0))
g(0) ≈ r(0)T D (22.34)
0.1
If g(0) ≥ 0 exit with h = 0, which should trigger a refactoring. Else set h 1 = 0, h 2 = 10
(say), h = 1.

2. Evaluate r (h).

3. Test for (22.32). If not satisfied, compute ĥ by restricted interpolation from data at h 1 and h;
set h 2 = h, h = ĥ, and return to 2.

4. Test for (22.31). If satisfied, exit. If not, compute ĥ by restricted interpolation from data at h 1
and h; reset h 1 = h, h = ĥ and go to 2.

If the exit h < 0.1 (say), the iteration step should be abandoned and the stiffness matrix refactored
at h = 0.
As stressed, a line search procedure should be initiated only when erratic or divergent behavior is
suspected. This being the case, the initial steps of the search procedure should take advantage of
previously computed information such as r(0) and r(1). Similarly (22.27), suitably safeguarded,
can be utilized for the interpolation process in steps 3 and 4; its key advantage being that it does
not require g derivatives.

§22.4. References

[22.1] OTTER, J. R. H., “Computations for Prestressed Concrete Reactor Pressure Vessels Using Dynamic
Relaxation,” Nucl. Struct. Engng., 1, 61–75, 1965.
[22.2] UNDERWOOD, P. G., “Dynamic Relaxation — A Review,” Ch. 5 in Computational Methods
for Transient Dynamic Analysis, (eds. T. Belytschko and T. J. R. Hughes), North-Holland,
Amsterdam, 1983.
[22.3] PARK, K. C., “A Family of Solution Algorithms for Nonlinear Structural Analysis Based on the
Relaxation Equations,” Int. J. Num. Meth. Engng., 18, 1337–1347, 1982.
[22.4] RIKS, E., “The Application of Newton’s Method to the Problem of Elastic Stability,”, J. Appl.
Mech., 39, 1060–1065, 1972.
[22.5] CRISFIELD, M. A., “Incremental/Iterative Solution Procedures for Nonlinear Structural Analy-
sis,” in Numerical Methods for Nonlinear Problems, Vol. 1, (eds. C. Taylor, E. Hinton and
D. R. J. Owen), Pineridge Press, Swansea, U. K., 261–290, 1980.
[22.6] CRISFIELD, M. A., “An Incremental-Iterative Algorithm that Handles Snap-Through,” Computer
& Structures, 13, 55–62, 1981.
[22.7] KUBÍČEK, M. and HLAVÁČEK, V., Numerical Solution of Nonlinear Boundary Value Prob-
lems with Applications, Prentice-Hall, Englewood Cliffs, New Jersey, 1983.

22–10
22–11 §22.4 REFERENCES

[22.8] BAKHVALOV, N. S., Numerical Methods, Mir Publishers, Moscow, 1975.

[22.9] ORTEGA, J. M. and RHEINBOLDT, W. C., Iterative Solution of Nonlinear Equations in
Several Variables, Academic Press, New York, 1970.
[22.10] BOGGS, P. T., “The Solution of Nonlinear Systems of Equations by A-Stable Integration Tech-
niques,” SIAM J. Num. Anal., 8, 767–785, 1971.
[22.11] BRANIN, F. H. and HOO, S. K., “A Method for Finding Multiple Extrema of a Function of n
Variables,” in Numerical Methods for Nonlinear Optimization, (ed. F. A. Lootsma), Academic
Press, London, 1972.
[22.12] KELLER, H. B., “Global Homotopies and Newton Methods,” in Recent Advances in Numerical
Analysis, (eds. C. de Boor and G. H. Golub), Academic Press, New York, 1978.
[22.13] WACKER, Hj. (ed.), Continuation Methods, Academic Press, NY, 1978.
[22.14] ALLGOWER, E. L., “A Survey of Homotopy Methods for Smooth Mappings,” in Numerical
Solution of Nonlinear Equations, (eds. E. L. Allgower et al.), Lecture Notes in Mathematics
878, Springer-Verlag, Berlin, 1981.
[22.15] PADOVAN, J., “Self-Adaptive Predictor-Corrector Algorithm for Static Nonlinear Structural Anal-
ysis,” Report NASA CR-165410 to Lewis Research Center, The University of Akron, Akron, Ohio,
1981.
[22.16] BATHE, K. J. and DVORKIN, E., “On the Automatic Solution of Nonlinear Finite Element Equa-
tions,” Computer & Structures, 17, 871–879, 1983.
[22.17] FELIPPA, C. A., “Finite Element Analysis of Three-Dimensional Cable Structures,” in Compu-
tational Methods in Nonlinear Mechanics, (ed. J. T. Oden et al.), The Texas Institute for
Computational Mechanics, University of Texas, Austin, Texas, 311–324, 1974.
[22.18] TRAUB, J. F., Iterative Methods for the Solution of Equations, Prentice-Hall, Englewood
Cliffs, New Jersey, 1964.
[22.19] FELIPPA, C. A., “Dynamic Relaxation and Quasi-Newton Methods,” in Numerical Methods for
Nonlinear Problems 2, (eds C. Taylor, E. Hinton, D. R. J. Owen & E. Oñate), Pineridge Press,
Swansea, U. K., 27–38, 1984.
[22.20] BASU, A. K., “New Light on the Nayak Alpha Technique,” Int. J. Num. Meth. Engng., 6,
152–153, 1973.
[22.21] BROYDEN, C. G., “Quasi-Newton Methods and their Application to Function Minimization,”
Maths. Comput., 21, 368–381, 1967.
[22.22] FELIPPA, C. A., “Procedures for Computer Analysis of Large Nonlinear Structural Systems,” in
Large Engineering Systems, (ed. A. Wexler), Pergamon Press, Oxford, 60–101, 1977.
[22.23] GILL, P. E., MURRAY, W. and WRIGHT, M. H., Practical Optimization, Academic Press,
London, 1981.

22–11
 21
.

Newton-Like
Methods

21–1
Chapter 21: NEWTON-LIKE METHODS 21–2

TABLE OF CONTENTS

Page
§21.1. Introduction 21–3
§21.2. Newton Iteration as a Dynamical System 21–3
§21.2.1. Corrective Process for Fixed λ . . . . . . . . . . . . 21–4
§21.2.2. Corrective Process for Varying λ . . . . . . . . . . 21–6
§21.3. Relaxed Newton Methods 21–6
§21.4. Damped Newton Methods 21–6
§21.5. Chord and Modified Newton Methods 21–7
§21.6. Quasi-Newton Methods 21–8
§21.7. *Convergence of Modified Newton 21–8
§21. Exercises . . . . . . . . . . . . . . . . . . . . . . 21–10

21–2
21–3 §21.2 NEWTON ITERATION AS A DYNAMICAL SYSTEM

§21.1. Introduction

The conventional Newton method (CNM) described in Chapter 20 is hindered by two major short-
comings:
High cost. The tangent stiffness matrix Kk = K(uk , λk ) has to be formed and factored at each
iteration step.
Low Reliability. Convergence to the desired solution is not guaranteed unless the initial estimate
is sufficiently close. The method may diverge, or converge to an unwanted solution. This is quite
likely in the vicinity of bifurcation points.
Because of these shortcomings many variations of CNM, collectively called Newton-like methods,
have been proposed and implemented over the past four decades, with varying degree of success.
Among the most important are
1. Relaxed Newton Methods (RNM): for reliability
2. Damped Newton Methods (DNM): for reliability
3. Modified Newton Methods (MNM): for efficiency
4. Quasi-Newton Methods (QNM): for efficiency
As can be seen by the large number of variations, no modification can be said to be uniformly superior
to others. The above list covers the most important so-called Newton-like methods. Question arise,
however, as to how far the offsprings can deviate from the parent and still be called Newton-like.
Authors have different opinions in this matter. To further complicate things, problem-adaptive
combinations of these techniques are often used in advanced nonlinear solvers.
Some of these variants, notably the Relaxed Newton methods (RNM), are more easily derived
by interpreting the Newton method in the context of a dynamical system. This interpretation is
discussed next.

§21.2. Newton Iteration as a Dynamical System

Figure 21.1 sketches what the chief goal of the Newton method as corrector is: to allow large
incremental steps by eliminating the drift error.
As discussed in previous Chapters, the incremental phase is driven by the first order rate form ṙ = 0,
in which the pseudo-time t is measured by an “increment clock.” Penalize the drift error by adding
a term proportional to the residual r:
ṙ + W r = 0, (21.1)
in which W is a positive-definite residual weighting matrix, which for the moment is left arbitrary.
This is called a first order corrective form, and also a first order relaxation form. It obviously
reduces to ṙ = 0 on an equilibrium path r = 0. The job of the penalty term W r is to force the
solution trajectories of (21.1) to approach r = 0 as the pseudotime t runs along a “corrective clock”
See Figure 21.2.
Figures 21.1 and 21.2 are a bit deceptive in that they depict corrective processes for a one-DOF
problem. A more realistic state of affairs can be observed in Figure 21.3, which depicts trajectories
of a corrective process on a constraint surface for the case of two DOFs.

21–3
Chapter 21: NEWTON-LIKE METHODS 21–4

λ 2 3 Equilibrium Path
Co ∆λ1
rre
ct
1 ∆u1

t
dic
Pre
Increment Control
∆λ0 Constraint

0 ∆u0
u
R

Figure 21.1. Sketch of how an incremental-iterative solution method works.

To bring explicitly the stiffness matrix and incremental load vector into play, insert ṙ = Ku̇ − qλ̇
into the above and transfer qλ̇ to the right hand side:

Ku̇ + Wr = qλ̇. (21.2)

Remark 21.1. A second order corrective form generalizes (21.1) by taking the second order differential
equation
r̈ + W1 ṙ + W2 r = 0, (21.3)
in which W1 and W2 are weighting matrices. These have different function: W1 provides damping while W2
is a “conditioner.” This more general form is not analyzed here.

§21.2.1. Corrective Process for Fixed λ

Suppose that we are at {uk , λk } at which the tangent stiffness Kk is nonsingular. We want to move
to a new state {uk+1 , λk } closer to r = 0 while keeping λ = λk fixed. This can be done by treating
(21.2) with the Forward Euler integrator

uk+1 = uk + h u̇k . (21.4)

where h is the integration steplength. The integrated corrector equation is

uk+1 = uk − h Fk Wk rk , (21.5)

where F = K−1 = (∂r/∂u)−1 is a flexibility matrix. Calling d = uk+1 − uk the correction in

displacements and passing K−1 to the right hand side,

Kk d = −h Wk rk (21.6)

If now we take
W = I, h=1 for any k, (21.7)

21–4
21–5 §21.2 NEWTON ITERATION AS A DYNAMICAL SYSTEM

λ
corrective clock
t=t 0 r=0
t=t K

k
loc
al c
ent
rem
inc

0
u
R

Figure 21.2. Figure 21.2. Pseudo time t running along an “incremental clock”
interspersed by a“corrective clock.” K is the total number of corrective iterations.

λ Incremental contraint
surface c = 0

r=0

u2

u1

Figure 21.3. The corrective process for two degrees of freedom. The challenge is to
end at a solution no matter where one starts on the constraint surface.

we obtain the conventional Newton method (CNM) for fixed λ, as can be easily verified. A variant
called Relaxed Newton, discussed below, results by letting h be adjustable.

21–5
Chapter 21: NEWTON-LIKE METHODS 21–6

§21.2.2. Corrective Process for Varying λ

Suppose next that λ is to be let vary while satisfying the scalar constraint c( u, λ) = 0 that
controls the increment size. The corrective equation can be generalized as
   
ṙ + Wr 0
= (21.8)
ċ 0

Inserting ṙ = Ku̇ − qλ̇ and ċ = aT u̇ + g λ̇, in which a = ∂c/∂u and g = ∂c/∂λ, one gets
   
Ku̇ + Wr qλ̇
= (21.9)
a u̇ + g λ̇
T
0

Treat (21.9) with the Forward Euler integrator on both u and λ:

uk+1 = uk + h u̇k , λk+1 = λk + h λ̇k (21.10)

Integrating (21.9) with (21.10), followed by setting W = I and h = 1, yields the conventional
Newton method for general increment control treated in the previous Chapter. The verification is
the matter of an exercise.

§21.3. Relaxed Newton Methods

One commonly used variant of CNM aims to increase the reliability but not necessarily lower the
cost per iteration. This is done by deriving CNM from the dynamical process described in the
previous section, and letting the steplength h be a variable. The
    
K −q d r
=− , (21.11)
a T
g η c

where supercript k is suppressed from K, q, etc., to reduce clutter. Solving for d and η as explained
in the previous Chapter, one then corrects

uk+1 = uk + hd, λk+1 = λk + hη. (21.12)

It is understood that h may change from integration to iteration, that is, h = h k .

The method (21.11)-(21.12) is called the relaxed Newton-Raphson method, or RNR.1 There are
three possibilities as regards h k :
1. If h k < 1, the iteration step is said to be underrelaxed and h k is an underrelaxation parameter.
2. If h k > 1, the iteration step is said to be overrelaxed and h k is an overrelaxation parameter.
3. If h k = 1 for all k, RNM reduces to CNM.
How is the steplength h chosen? Rules to this effect are discussed in Chapter 22.

1 Some authors called this the damped Newton-Raphson method but that name is reserved here for the variant discussed
in the next section.

21–6
21–7 §21.5 CHORD AND MODIFIED NEWTON METHODS

§21.4. Damped Newton Methods

The Relaxed Newton Methods provide gains in reliability as long as the stiffness matrix is not
singular or ill-conditioned. But it does not help in the vicinity of critical points. For example if
K is exactly singular, system (21.11) is not solvable by the double RHS method discussed in the
previous Chapter, and the variable steplength device does not help.
Critical points may come in many flavors. In order of increasing traversal difficulty: isolated limit
points, isolated bifurcation points, initially singular structures, and clustered limit and/or bifurcation
points.
For the less difficult cases, moving away slightly from the singularity often works. Much tougher
is the case when stiffness matrix at the start of the analysis, or of an analysis stage, may be highly
singular. This happens, for instance, in some cable, pneumatic and biological structures that are
mechanisms in the reference configuration and acquire stiffness as they deform. For such cases a
variants collectively known as the Damped Newton Method or DNM, can be effective at the coist
of programming complexity. [The name of Regularized Newton is also used.]
DNM overcomes the singularity problem by adding a diagonal correction to the stiffness matrix.
Instead of (21.10) one solves
    
K + γ D −q d r
= , (21.13)
a g η c

where D is a nonnegative diagonal matrix γ ≥ 0 is a “numeric damping” coefficient, and k

supercripts have been omitted. The correction can then be applied with a steplength h:

uk+1 = uk + hd, λk+1 = λk + hη. (21.14)

If the damping coefficient γ is zero and h is unity, the CNM results. As γ is increased the method
approaches steepest descent if D = I and scaled steepest descent for general D. This has been useful
in conjunction with cable net structures that must traverse highly singular regions. Two choices for
D tried in that case are
rT Kr
D = βI, β= T
r r
T (21.15)
r K Kr K
D = βD K , β=
rTK r
where D K = diag K and r K = D−1
K r. Practical values for the damping coefficient γ may be
characterized as follows:

γ ≥1 very heavy damping

1 ≥ γ ≥ 0.1 heavy damping
0.1γ ≥ 0.01 moderate damping
0.01γ ≥ 0.001 light damping
The best results for cable net structures were obtained with light damping. Once the structure
acquires sufficient stiffness by deforming, the correction terms may be removed by setting γ = 0.

21–7
Chapter 21: NEWTON-LIKE METHODS 21–8

§21.5. Chord and Modified Newton Methods

The problem of high computational cost of CNR per step can be alleviated if the same stiffness
matrix is maintained for several iteration steps. This general class of methods, collectively known
as chord methods is based on the iteration scheme
    
K −q̄ d r
=− , (21.16)
aT g η c

uk+1 = uk + d, λk+1 = λk + η (21.17)

Here K and q̄ denote an approximation to K and q in some sense, which is maintained fixed for
several or all iteration steps. On the other hand r and c are changed at each iteration. Several
variants result of this general scheme result according to two criteria: (a) How K and q̄ are chosen
and updated. (b) How a and g are chosen and updated.
Two specializations of the chord method have proven effective in practice. If K = Kn , which is
the stiffness matrix at the start of the n th increment, which is kept fixed thereafter, the modified
Newton method (MNM) method results. There is a variant called delayed modified Newton method
(DMNM) for which K = K0 , which is the stiffness matrix evaluated after the predictor step, and
which again is kept fixed for all k.
Updating versions of MNM and DMNM, identified by acronyms UMNM and UDMNM, respec-
tively, emerge if K is allowed to vary during the iterative process. Several strategies to that effect
can be devised. Only three, ranging from the simplest to the most sophisticated, are mentioned
here:
1. Periodic update: Recompute K every m ≥ 1 iterations. The “period” m is chosen on the basis
of prior experience, relative computational cost of factorization versus solving, etc. Obviously
m = 1 gives back the conventional Newton method.
2. Residual monitoring: If the residual norm ||rk || does not steadily decreases over a certain
“subperiod” m ∗ , K is recomputed. Typically m ∗ = 3 or 4, which allows for “residual spikes”
common as K is reset. This strategy is best combined with the previous one by choosing m as
a multiple of m ∗ .
3. Progressive update: This merges chord methods with a nonunitary steplength h.
§21.6. Quasi-Newton Methods
Quasi-Newton (QN) methods represent a refinement of the MNM methods. The stiffness K is
updated at each iteration step with rank-one or rank-two matrices built up from information from
the previous iteration. In this way a better approximation of the actual stiffness matrix is obtained
while still avoiding revaluation and factorization.
The idea comes from the field of optimization, where QN methods (also called variable metric
methods) have enjoyed great success. They were proposed for solving nonlinear structural prob-
lems in the late 1970 with high hopes. Evidence shows, however, that the moderate improvements in
reducing the number of iterations to convergence does not compensate for the increase in program-
ming complexity and storage. The idea has some uses, however, in the derivation of accelerator
and secant formulas presented in Chapter 22.

21–8
21–9 §21.7 *CONVERGENCE OF MODIFIED NEWTON

§21.7. *Convergence of Modified Newton

(ASEN 5107 students pls ignore this advanced material. It is placed here for eventual development)
Assuming for simplicity that λ is kept fixed, then the limit relaxation equation becomes

Ku̇ = −K(x) (21.18)

where x = u − u(∞) is the distance to the equilibrium solution at t = ∞. This can be modally decomposed
as
ẏ = −µy (21.19)
in which µ are the roots of the symmetric eigenproblem K0 z = Kz and y are modal amplitudes. The
appropriate eigenvalue for the Newton direction u̇ can be estimated by the Rayleigh quotient

u̇T Ku̇
µ= (21.20)
u̇T Ku̇

The structural behavior can be characterized as follows.

1. If µ < 1, the structure is softening in the mode y;
2. If µ > 1, the structure is hardening in the mode y.
For the MNM to converge,
|κ| = |1 − hµ| < 1 (21.21)
If the structure softens, MNM converges but the converges rate deteriorates unless h is increased (overrelax-
ation). If the structure hardens, MNM diverges unless h is cut (underelaxation). But the structure hardens in
some modes while softening in others, MNM cannot be continued, and a refactoring of the stiffness matrix is
called for.
The preceding observations are well know to experienced investigators. They have observed that MNM works
quite well in problems when the structure experiences overall softening.

Remark 21.2. To reduce the variation of µ, one may reduce the incremental step, or proceed to reform the
stiffness matrix. In some programs the strategy control attempts to cut the incremental steplengths; after two
or three unsuccesful attempts the stiffness is reformed.

21–9
Chapter 21: NEWTON-LIKE METHODS 21–10

Homework Exercises for Chapter 21

Newton-Like Methods

EXERCISE 21.1 [A:10] (Very easy, just to get acquanted with the relaxation equation). Verify that (21.4) and
(21.5) followed by W = I and h = 1 lead to the Conventional Newton Method (20.25) for load control.

EXERCISE 21.2 [A:20] Starting from (21.9) and (21.10) derive the general form of the Relaxed Newton
method. Verify that if W = I and h = 1 this reduces to the Conventional Newton system for general
increment control defined by (20.12)-(20.14).

EXERCISE 21.3 [A:20] Find out which W leads to the Damped Newton system (21.13).

21–10
 20
.

Conventional
Newton
Methods

20–1
Chapter 20: CONVENTIONAL NEWTON METHODS 20–2

TABLE OF CONTENTS

Page
§20.1. Introduction 20–3
§20.2. Stage Analysis Review 20–3
§20.3. Problem Statement 20–3
§20.4. The Corrective Phase 20–4
§20.5. Solving the Newton Systems 20–5
§20.6. Termination Tests 20–5
§20.7. The Ordinary Newton Method 20–6
§20. Exercises . . . . . . . . . . . . . . . . . . . . . . 20–10

20–2
20–3 §20.3 PROBLEM STATEMENT

§20.1. Introduction

In the overview of solution methods given in Chapter 16 it was noted that solution methods based on
continuation generally included two phases: incremental and corrective. In the purely incremental
methods covered in Chapter 17–19 the corrective phase is absent. If the corrective phase is present,
the incremental formula simply functions as a predictor that provides a starting point for the
corrective iteration. The purpose of this iteration is to eliminate (or at least reduce) the drifting
error by moving towards the equilibrium path along the constraint hypersurface.
Solution methods that include a corrective phase will be collectively called corrective methods,
although perhaps a more appropriate name would be predictor-corrector methods. There are purely
corrective methods that lack a predictor phase entirely (for example, the orthogonal trajectory
accession method) but they have not proven important in practical applications.
The most important class of corrective methods pertains to the Newton-Raphson method and its
numerous variants: modified, modified-delayed, damped, quasi, and so forth. These are collectively
called Newton-like methods, and only require access to the past solution. In the present section we
study the conventional Newton method under general increment control.

§20.2. Stage Analysis Review

Let us recall that our purpose is to solve the residual equations

r(u, λ) = 0 (20.1)

over a loading stage as the control parameter is incremented from 0. As previously discussed the
additional equation that makes (20.1) determinate is the increment constraint equation

c(u, λ) = 0. (20.2)

Starting from λ = 0, we want to calculate a series of solutions

u0 , λ0 , u1 , λ1 , ... un , λn ... (20.3)

that characterizes numerically the response u = u(λ) while satisfying the residual equations (20.1)
within prescribed accuracy.
The purely incremental methods covered in the three previous Chapters compute a sequence of
values such as (20.3) by direct integration of the first-order rate equations Ku = q or Kr e f u =
q − f  . The methods considered here implement a corrective phase in which one iterates for
equilibrium while satisfying the increment constraint. The starting point for the corrective phase
is the solution predicted by the incremental method. Consequently, these methods are often called
incremental-iterative methods.

20–3
Chapter 20: CONVENTIONAL NEWTON METHODS 20–4

§20.3. Problem Statement

Assume that n incremental steps of the stage analysis have been performed. The last accepted
solution is
un , λn (20.4)
We want to compute the solution
un+1 , λn+1 (20.5)
that satisfies the nonlinear algebraic system

r(un+1 , λn+1 ) = 0,
(20.6)
c(un , λn ) = 0,
where
un = un+1 − un , λn = λn+1 − λn (20.7)
Although the above increment constraint is a special case of (20.2), it befits those most commonly
used in practical calculations.
The predicted solution
u0n , λ0n , (20.8)
is typically obtained by performing an incremental step as described in Chapters 17 and 18.

§20.4. The Corrective Phase

All that computations that follow pertain to the n th incremental step. Hence for simplicity we shall
omit the subscript n from the formulas.
Starting from the predicted approximation (20.8),

u0 ≡ u0n , λ0 ≡ λ0n (20.9)

the conventional Newton method applied to (20.6) generates a sequence of iterates

uk , λk , (20.10)

where k = 1, 2 . . . is an iteration step index.

The conventional Newton method is based on the truncated Taylor expansion of the system r = 0,
c = 0 about (uk , λk ):
∂r k+1 ∂r k+1
rk+1 = rk + (u − uk ) + (λ − λk ) + H.O. = 0,
∂u ∂λ (20.11)
∂c ∂c k+1
c k+1
= ck + (uk+1 − uk ) + (λ − λk ) + H.O. = 0.
∂u ∂λ
where ‘H.O.’ denote higher order terms that are quadratic or higher in the changes uk+1 − uk and
λk+1 − λk , and all derivatives are evaluated at (uk , λk ). Discarding such terms and recalling that
K = ∂r/∂u, q = −∂r/∂λ, aT = ∂c/∂u, g = ∂c/λ, we obtain for the corrections

d = uk+1 − uk , η = λk+1 − λk (20.12)

20–4
20–5 §20.6 TERMINATION TESTS

the linear algebraic system

    
K −q d r
=− , (20.13)
aT g η c

where
∂r ∂r ∂c ∂c
K= , q=− , aT = , g= , (20.14)
∂u ∂λ ∂u ∂λ

and all known quantities are evaluated at uk , λk . Note that for notational simplicity this superscript
has been kept out of d, η, r, etc, unless it is desirable to make the dependency on the iteration index
k explicit. If the tangent stiffness matrix K is of order N , the coefficient matrix of the linear system
(20.13) has order N + 1. This matrix is called the augmented stiffness matrix.
Note that although generally K is symmetric and sparse, the augmented stiffness is generally un-
symmetric (but see Exercise 20.2), and its sparseness is detrimentally affected by the augmentation.
It is therefore of interest to treat the linear system (20.13) with techniques that preserve those at-
tributes. The solution procedures described below make use of auxiliary systems of equations to
achieve that goal. The number of auxiliary systems depends on whether the tangent stiffness K
is nonsingular (regular points) or singular (critical points). For the latter we have to distinguish
between limit points and bifurcation points. In the present section we shall concentrate on the
treatment of regular points.

§20.5. Solving the Newton Systems

Recall from Chapter 4 that regular points of the system (20.1) are equilibrium solutions (u, λ) at
which the tangent stiffness matrix K is nonsingular. If this property holds, we can perform forward
Gauss elimination on (20.13) to get rid of d and produce the following scalar equation for η:

(g + aT K−1 q) η = −c + aT K−1 r. (20.15)

Let dr and dq denote the solution of the symmetric linear systems

Kdr = −r, Kdq = q. (20.16)

Then
c + aT dr
η=− , d = dr + ηdq . (20.17)
g + aT dq

It is seen that two right hand sides, r and q, have to be generally solved for at each Newton step. The
number reduces to one for k > 1, however, if modified Newton is used so that K is held fixed for
several steps and q does not vary. The last assumption holds in structural mechanics applications
if the loading is conservative and proportional. (The modified Newton method is described in the
next Chapter).

20–5
Chapter 20: CONVENTIONAL NEWTON METHODS 20–6

§20.6. Termination Tests

At which point should we stop the Newton iteration? There are several convergence criteria that
can be applied.
1. Displacement convergence test. The change in the last correction d of the state vector u, as
measured in an appropriate norm, should not exceed a given tolerance d . For example, using
the 2-norm (Euclidean norm) 
d = dT d ≤ d . (20.18)

2. Residual convergence test. Since the residual r measures the departure from equilibrium,
another appropriate convergence test is

r ≤ r . (20.19)

Some comments are now in order.

1. The two tests may be applied in an “and” or “or” matter as iteration stopping criterion. It is
also possible to combine both tests in the form of an “work change” criterion, for example

|rT d| ≤ d r (20.20)

2. Since d and r have usually physical dimensions, so do necessarily d and r . For a general
purpose implementation of Newton iteration this dependency on physical units is undesirable
and it is more convenient to work out with ratios that render the r and d dimensionless. For
example:
r
≤ r (20.21)
r0 
where r0 is the residual after the predictor step; now r can be dimensionless. A similar ratio
can be used for the displacement convergence test, but here the reference value should be a
total or accumulated displacement; for example:
d
≤ d (20.22)
u0 

3. Divergence Safeguards. The Newton iteration is not guaranteed to converge. There should
therefore be divergence detection tests that will cause the iteration to be interrupted. For ex-
ample, turning the above ratios around, divergence may be diagnosed if either of the following
inequalities occur:
r d
≥ gr , ≥ gd (20.23)
r 
0 u0 
where gr and gd are “dangerous growth” factors, for example gr = gd = 1000.
Occasionally the Newton iteration will neither diverge not converge but just “bounce around”
(oscillatory behavior). To avoid excessive wheel spinning in such cases it is always a good
practice to put a maximum number of iterations per step in the program. Typical limits might
be 20 to 50.

20–6
20–7 §20.7 THE ORDINARY NEWTON METHOD

§20.7. The Ordinary Newton Method

The Newton iteration discussed in the mathematical literature on solving nonlinear systems assume
that λ is held constant. This corresponds to λ control or load control in our terminology. As
previosuly explained, fixing λ makes critical points impassable. However, the resultant method
provides good examples to watch the typical behavior of the Newton iteration process.
If λ is kept constant the incremental step constraint is λn = n , which has derivatives 0 and 1
with respect to u and λ, respectively. System (20.13) simplifies to
    
K −q d r
=− , (20.24)
0 1 η 0
where c = 0 because the constraint is satisfied exactly. Since η = 0, the “bordering” disappears
and the Newton iteration reduces to
uk+1 = uk − (Kk )−1 rk , λk+1 = λk = λn + n (kept fixed) (20.25)
Note that the incremental load vector q disappeared entirely. This is the method found in standard
numerical analysis texts. This version is used in the examples that follow.
√
Example 20.1. The computation of the square root + a of a scalar number a > 0 by Newton iteration is
set up as follows. The square root satisfies the equation r (x) = f (x) = x 2 − a = 0. Starting from an initial
value x0 > 0, the Newton iteration computes
x k+1 = x k − f (x k )/ f  (x k ) = x k − ((x k )2 − a)/(2x k ). (20.26)
where prime denotes derivative with respect to x. The results for a = 3 and x0 = 1 are illustrated by the
Mathematica program below.

f[x_,a_]:=x^2-a; Df[x_]:=2*x;
a=3.; xk=1.; Print["x0=",xk];
For[k=0, k<6, k++, xkp1=xk-f[xk,a]/Df[xk];
Print["x",k+1,"=",xkp1//InputForm]; xk=xkp1];

x0=1.
x1=2.
x2=1.75
x3=1.732142857142857143
x4=1.732050810014727541
x5=1.732050807568877295
x6=1.732050807568877294
√
After six cycles the iteration yields 16 places of accuracy for 3. Note that the number of exact digits roughly
doubles from k = 2 onwards. This is typical of the Newton iteration once it “locks in” a root because the
process has asymptotically quadratic convergence.
The numerical process shown above is actually that used by mathematical software libraries of languages like
Fortran or C for the computation of the square root function. However the initial value is determined by a
scaled rational interpolant that gives 2-4 digits of accuracy for x0 ; as a result only 2 or 3 cycles are needed to
achieve double-precision accuracy for most inputs.

20–7
Chapter 20: CONVENTIONAL NEWTON METHODS 20–8

Example 20.2. This example is more typical of a structural application. Consider the residual equation for a
2-DOF system    
r1 u 1 + 2u 31 − u 22 − 2λ
r(u, λ) = = (20.27)
r2 3u 2 − 2u 1 u 2 − λ
The tangent stiffness matrix is
∂r  
1 + 6u 21 −2u 2
K= = (20.28)
∂u −2u 2 3 − 2u 1
For λ = 1 the residual equations have three real roots which to six digits of accuracy are
     
1 1.18400 1.66726
u1 = , u2 = , u3 = . (20.29)
1 1.58227 −2.98938

The following Mathematica programs starts from the initial values u 01 = u 02 = 0.8 and quickly finds the
nearest root u 1 = u 2 = 1, delivering 16 digits of accuracy after 5 cycles:

ClearAll[r,u1,u2,lambda];
lambda=1;
r[u1_,u2_]:={{u1+2*u1^3-u2^2-2*lambda},{3*u2-2*u1*u2-lambda}};
Kt[u1_,u2_]:={{1+6*u1^2,-2*u2}, {-2*u2,3-2*u1}}
uk={{0.8},{0.8}};
Print["Starting v0=",uk//InputForm];
For [k=0, k<5, k++,
{u1,u2}={uk[[1,1]],uk[[2,1]]};
ukp1 = uk - Inverse[Kt[u1,u2]].r[u1,u2];
Print["Cycle k= ",k," u",k+1,"=",ukp1//InputForm];
uk=ukp1 ];
{u1,u2}={uk[[1,1]],uk[[2,1]]};
Print["Final residual=",r[u1,u2]];

Starting v0={{0.8}, {0.8}}

Cycle k= 0 u1={{1.025426944971537002}, {0.9719165085388994309}}
Cycle k= 1 u2={{1.001827210881738689}, {1.005246766090385063}}
Cycle k= 2 u3={{0.999984431106495672}, {0.9999493398104184104}}
Cycle k= 3 u4={{0.9999999985779087286}, {0.9999999955786530219}}
Cycle k= 4 u5={{0.9999999999999999893}, {0.9999999999999999659}}
-18 -17
Final residual={{-7.48099 10 }, {-1.25767 10 }}

But changing the initial values to u 0 = 0.8 and u 1 = 1.1, which is even closer to the u 1 = u 2 = 1 root, the
process converges to the second root in (20.29), reaching 16 digits of accuracy after 8 cycles:

ClearAll[r,u1,u2,lambda];
lambda=1;
r[u1_,u2_]:={{u1+2*u1^3-u2^2-2*lambda},{3*u2-2*u1*u2-lambda}};
Kt[u1_,u2_]:={{1+6*u1^2,-2*u2}, {-2*u2,3-2*u1}}
uk={{0.8},{1.1}};
Print["Starting u0=",uk//InputForm];
For [k=0, k<5, k++,

20–8
20–9 §20.7 THE ORDINARY NEWTON METHOD

{u1,u2}={uk[[1,1]],uk[[2,1]]};
ukp1 = uk - Inverse[Kt[u1,u2]].r[u1,u2];
Print["Cycle k= ",k," u",k+1,"=",ukp1//InputForm];
uk=ukp1 ];
{u1,u2}={uk[[1,1]],uk[[2,1]]};
Print["Final residual=",r[u1,u2]];

Starting u0={{0.8}, {1.1}}

Cycle k= 0 u1={{1.188636363636363636}, {1.325}}
Cycle k= 1 u2={{1.215044904251438747}, {1.718220285975100653}}
Cycle k= 2 u3={{1.189744133901914791}, {1.602104015495415557}}
Cycle k= 3 u4={{1.184187697535010493}, {1.582880662857412937}}
Cycle k= 4 u5={{1.183998614947570215}, {1.582271181104649434}}
Cycle k= 5 u6={{1.183998417328768284}, {1.582270556284133866}}
Cycle k= 6 u7={{1.183998417328558547}, {1.582270556283474425}}
Cycle k= 7 u8={{1.183998417328558548}, {1.582270556283474426}}
-19
Final residual={{0.}, {-8.67362 10 }}

This illustrates the “finicky” nature of Newton iteration. It can do (and often does) the unexpected, such as
diverging or converging to the “wrong” root. In fact the a whole subset of fractal or chaotic mathematics
is devoted to the understanding of “domains of attraction” of roots. Because of this capricious behavior, in
practical use of the Newton corrector numerous safeguards are implemented to avoid surprises. But the whole
subject is too lenghty for coverage in an introductory treatment.

20–9
Chapter 20: CONVENTIONAL NEWTON METHODS 20–10

Homework Exercises for Chapter 20

Conventional Newton Methods

EXERCISE 20.1 (A:15) If the case of load control, the incremental stepsize constraint reduces to

c = λn − n = 0 (E20.1)

where n is prescribed. After the predictor step, c = 0. Show that the Newton method reduces to a standard
form that requires only the solution of one of the auxiliary systems (20.16).

EXERCISE 20.2 (A:15) What algebraic constraint c = 0 would make the augmented tangent stiffness sym-
metric?

EXERCISE 20.3 (C:25) Enlarge the program that solves Exercise 18.2 by using conventional Newton as
corrector, the incremental stepsize constraint

c = |sn | − n = 0, (E20.2)

(where = n is kept fixed throughout) and the termination tolerance |rk /r0 | ≤ 10−3 or 20 iterations, whichever
occurs first. Suggested values for n for all n is 0.01, but the results should be insensitive to that choice.

EXERCISE 20.4 (A:30) (Advanced, paper level) Obtain the expression of the augmented stiffness if the
orthogonal trajectory accession method is used and write out the Newton algorithm for this case.

20–10
 19
.

Pseudo-Force
Incremental
Methods

19–1
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–2

TABLE OF CONTENTS

Page
§19.1. Pseudo Force Formulation 19–3
§19.2. Computing the Reference Stiffness and Internal Force 19–4
§19.3. Integration of Pseudo-force Rate Equation 19–5
§19.3.1. Forward Euler Integration under Load Control . . . . . . 19–5
§19.3.2. Pseudo-Force Extrapolation . . . . . . . . . . . . 19–6
§19.3.3. Iterative Improvement . . . . . . . . . . . . . . 19–6
§19.4. Numerical Stability 19–7
§19.5. Accuracy Control 19–8
§19.6. Secant Estimation of n  19–8
§19.7. General Increment Control 19–8

19–2
19–3 §19.1 PSEUDO FORCE FORMULATION

The disadvantages of purely incremental methods in terms of solution “drift” and high computational
expense were recognized as applications of nonlinear finite element analysis expanded to cover
wider classes of problems. Many techniques aimed at avoiding these difficulties were proposed
and tested during the late 1960s and early 1970s. From the mass of experience accumulated during
this period, two principal strategies emerged.
For path-independent, smooth nonlinearities typified by finite deflection and nonlinear-elastic be-
havior, the power of Newton-like corrective methods was eventually recognized. In such problems,
incremental methods were relegated to the secondary role of predictors for starting a corrective
process.
For path-dependent material nonlinearities, however, purely incremental methods have remained
important because of the reasons noted in Chapter 18. These methods underwent modifications
aimed primarily at reducing the computational expense while retaining the advantages of numerical
stability, implementation simplicity and physical transparency. Unfortunately these goals, being
largely contradictory, can only be met half-way. The most successful attempt in this direction has
been the development of pseudo-force incremental methods, which are covered below.

§19.1. Pseudo Force Formulation

In the pseudo-force reformulation of incremental methods the pervasive role that the tangent stiffness
matrix plays in the methods discussed in Chapters 17-18 is relaxed. Instead the deviation from a
“reference linear response” is collected in a pseudo-force vector. This approach allows a reference
stiffness matrix to be reused over many incremental steps. Since this avoids having to repeatedly
assemble and factor the tangent stiffness matrix, the gain in speed per step over the conventional
incremental methods may be very substantial in two- and three-dimensional problems. There is no
free lunch, however, for the speed gain is counterbalanced by two disadvantages:

1. Pronounced accuracy loss as nonlinearities become severe, which may force extremely small
increments to be taken.

2. Increasing danger of numerical instability, especially in “hardening” portions of the response.

The pseudo-force method can be explained more conveniently by starting from the “force balance”
form (4.9) of the residual:
r = f − p = 0, (19.1)
where f and p are the internal and external (applied) force vectors, respectively. Furthermore for
simplicity we shall assume the separable form (4.10), that is f(u) = p(λ).
Decompose the internal residual force f as follows:

f = Kr e f (u − ur e f ) + n, (19.2)

where Kr e f is a nonsingular reference stiffness matrix that is kept fixed as long as possible and ur e f
is the state at λ = λr e f (usually 0). The deviation

n = f − Kr e f (u − ur e f ), (19.3)

19–3
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–4

is called the pseudo-force vector. Inserting (19.2) in the residual expression (19.1) we get

r = Kr e f (u − ur e f ) + n − p = 0. (19.4)

Differentiating (19.4) with respect to λ yields the rate form

r = Kr e f u + n  − q = 0, (19.5)

where q = ∂p/∂λ is the incremental load vector. Since Kr e f is assumed nonsingular, solving for
u yields
u = Kr−1 
e f (q − n ) = v ,
f
(19.6)
Vector v f is called the pseudo incremental velocity. It plays a similar role to that of v = K−1 q in
the conventional incremental methods discussed in Chapters 14 and 15.
Remark 19.1. The pseudo-force vector n may be viewed as a “force deviation from the linear reference
response” Kr e f (u − ur e f ). Now if the structural response is linear so that

K0 u = p, (K0 constant), (19.7)

choosing Kr e f equal to the linear stiffness K0 and ur e f = 0 gives n = 0. For this reason many authors call n
the nonlinear force vector.

Remark 19.2. Differentiating (19.3) with respect to u gives, for a separable residual,
∂r ∂n
= K = Kr e f + Knon , Knon = . (19.8)
∂u ∂u
If Kr e f is kept equal to the linear stiffness throughout, Knon is called the nonlinear stiffness matrix.

§19.2. Computing the Reference Stiffness and Internal Force

A very common choice for Kr e f , though far from the only one, is the stiffness K0 at the start λ = 0
of the response-calculation stage. Choosing ur e f = u0 accordingly, (19.3) becomes

r = K0 (u − u0 ) + n − p = 0. (19.9)

Response calculation procedures based on the choice (19.9) are sometimes called the initial stiffness
method.
Another common strategy is to start with the initial stiffness and continue with as many incremental
steps as possible, resetting Kr e f = Kn if numerical stability or accuracy problems are detected at
the (n + 1)th step. Still a third approach is to keep Kr e f for a preassigned number of incremental
steps, say m, and update Kr e f at steps 0, m, 2m, . . ., unless numerical stability or accuracy problems
are encountered.
The effectiveness of pseudo-force methods in finite element programming depends largely on the
ability to compute the internal force vector f directly on an element-by-element basis. Recall, for
example, that in the core-congruential formulation of the Total Lagrangian (TL) description,

f= GT si bi d V0 , (19.10)
V0

19–4
19–5 §19.3 INTEGRATION OF PSEUDO-FORCE RATE EQUATION

where V0 is the reference volume, G the transformation between displacement gradients and physical
degrees of freedom, si are PK stresses in the current configuration, and bi are the vectors defined
in §8.5. Note that expression does not explicitly involve material properties, and is consequently
applicable to problems with material nonlinearities. More precisely: the only requirement for using
(19.9) is the availabilibity of the stresses si in the current configuration whereas the procedure by
which such stresses are obtained is irrelevant.
Remark 19.3. In finite element work the synonyms initial force method, initial stress method and initial strain
method have been associated with restricted versions of what we call here pseudo-force methods. These names
focus attention on various physical interpretations of the calculation of the f term. A heated controversy as to
which version was the best took place in the late sixties; from current perspective such arguments have only
historical interest.

§19.3. Integration of Pseudo-force Rate Equation

In this subsection we assume that the calculation of f, given the necessary ingredients to apply
(19.5), is more practical than that of f  . An estimation of this rate is done through finite difference
approximations. This leads to very simple and fast implementations at the cost of numerical
reliability and accuracy. A more expensive but reliable alternative technique for evaluating f is
discussed later.

§19.3.1. Forward Euler Integration under Load Control

The simplest incremental algorithm results on treating (19.5) by the forward Euler method with a
backward-difference estimation of f  and assuming that the increments of λ are prescribed:

nn = fn − Kr e f (un − ur e f ),
nn − nn−1
nn = ,
λn−1 (19.11)
un = Kr−1 
e f (qn − nn ) λn = vn λn ,
f

un+1 = un + un , λn+1 = λn + λn .

where as usual we denote

fn ≡ f(un ), nn = n(un ), etc.
The pseudo incremental velocity vn = Kr−1 
f
e f (qn − nn ) plays the role of the incremental velocity
vector, as can be seen by comparing the advancing equations (17.7) for purely incremental methods.
The scheme (19.11) applies if n ≥ 1 and as long as Kr e f and ur e f are kept fixed. At the start
n = 0, nn−1 = n−1 is not known. But if Kr e f = K0 and ur e f = u0 , n = 0. The same condition
is applicable when Kr e f and ur e f are reset, if one chooses the tangent stiffness at ur e f as reference
stiffness.
This advancing scheme has poor accuracy characteristics unless nonlinearities are mild (say within
±20% of the reference response). There are three ways of improving accuracy: pseudo-force ex-
trapolation, iteration and resetting the reference stiffness. The first two are described in subsections
below.

19–5
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–6

Example 19.1. Solve the residual equation

r = 5u − u 3 − λ = 0, for λ = 4, (19.12)

using the pseudo-force method (19.11) with two λ increments of 2.0, K r e f = K 0 = 5 and u r e f = u 0 = 0. The
exact solution on the fundamental path is v(λ) = v(4) = 1. Since q = −∂r/∂λ = 1, and n = 5u − u 3 − 5u =
−u 3 , the rate form is r  = 5u  − (u 3 ) − 1 = 0.
The first increment, with λ0 = 2 specified, is

n 0 = 0, u 0 = K r−1 
e f (q0 − n 0 ) λ0 = 0.4, u 1 = u 0 + u 0 = 0.4, λ1 = λ0 + λ0 = 2.

The second increment, with λ1 = 2 again, is

n 1 = −(u 31 − u 30 )/ λ0 = −0.032, u 1 = K r−1 

e f (q1 − n 1 ) λ0 = 0.4128,

u 2 = u 1 + u 1 = 0.8128, λ2 = λ1 + λ1 = 4.

Repeating these computations with 4, 8 and 17 equal increments of λ gives 0.8461, 0.8841 and 0.9190,
respectively. As can be observed the accuracy attained is low. Table 19.1 compares these values with those
obtained with other methods.

§19.3.2. Pseudo-Force Extrapolation

Accuracy improves if the first of (19.10) is replaced by a central difference estimator:
nn+1 − nn−1
nn = (19.13)
λn + λn−1
Since nn+1 is not known, it has to be predicted by extrapolation. The simplest extrapolator is
λn
nn+1 ≈ nn+1
P
= n(un+1
P
), with P
un+1 = un + (un −un−1 ) . (19.14)
λn−1

For constant λ, the predicted un+1

P
is simply 2un − un−1 . The advancing algorithm is identical to
(19.11) with the second equation replaced by (19.13) and (19.14).
The result of applying this technique to the example equation (19.12) with fixed λ increments is
presented in Table 17.1. The accuracy obtained now is similar to that of the conventional purely
incremental method with Forward Euler. Also given there are the results of using the midpoint rule
(17.8), which as can be seen delivers higher accuracy.

§19.3.3. Iterative Improvement

Another way to improve accuracy while avoiding the reset of Kr e f is to iterate on un+1 while keeping λn
fixed. To derive an iteration scheme, write the residual form (19.4) at n + 1 and n, subtract, and solve for un :

un = Kr−1
e f (nn+1 − nn ) − qn λn (19.15)

Let k be an iteration step index and u1n+1 be the value obtained from the increment equation. The resulting
iterative scheme is 
ukn = Kr−1e f (nn+1 − nn ) − qn λn ,
k
k = 1, . . . (19.16)
n+1 = un + un .
uk+1 k

19–6
19–7 §19.4 NUMERICAL STABILITY

Table 19.1 Computed Incremental Solutions for (19.12)

Load Control, Equal λ Increments

Steps PFI-FE PFI-FE-X PI-FE PI-MR

1 0.8000 0.8000 0.8000 0.8850
2 0.8128 0.8512 0.8425 0.9360
4 0.8461 0.8835 0.8884 0.9760
8 0.8841 0.9139 0.9276 0.9888
17 0.9190 0.9407 0.9565 0.9964
32 0.9470 0.9617 0.9755 0.9990
64 0.9675 0.9768 0.9868 0.9997
128 0.9812 0.9867 0.9931 0.9999
PFI-FE: Pseudo-force incremental with Forward Euler (19.11)
PFI-FE-X: PFI-FE with extrapolation (19.13)-(19.14) for nn
PI-FE: Purely incremental with Forward Euler (17.7)
PI-FE: Purely incremental with Midpoint Rule (17.8)

Using the fact that q is independent of u, this can be rewritten in the “subincremental” form
−1 k
n+1 = un+1 − Kr e f rn+1 .
uk+1 (19.17)
k

If Kr e f ≡ Kn , this is precisely the modified Newton-Raphson (MNR) method with unit steplength. This
shows that the the iterated pseudo-force incremental method is a MNR method with an arbitrary selection of
reference stiffness. The properties of these methods are investigated in Chapters that deal with Newton-Raphson
corrective methods.

§19.4. Numerical Stability

The purely incremental tangent-stiffness methods studied in Chapters 17-18 are highly stable if some mild
precautions are heeded. On the other hand, pseudo-force methods are much less robust. For a single degree
of freedom, the homogeneous model equation corresponding to (17.10) is
∂n 
v  = K r−1  −1
e f n = Kr e f v = Av  , (19.18)
∂v
where A = K non /K r e f . Unlike (17.10) this is no longer a differential equation in v, v  but a difference equation
in v  , with A as amplification number. The iteration process to solve this equation is stable if

|A| < 1. (19.19)

This condition is independent of the stepsize λ. It is seen that the key for numerical stability is that the
reference stiffness “dominates” the nonlinear stiffness in the sense (19.20). That is, from that standpoint it is
better to overestimate K r e f .
The generalization of (19.19) to N degrees of freedom is
∂n 
u = Kr−1  −1
e f u = Kr e f u = Kr−1 
e f Knon = Au . (19.20)
∂u

19–7
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–8

Stability is controlled by the N × N amplification matrix

∂n 
A = Kr−1 = Kr−1
e f Knon . (19.21)
ef
∂u
Assume that this amplification matrix A has eigenvalues µi , i = 1, . . . N , and let µ = maxi |µi |. The condition
for numerical stability is
µ < 1. (19.22)

This condition holds regardless of the stepsize λ. Likewise, the iteration (19.17) converges only if the
condition (19.22) holds.
A practical estimator for this eigenvalue is
|Kr−1
e f (nn − nn−1 )|
µ≈ . (19.23)
|un − un−1 |

Remark 19.4. This result is another aspect of the close relationship between pseudo-force incremental methods
and modified-Newton corrective methods. If nonlinearities are substantial, the method diverges regardless of
the increment length used.

Remark 19.5. It is beneficial from a stability standpoint to have Kr e f “dominate” Knon . This happens in
softening structures when the elastic stiffness is selected as Kr e f , and explains the success of the method in
plasticity analysis. On the other hand, if the structure hardens as λ increases (examples: cable and pneumatic
structures), the stability condition is easily violated.

Remark 19.6. The simplest cure to numerical instability is to recompute the reference stiffness. Another (as
yet unexplored) possibility is to correct Kr e f with a rank-one matrix.

§19.5. Accuracy Control

Accuracy control may be effected as in the case of conventional incremental methods if one sub-
stitutes v f for v.
§19.6. Secant Estimation of n 
The finite difference estimators for n described in §17.3 are easy to implement but decidedly suffer
from a lack of robustness unless the problem is only mildly nonlinear. An alternative estimate
of n  can be obtained through the following “secant approximation” technique. This estimate is
computationally slower but more reliable. Recall that
n = p − Kr e f (u − ur e f ). (19.24)
n = q − Kr e f u = q − Kr e f (K−1 q) = (I − Kr e f K−1 )q. (19.25)
We now replace the exact inverse of the tangent matrix by a secant approximation:
K−1 ≈ Fs . (19.26)
Fs is a low-rank correction (typically rank one or two) of Fr e f = Kr−1
e f that is constructed on the
basis of the following increments:
us = un−1 = un − un−1 , rs = r(un , λn ) − r(un−1 , λn ), (19.27)

19–8
19–9 §19.7 GENERAL INCREMENT CONTROL

Note that the residual rs is computed by holding λ constant and equal to λn and is not rn − rn−1 .
Two choices for the inverse stiffness secant approximant are the Davidon-Fletcher-Powell (DFP)
rank-two update formula

us usT Fr e f rs rsT Fr e f
Fs = Fr e f + − , (19.28)
usT rs rsT Fr e f rs

and Davidon’s rank-one update formula

( us − Fr e f rs )( us − Fr e f rs )T
Fs = Fr e f + . (19.29)
( us − Fr e f rs )T rs

These formulas are collectively called Quasi-Newton updates in the numerical analysis literature,
although to be more precise what we have shown above is just the first member of such updates.

§19.7. General Increment Control

The preceding developments assume that λn is prescribed. But we can readily extend the pseudo-
force technique to general increment control by following the procedures discussed in Chapter 18.
For that we must replace vn by
vnf = Kr−1
e f (qn − ṅn ). (19.30)
where the superposed dot denotes derivative respect to the pseudotime parameter t pertinent to the
increment strategy chosen. Of particular importance is arclength control, in which t becomes the
arclength s. Criteria for stability and accuracy are readily converted to this case.

19–9
 18
.

Purely Incremental
Methods:
General Control

18–1
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–2

In this Chapter we continue the development of the purely incremental methods under general
incremental control conditions. A general nonlinear transformation of state and control parameters
to pseudo-time is introduced. Computable forms of this constraint are specialized to arclength
and hyperelliptic control. The computational implications of these decisions are discussed, and a
practical implementation of hyperplane-distance arclength control outlined.

§18.1. Parametric Form

In previous Chapters it was noted that continuation solution methods under λ (load) control have
difficulty traversing limit points. This shortcoming can be circumvented through the use of more
general increment control schemes. This generalization can be practically effected by adjoining
algebraic constraint equations such as those listed in Chapter 16.
These more general forms of increment control can be described on a uniform basis as follows.
Because the main effect of enforcing the constraint c(u, λ) = 0, is to link the increments of u
and λ, we express the response in the pseudo time parametric form

u = u(t), λ = λ(t). (18.1)

If t ≡ λ we would of course regress to the λ-control parametrization

u = u(λ), (18.2)

of the equilibrium path, and nothing would be gained. But the pseudotime t in (18.1) is now at our
disposal and we can try to do better. Differentiating (18.1) with respect to t we get

du = u̇ dt, dλ = λ̇ dt = (1/ f ) dt, (18.3)

where we have called for convenience

1 dt
f = = . (18.4)
λ̇ dλ

§18.1.1. Requirements
The incremental path equation becomes
∂r(u, λ) ∂r ∂r
ṙ = = u̇ − λ̇ = Ku̇ − q/ f = 0. (18.5)
∂t ∂u ∂λ
At limit points dλ = 0. By “smooth traversal” of a limit point is meant that t varies regularly as a
“vehicle odometer” as that limit point is crossed. Consequently f = dt/dλ must go to infinity at
a limit point. This reasoning shows why t ≡ λ does not work, because if so f = 1 everywhere. It
also follows that the relation between λ and a useful t must be necessarily nonlinear, for if we take
t = c1 λ + c2 with c1 = 0, f is always 1/c1 .
It is computationally desirable, however, that in “almost linear” portions of the response f ap-
proaches a finite value, because in the limit we would like to recover the conventional methods of
linear structural analysis.

18–2
18–3 §18.1 PARAMETRIC FORM

§18.1.2. The Arclength Choice

Now which quantity becomes infinite at limit points? An obvious choice is the incremental velocity
v = K−1 q or, more precisely, some norm of it. A particularly attractive choice in light of its
geometric significance is 
f = 1 + vT v, (18.6)
This corresponds to taking
dt ≡ ds = f dλ = v du + dλ, (18.7)
As shown in Chapter 4, ds is the differential arclength of the response curve in state-control space.
Consequently s is the arclength traversed along this curve measured from some arbitrary point,
such as the last solution. Note the following attractive features of this choice:

(a) At limit points f → ∞, as required.

(b) At turning points where v → 0, f → 1, so (unlike state control) traversal controlled by

increasing arclength is not affected.

(c) If the response is linear f maintains a constant value.

The arclength choice is no panacea, however, because three computational problems remain, listed
below in order of increasing difficulty.

(d) The exact arclength sn from the last computed solution at Pn (un , λn ) to another point P(u, λ)
on the equilibrium path is given by the path integral
  
sn = ds = 1 + vT v ds. (18.8)
Pn →P Pn →P

But as written this expression is not directly computable because it requires knowledge of
v = K−1 q along the equilibrium path that emanates from the last solution, which is precisely
what we want to compute. Consequently an increment control constraint such as

sn =
n , (18.9)

makes no computational sense. This minor difficulty is eliminated by using an approximation

sn to the arclength. The approximation is directly or indirectly effected by introducing an
appropriate constraint equation, as discussed below.

(e) The choice (18.6) intermixes state parameters, which have generally physical dimension (of
displacement), with the dimensionless scalar 1. This mixture can introduce difficulties in
that the computed solutions are not invariant with respect to the choice of state parameter
dimensions. It can be corrected by employing appropriate state-scaling techniques as discussed
in Chapter 4.

(f) Traversal of bifurcation points remain difficult or impossible without additional “tricks”. The
key roadblock is that the continuous parametrization (18.1) breaks down at bifurcation points,

18–3
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–4

because it cannot simultaneously represent the two or more branches that intersect there.1
Circumventing this difficulty within the context of purely incremental methods is not easy,
and consideration is deferred to later Chapters.
We now study two computable approximations to the arclength parametrization.

§18.2. Hyperplane Distance Control

The simplest computable approximation to the exact arclength constraint sn =

n is sn =
n ,
where sn is the hyperplane distance (4.23) to the last solution un , λn . Thus the increment control
constraint is 
sn = (vnT un + λn )/ f n =
n , where f n = 1 + vnT vn . (18.10)

in which
n , which controls the magnitude of the n th incremental step, is either specified or auto-
matically adjusted through some accuracy control rule, as discussed below.
Substituting un = vn λn into the above we get the Forward Euler incremental scheme


n
vn = K−1
n qn , f n = 1 + vnT vn , λn = ± ,
fn (18.11)
un+1 = un + vn λn , λn+1 = λn + λn .

It remains to select the sign of λn and the magnitude of

n . The proper sign for λn can be
chosen according to the criteria discussed in §16.7. If the positive external work criterion (16.16)
is used, λ is chosen to have the sign of qT v. As noted there, this simple rule fails at bifurcation
and turning points, wherein the angle criterion discussed there should be used.
As for the magnitude of
n , two possibilities exist. Either
n is kept constant and equal to the given

0 , or it may be automatically adjusted in an adaptive scheme. If the latter strategy is adopted,

parameter  is used to bound the local error as explained in §16.5. Without going through the
derivation, which is given in §16.4 below, the result is

2
n−1

n = (18.12)
an−1

where an−1 is estimated by (16.23) in which v is replaced by v/ f . This must be complemented by

a minimum arclength travel distance condition:

n ≥
min , (18.13)

which avoids “getting stuck” at limit points under certain conditions, as well as a maximum arclength
travel distance condition

n ≤
max , (18.14)
to avoid surprises in hitting rapidly changing portions of the response.

1 A second order rate form is in fact required in the vicinity of a simple bifurcation point.

18–4
18–5 §18.3 GLOBAL HYPERELLIPTIC CONSTRAINT

One simple way to define these minmax values is to set

0

min = ,
max =
0
f ac , (18.15)

f ac
where the factor
f ac , which is typically 1 to 20, is part of the input data. If
f ac = 1 this rule
effectively forces fixed a fixed
n =
0 .
Remark 18.1. The essential difference between the arclength and load control schemes is the arclength
form does not have trouble crossing limit points. As the limit point is approached, λ is driven to zero and
automatically changes sign upon crossing it. The only numerical danger is that of hitting the singularity exactly
so that the factorization of K fails; however, handling of this emergency in the computer implementation is
not difficult.

Remark 18.2. In a poorly scaled problem the previous scheme should be modified by replacing f by its scaled
equivalent: 
fn = 1 + vnT S2 vn , (18.16)

where S2 is a diagonal scaling matrix. Some appropriate choices for S2 are discussed in the next Chapter.

§18.3. Global Hyperelliptic Constraint

As a second example, if we adopt the unscaled global hyperelliptical constraint (16.27) reproduced
here for convenience
an2 unT un + bn2 =
2n , (18.17)
where an , bn and
n are scalars given at each step, we get the formula

n
λn =  , (18.18)
± an2 vnT vn + bn2

which is followed by solving for un .

For poorly scaled problems, or problems in which u collects quantities of different physical mag-
nitudes, we should used the scaled form
an2 unT S2 un + bn2 =
2n ,
where the diagonal matrix S2 is chosen to take care of scaling. Then

n
λn =  (18.19)
± an2 vnT S2 vn + bn2

One interesting possibility is to choose S2 = K. This may be viewed as an energy constraint

because vT Kv = qT v is external incremental work. With that choice, and adopting the positive-
external-work criterion to choose the sign of λ, we have

λn =  sign(qnT vn ). (18.20)

+ an |qn vn | + bn
2 T 2

18–5
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–6

Remark 18.3. Equation (18.19) represents a “closed” constraint surface (a hyperellipse in state-control space)
and thus it may be thought of as “safer” than the hyperplane distance constraint, which is “open”. Although
this consideration has some merit if a corrective phase that moves on that surface follow, it has no weight
in a purely incremental method. One drawback of the hyperelliptic form is the need to choose an and bn in
addition to
n at each step. Although this adds flexibility, it can complicate the implementation and require
more run-time decisions.

Remark 18.4. The scalar uT K u is not necessarily positive if K is indefinite; hence the need for taking
the absolute value of qT v in (18.20).

Remark 18.5. Diagonal scaling on displacement increments provides an intermediate choice between unscaled
and K-scaled forms. That is, choose S2 = diag(K).

§18.4. Accuracy Control

Assuming that t = s so that arclength increment control is used, the Taylor expansion about a solution point
reads  dv 
f
   
u vf  ds 
= s + 12   (s)2 + O(s 3 ), (18.21)
λ (1/ f ) d(1/ f )
ds
√
where v f = v/ f = v/ 1 + vT v. The truncation error is the quadratic term in s. Proceeding as in the load
control case we obtain    
ev |dv f /ds|
= 12 (s)2 . (18.22)
eλ |d(1/ f )/ds|
where ev and eλ are the norms of the local truncation errors associated with u and λ, respectively. We shall focus
on controlling accuracy by monitoring ev , since sufficient accuracy on λ generally follows. Again defining as
in §14.5 the desired local accuracy level by the scalar , we arrive at the rule: adjust
n by
2|sn 1 |
|sn | ≤ (18.23)
an−1
in which an−1 is estimated by
vn / f n − vn−1 / f n−1
an−1 = , (18.24)
vn / f n
Observe that this is the same as the adjustment rule (14.23)–(14.24) but with v replaced by v f = v/ f .

§18.5. Numerical Stability

Following the same procedure as §14.4 it can be shown that the linearized problem that governs numerical
stability becomes  du   ∂v 
f ∂v f
 
 ds   ∂u ∂λ  u
 dλ  =  ∂(1/ f ) ∂(1/ f )  λ , (18.25)

ds ∂u ∂λ
√
in which v f = v/ f = v/ 1 + vT v. The amplification matrix A is that shown in brackets above. So far this
problem appears to be analytically intractable and no general conclusions may be drawn so far. However,
numerical evidence show that arclength-control incremental methods are as stable as those using load control.
Thus the empirical rule of Chapter 14 may be used; namely if
is adjusted for accuracy, setting  < 1 takes
care of stability.

18–6
18–7 §18.8 ASSESSMENT OF PURELY INCREMENTAL METHODS

§18.6. What Happens at a Limit Point

Suppose that the computed solution process approaches an isolated limit point at which K has the
null eigenvector z normalized to unit length, that is

Kz = 0, zT z = 1. (18.26)

As the critical point is approached, it can be shown (using the theory of inverse iteration) that

dλ → 0, v = K−1 q → αz, α → ±∞, (18.27)

That is, v tends to become parallel to z although its norm goes to (plus or minus) infinity. But the
f -normalized v does approach the normalized eigenvector:
v
v f = 1/ f v = √ → ±z. (18.28)
1 + vT v
and the incremental equations approach
 
u ±z
=
(18.29)
λ 0
Therefore the f -normalization of v automtically take care of aligning the incremental direction
normally to the λ axis.
Remark 18.6. Some minor safeguards remain: if the tangent stiffness is evaluated at or very near the limit
point, K(u) may be numerically singular, in which case a simple remedy is to change u by a tiny amount and
try again (a more theoretically sound approach based on penalty spring stabilization is discussed later). This
simple technique is used in the computer program for Exercises 18.1–18.5. Also if
is automatically adjusted
by accuracy requirements it must not be allowed to fall under a minimum value.

§18.7. An Automated Incremental Algorithm

We are now in a position to give the outline of a arclength-controlled incremental algorithm essen-
tially based on the scheme described previously. This is done in Table 18.1. The steps listed therein
are for a single stage. (In programs that accept multistage analysis, “stop” is replaced by a save,
exit, check error conditions and recover if possible, reset and restart process with the next stage.
Although programming that sequence can be elaborate, it does not involve conceptual difficulties.)

§18.8. Assessment of Purely Incremental Methods

The simplest load-controlled incremental scheme described in Chapter 17 has been extensively
used as a stand-alone method (that is, not combined with equilibrium corrections) in early imple-
mentations (1955-1965) of finite-element-based nonlinear structural analysis. It survives in this
primitive form in a surprisingly large proportion of finite element computer programs; particularly
those aimed at treating highly nonlinear material behavior. From current perspective it suffers from
two serious disadvantages.
Drift Error. The residual force vector r never enters the calculations. Consequently, the deviation
from the equilibrium path due to the propagation and accumulation of local integration errors cannot

18–7
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–8

Table 18.1 - Arclength-Controlled Incremental Solution

Forward Euler Procedure for a Single Stage
At λ = 0 we know u = u0 . We want to advance λ in a external-work-increasing process
until either the norm of the state vector u exceeds u max , the magnitude of λ exceeds
λmax , the number of incremental steps exceeds n max , or an impassable bifurcation point
is reached. Parameters  < 1, and
0 are specified. Set n = 0 and perform the following
steps.

Step 1. Form and factor stiffness matrix Kn . If the factorization fails on account of
singularity, perturb u by a tiny amount and repeat. If this failure repeats after
a certain number of tries, stop with appropriate error message.
 side qn and solve Kn vn = qn for the incremental velocity vn .
Step 2. Form right-hand
Form f n = 1 + vnT vn .
Step 3. If an adaptive stepsize scheme is used, adjust
n as per §18.2, else keep
n =
0 .
Set λn =
n / f n , and give it the sign of qnT vn .
Step 4. Compute un = vn λn . Advance un+1 = un + un and λn+1 = λn + λn .
If adaptive
control is used, save v f to be used in the estimation of an−1 .
Step 5. If |un+1 | exceeds u max , or |λn+1 | exceeds λmax , or n exceeds n max , stop. Else
set n ← n + 1 and return to Step 1.
If scaling of v is introduced because of the physical dimensionality discrepancy between
λ and u, the simplest  implementation is to define a reference length L r e f as part of the
inputs. Then f n = 1 + ṽnT ṽn , in which ṽn = vn /L r e f .

be eliminated by corrective iteration. In practical terms, this means that realistic estimates of the
response tracing accuracy can be obtained only by rerunning the problem with several increment
sizes.

Computational Expense. To keep the drifting error down, many small steps may be required,
particularly in “difficult” regions of the response. But at each step the stiffness matrix must be
formed and factored. This can be an expensive proposition in two- and three-dimensional problems.
To reduce the stiffness recalculation cost, the “pseudo-force” methods discussed in Chapter 19 have
been used extensively in plasticity and viscoelasticity calculations. These methods are affected,
however, by serious numerical stability difficulties.
What are the advantages? First, given the early applications of finite element methods (see Remark
below), incremental methods are quite easy to program as extensions of linear analysis codes. The
absence of the residual vector is in fact helpful, since linear analysis does not require it whereas
the stiffness matrix is always available. And the underlying concept (follow the physics) is readily
understood by practicing engineers. The virtues of simplicity and physical transparency should

18–8
18–9 §18.8 ASSESSMENT OF PURELY INCREMENTAL METHODS

never be underestimated!
Second, in problems that exhibit strong path-dependency, as typified by flow plasticity models,
the many-small-steps requirement is not necessarily a hindrance, since the increments have to be
kept small anyway to avoid unacceptable element-level errors (for example, yield surface drift) in
material-law calculations.
Finally, the high frequency of stiffness matrix assembly and refactorization has a silver lining: it is
difficult to miss critical points. The fact that purely incremental methods have trouble tracing post-
buckling branches beyond bifurcation points is irrelevant to situations in which the determination
of such points, rather than traversing them, is the main objective of the analysis.

Remark 18.7. The preference of incremental over corrective methods in the early implementation of nonlinear
finite element analysis has historical roots. Finite element methods were invented in the aircraft/aerospace
industry where linear analysis dominates. (This also helps to explain the initial popularity of the force method,
which is ill-suited for nonlinear analysis.) Furthermore, early excursions in the nonlinear world involved fairly
mild nonlinearities. Thus, developers of displacement-based finite element codes passed naturally from the
linear stiffness equations
Ku = q (18.30)
to the incremental form
K u = q,
(18.31)
unew = uold + u

and gave scant thought to the nonlinear equilibrium equations.

18–9
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–10

Homework Exercise for Chapters 16-18

Incremental Solution Methods

Exercises 18.1 and 18.2 pertain to the response analysis by purely incremental methods, of the nonlinear
response of the two-bar arch example structure of Chapter 8, which was treated by the TL description.
The structure properties are: span S = 2, height H = 1 (hence the rise angle is α = 45◦ ), elastic modulus
E = 1 and cross section area A0 = 1. The applied loads are f X = 0 and f Y = λF where F = −1 is constant.
This gives a downward vertical applied load if λ > 0. For this rise angle, the first limit point of the Total
Lagrangian model analytically occurs at load level λ L ≈ 0.136 and a vertical deflection of u Y L ≈ −0.425.
Traversal of limit points is done through a positive-work advancing criterion: qT v > 0.
Exercises 18.1 and 18.2 should be done with the Mathematica Notebook called IncSolTwoBarArch.nb,
which has been posted on the course Web site. This is a very rough conversion from the ancient Fortran code
used in previous offerings. [The conversion took much longer than anticipated despite the simplicity of the
code.]
Despite the roughness of this implementation, it can be used to illustrate the behavior of two integrators:
Forward Euler (FE) and Midpoint Rule (MR), combined with three increment control strategies: Load Control
(LC), Displacement Control (DC) and Arclength Control (AC). [The Classical Runge Kutta (RK4) integrator
will be added later as an Exercise.] One significant advantage of Mathematica over Fortran is the availability
of built-in graphics. Thus response plots, for example, can be immediately generated as part of the output,
helping quick visualization of method performance.

EXERCISE 18.1 [D:15] Prepare a hierarchical diagram of Cells 1-12 of the IncSolTwoBarArch.nb Note-
book, beginning with the main program given in Cell 12. Note which module calls which and write down the
purpose of each module in one or two lines along the module name. Return this diagram as answer to the
homework.*

EXERCISE 18.2 [C:20] Run the four scripts (main programs) in Cells 12 through 15 of IncSolTwoBarArch.nb.
Briefly explain what they do, and how the four method combinations driven by them stack up in terms of (i)
robustness in traversing limit points, (ii) accuracy in locating the first limit point (analytical values are given
above), and (iii) accuracy in crossing λ = 0 at u Y = −1. Please attach the four response plots of u Y versus λ
to your returned homework (do not bother with printouts). (I forgot to include method labels in the ListPlot
commands to identify which method is which; these may be written by hand).
Note: before you can run those scripts, Cells 1-11 should be initialized. A quick way to accomplish that in
version 3.0 is to click Kernel → Evaluate → Evaluate Initialization. Do this twice to get rid of error message
boxes.

* If you are not sure of what a hierachical diagram is, go to the IFEM Web page:
http://caswww.colorado.edu/courses.d/IFEM.d/Home.html
and look up Exercise 15.7 in Chapter 15.

18–10
18–11 Exercises

EXERCISE 18.3 [C:25] Consider the one-DOF nonlinear problem (not a structure) governed by the force
residual equation
(λ − 1)2 + (u + 1)2 − 2 = 0 (E18.1)
√
This represents a circle of radius 2 and center (1, −1) in the (λ, u) plane. Suppose that one starts at the
reference state u = λ = 0 and tries to trace the response, with a constant stepsize, by going around the circle
counterclockwise and eventually returning to the origin. In doing so you need to cross two turning points and
two limit points. While traversing limit points is easy using Arclength Control and a postive work criterion,
turning points are trouble.
Program from scratch an incremental solution to this problem using the Midpoint Rule, Arclength Control,
constant
, and the angle criterion2 to keep a positive traversal direction, and try it.3 If your implementation
works all the way you are ahead of all commercial nonlinear FEM codes in this regard.

EXERCISE 18.4 [C:25] The fixed-step classical 4th order Runge Kutta integrator (RK4) is implemented in
Cell 10 of the IncSolTwoBarArch.nb Notebook.4 It is identified as integ="RK4". Check visually if the
implementation is correct by comparing to the description, for example, in Numerical Recipes.
Run the same arch problem as in Exercise 8.2 using RK4 combined with Load Control and Arclength Control;
experiment with ell and nmax and report on whether you can obtain high accuracy (see 8.2 for how to assess
it) with a fairly large stepsize, say ell=0.1.

EXERCISE 18.5 [C:25] It was √ shown in Chapter 8 that if the rise angle α of the two-bar arch example structure
exceeds 60◦ , i.e. tan α > 3, it will fail by bifurcation first. For example S = 2 and H = 2 would do it
because tan α = 2. However if you set those inputs and run the programs in IncSolTwoBarArch.nb the
bifurcation point will be completely masked; you will see only the symmetric solution passing two limit points.
One simple technique to make “dumb incremental solvers,” like those provided in the Notebook, pay attention
to bifurcation points is to inject artificial imperfections. This can be done, for example, by putting a fictitious
but tiny load system that disturbs the symmetric response. For example, define the reference crown load in
force as {-0.001,-1} instead of {0,-1}. Set S = H = 2 and play with the tiny X -force, the increment
length ell and number of steps nmax until you see a decent tracing of bifurcation post-buckling: at a certain
load level u X will increase rapidly, signaling that the arch is buckling horizontally. To see that better, do the
ListPlot of u X versus λ collected in list uXvslambda found near the bottom of the driver cell. Comment on
what combination of method and solution parameters let you succeed.
Note (1). Bifurcation experiments can be found in the Notebook IncBifSolTwoBarArch.nb also posted on
the web site.
Note (2). If you take the response long enough you may be able to have the structure return to the primary
symmetric path upon passing through the second bifurcation point, but that may take lots and lots of steps
since the implementation uses a constant stepsize
.

2 The computation of the angle between vn and vn−1 is illustrated in the posted Notebook. It is saved as part of the solution
table, although it is not used in the solution procedure therein. The variable is called a or an and is actually the cosine
of that angle, which is simply the dot product of those velocity vectors normalized to unit length.
3 The positive-work criterion qT v > 0 fails because one needs to release work along some parts of the response trajectory.
In fact the net work on doing a complete circle is zero.
4 For a description of the RK4 algorithm see any book on numerical methods for ODEs. For example the widely used
Numerical Recipes in Fortran; it is presented in Section 16.1 of the second edition.

18–11
 17
.

Purely Incremental
Methods:
Load Control

17–1
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–2

TABLE OF CONTENTS

Page
§17.1. Governing Differential Equation 17–3
§17.2. Forward Euler Integration 17–3
§17.3. More Accurate Integration 17–4
§17.4. Numerical Stability of Forward Euler 17–5
§17.5. Accuracy Monitoring 17–7

17–2
17–3 §17.2 FORWARD EULER INTEGRATION

Incremental methods calculate the nonlinear response through the numerical integration of a rate
form of the equilibrium equations as the stage control parameter λ is varied. In the nomenclature
introduced in Chapter 16, we can characterize these as predictor-only methods: no corrective
iterations to recover equilibrium are performed. They are also known as step-by-step, initial-
value or marching methods in the engineering literature. The qualifier “purely” distinguishes these
incremental methods from those that make use of the pseudo-force concept, and which are covered
in Chapter 19.
The present Chapter emphasizes purely incremental methods in which the first-order rate equations
are integrated by a forward Euler scheme. Furthermore, for simplicity we focus on the simplest
increment control strategy: load control, in which λ is treated as an independent variable. This
restriction allows subjects such as stability and accuracy to be discussed in a straightforward manner.
An arclength-parametrized version, which allows the introduction of more robust increment control
techniques and the automatic traversal of limit points, is presented in the following Chapter.

§17.1. Governing Differential Equation

Recall the first-order rate equation ṙ = Ku̇ − qλ̇ = 0 specialized to t ≡ λ:

r = Ku − q = 0, (17.1)

where primes denote differentiation with respect to λ. If the stiffness matrix is nonsingular, this
equation uniquely relates the differential of u to that of λ:

du
u = = K−1 q = v, (17.2)
dλ

where as usual v denotes the incremental velocity vector. Purely incremental methods with λ as
independent variable are based on the numerical integration of (17.2) to generate an approximate
response u = u(λ) given the initial condition

u = u0 at λ = 0. (17.3)

Remark 17.1. The exact integral of (17.2) with the initial conditions (17.3) is

r(u, λ) = r0 (17.4)

where r0 = r(u0 , 0). Thus an initial equilibrium error does not decay even if the integration were carried out
exactly. This is the source of the drifting error that afflicts purely incremental methods. The error committed
at each step moves the equilibrium point to a neighboring curve in the incremental flow (see Figure 16.2).
Consequently the solution may “drift away” quickly when the incremental flow paths “flare out” from the
equilibrium path.

17–3
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–4

§17.2. Forward Euler Integration

In the remaining subsections of this Chapter we consider that the incrementation process is controlled
directly by varying the stage parameter λ, which thus assumes the role of independent variable.
This is tantamount to using the λ-control increment discussed in Chapter 16. This restriction is
removed in the next Chapter.
The simplest incrementation scheme is obtained by using the forward Euler integrator

un+1 = un + λ un , (17.5)

def
where n is the incremental step index, un = u(λn ) and

λn = λn+1 − λn , (17.6)

is the stage parameter stepsize. Treating (17.2) with this integrator yields the scheme

un = K−1n qn λn = vn λn ,

(17.7)
un+1 = un + un .

In the actual computer implementation of (17.7) the linear system Kn vn = qn is preprocessed by

assembling and factoring Kn . The right hand side qn is solved for to get vn . This is multiplied by
λn , which is either prescribed or (better) adjusted by the stepsize-control techniques discussed
below.
Remark 17.2. As discussed in Chapter 16, (17.7) is also the usual predictor for incremental-corrective methods.

§17.3. More Accurate Integration

To increase accuracy, more refined integration formulas have been proposed. An attractive second-
order choice is the explicit midpoint rule (also called Heun’s rule by some authors):

un+1/2 = un + 12 K−1
n qn λn ,
def def
Kn+1/2 = K(un+1/2 ), qn+1/2 = q(un+1/2 ), (17.8)
un+1 = un + K−1
n+1/2 qn+1/2 λn .

This scheme was used in the author’s thesis1 to treat problems with combined geometric and
material nonlinearities. The midpoint rule has attractive features for flow-plasticity studies, since
local elastic unloading can be detected during the first “trial” step and Kn+1/2 adjusted accordingly.
The same feature can be used to advantage in bifurcation analysis if a stiffness-determinant change
is detected between Kn and Kn+1/2 . But note that the stiffness matrix has to be formed and factored
twice per incremental step.

1 C. A. Felippa, Refined Finite Element Analysis of Linear and Nonlinear Two-dimensional Structures, Ph.D. thesis, Dept.
of Civil Engineering, University of California, Berkeley (1966)

17–4
17–5 §17.4 NUMERICAL STABILITY OF FORWARD EULER

Natural extensions of (17.8) are third and fourth-order Runge-Kutta (RK) formulas, which require
three and four stiffness evaluations and factorizations per step, respectively. These more refined
methods, however, are rarely used in structural mechanics for the amount of work per step is
considerable. Remark 17.8, however, indicates a possible niche for the classical fourth-order RK
in nonconservative problems.

§17.4. Numerical Stability of Forward Euler

Can the integration process (17.7) become numerically unstable? The subject is rarely mentioned
in the finite element literature. For simplicity we begin with the one-degree-of-freedom counterpart
u  = v of u = v. The right-hand side v = K −1 q is Taylor-series expanded in u = u − u n about
u n as
∂v
u  = u n + µ u + O(u 2 ) with µ = . (17.9)
∂u
For the linearized stability analysis only the homogeneous part of (17.9) is retained, which yields
the model equation
u  = µu. (17.10)

Consider the case in which µ is negative real and h = λ > 0. Then the solution u = u(λ) of
the model equation is exponentially decreasing as λ increases. The forward Euler integration is
absolutely stable2 if
|1 + hµ| ≤ 1, or h ≤ −2/µ. (17.11)
If h exceeds this value, the computed solution exhibits oscillatory instability. If µ is positive real the
solution of the model equation grows exponentially as λ increases and the forward Euler integration
is “relatively stable” for all h > 0.
Now if λ decreases so that h = λ < 0 the roles are reversed (cf. Remark 17.3). The stability
condition is h ≥ −2/µ if µ > 0. If λ is a load parameter, loading and unloading sequences may
be viewed as equally likely; consequently a safe stability constraint is

2
|λ| ≤ . (17.12)
|µ|

For the general system (17.2), let µi (i = 1, 2 . . . N , N being the number of degrees of freedom)
be the eigenvalues of the so-called amplification matrix

∂v ∂(K−1 q)
A= = . (17.13)
∂u ∂u
It is shown in Remark 5.8 that this matrix, although generally unsymmetric, has real eigenvalues if
the problem is conservative, i.e. K is the Hessian of a potential (u, λ) for fixed λ, and K is positive

2 C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall, Englewood Cliffs, N.J.
(1971)
L. Lapidus and J. H. Seinfield, Numerical Solution of Ordinary Differential Equations, Academic Press, New York
(1971).

17–5
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–6

definite. The eigenvalues are given by the eigensystems (5.22) or (5.23). Under such conditions a
safe increment is given by (17.12), where now

µ = µmax = max |µi |, i = 1, . . . N . (17.14)

i

If K does not depend on λ (as in the linear case) all µi vanish and the increment is unrestricted. If
K depends midly upon λ, eigensystems (5.21)–(5.22) show that the largest |µi | are associated with
the smallest eigenvalues of K, i.e. the fundamental stiffness modes (see also Remark 17.4).
Of course the actual calculation of all µi at each step would be a formidable computational task.
But the following finite-difference “path” estimate is easily obtained:
vn+1 − vn  vn+1 − vn  1 an
µ≈ = = , (17.15)
un+1 − un  vn  λn λn
where
√ an = vn+1 − vn /vn  and  denotes the 2-norm or Euclidean norm of a vector: x =
xT x. (an may be viewed as a kind of “incremental acceleration”.) Unfortunately this quantity
is not available until the n th step is completed, and to get a practical estimate we replace it by the
previous step estimate:
an−1 vn − vn−1 
µ≈ , an−1 = . (17.16)
λn−1 vn−1 
Insertion into (17.12) yields the stability condition

2 |λn−1 |
|λn | ≤ . (17.17)
an−1

If A can have complex eigenvalues, however, this simple rule does not apply (see Remark 17.8).
Remark 17.3. The fact that h = λ can be either positive or negative is a distinguishing feature of incremental
static analysis. In the numerical integration of actual dynamical systems, the time increment h = τ is never
negative; therefore stability results from conventional ODE theory should be used with caution.

Remark 17.4. Having the largest µi associated with the smallest eigenvalues of K represents another note-
worthy difference with dynamic problems. In the latter, the stability limit of explicit integrators such as (17.7)
is determined by the highest frequencies of the system. As discussed in Chapters dealing with dynamic relax-
ation methods, the difference is due to the fact that the matrix multiplying the highest derivative is the mass
(stiffness) in a dynamic (static) problem.

Remark 17.5. If the midpoint rule (17.8) is used, the stability limit remains the same for real µi .

Remark 17.6. Replacing vn−1  by vn  in (17.16) is inconsequential, as it is only a gross estimate. In fact,
choosing the smallest of the two norms would be the more conservative policy. A more serious objection is
the choice of the 2-norm unless the problem is well scaled and all degrees of freedoms have common physical
dimension. Otherwise a diagonal scaling matrix may have to be introduced as discussed in Chapter 4; for
example
|(vn − vn+1 )T S2 (vn − vn+1 )|
2
an−1 = . (17.18)
min(|vnT S2 vn |, |vn−1
T
S2 vn−1 |)

17–6
17–7 §17.5 ACCURACY MONITORING

Remark 17.7. As K approaches singularity, v → ∞ and the stable λ approaches zero. This is an
indication of the problems encountered by this type of incremental method at critical points. “Flat” limit
points can be traversed using the step-controlled parametric formulation discussed in the next Chapter. At
those points λ changes sign. But as the limit point becomes progressively sharper, numerical difficulties
increase. Bifurcation points, which in some sense may be viewed as infinitely sharp limit points, are difficult
to traverse without resort to either perturbation or special techniques that necessarily involve buckling mode
estimation, as discussed later.

Remark 17.8. If the problem is not derivable

√ from a potential or K is not positive definite, A may have complex
eigenvalues such as µi = αi + jβi , j = −1. Let h = λ. If hαi < 0, the appropriate stability condition
for forward Euler reads
(1 + hαi )2 + h 2 βi2 ≤ 1. (17.19)
If the imaginary component βi dominates, the stable stepsize may be sharply reduced over that of the potential
case, and if αi = 0 there is no stable h. It is not difficult to construct “load follower” problems that yield almost
imaginary µi . The morale is that purely incremental methods should be used with caution in nonconservative
problems. For this class of problems, third- and fourth-order Runge-Kutta methods do enjoy a substantial
stability edge; see, for example, the stability charts on page 120 of Lapidus and Seinfeld (referenced cited in
footnote 2).

§17.5. Accuracy Monitoring

If the response is twice differentiable, the local truncation error of the Euler integrator (17.5) at a
regular point is easily obtained from the truncated Taylor expansion about (un , λn ) as3
e = 12 (λ)2 uξ = 12 (λ)2 vξ , (17.20)
where the symbol vξ = uξ denotes the second derivative ∂ 2 u/∂λ2 = ∂v/∂λ evaluated at points
ξi ∈ [λn , λn+1 ], which generally differ from component to component. To assess the magnitude
e = e of this error we need an estimate of the norm of v . An obvious finite-difference estimate
for this quantity is (vn+1 − vn )/λn but — as in the stability analysis – λn and vn+1 are not
available until the step is completed. For a practical estimation of e we are forced to use previous
step values:
vn − vn−1 
e = ||e|| ≈ 12 (λn )2 . (17.21)
λn−1
For accuracy monitoring a convenient dimensionless measure is the ratio  of e to the increment
length un  = λn vn :
 
 λ  vn − vn−1 
 n 
 = 12   an−1 with an−1 = . (17.22)
 λn−1  vn 
Here we again denote by a a ratio similar to the one in (17.16) but with vn  in the denominator,
which is an inconsequential change. To strive for uniform local accuracy the basic idea is to specify
 (say  ≈ 0.1) and adjust the stepsize so that

2 |λn−1 |
|λn | ≤ . (17.23)
an−1

3 P. Henrici, Error Propagation for Difference Methods, Wiley, New York (1963).

17–7
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–8

This increment size must also be subjected to other bounds provided by rules discussed later. Similar
accuracy monitoring techniques can be devised for more complicated integration schemes such as
the midpoint rule.
Remark 17.9. Comparing (17.23) with (17.17) furnishes a simple rule: choosing  < 1 takes care of stability
if the roots of A are real. In any case the similarity between the stability and accuracy control rules is striking.

Remark 17.10. For poorly scaled problems the use of a weighted norm, as in (17.18), is advisable.

Remark 17.11. Given bounds on e at each step, bounds on the accumulated drift error can be obtained but
they are usually so pessimistic as to be useless, unless some detailed problem information is available; see
e.g. Gear (loc.cit. in footnote 2). The only reliable way to assess global accuracy is to rerun the problem with
several values of , for example  = 0.2, 0.1 and 0.05.

Remark 17.12. This truncation error monitoring technique works in static nonlinear analysis because e
is controlled by physically relevant low-frequency modes. It fails on direct time integration of dynamical
problems — see e.g. Park4 — because e is then controlled by physically irrelevant high-frequency modes. In
numerical analysis parlance, problems in structural dynamics are said to be stiff.

4 K. C. Park and C. A. Felippa, Direct Time Integration Methods in Nonlinear Structural Dynamics, Comp. Meth. Appl.
Mech. Engrg., 17/18, pp. 277–313 (1979)
K. C. Park, Time Integration of Structural Dynamics: A Survey, Ch. 4.2 in Pressure Vessels and Piping Design Technology
— A Decade of Progress ASME, New York (1982).

17–8
 16
.

Overview of
Solution Methods

16–1
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–2

TABLE OF CONTENTS

Page
§16.1. Introduction 16–3
§16.1.1. Stages, Increments and Iterations . . . . . . . . . . . 16–3
§16.1.2. Why Incrementation? . . . . . . . . . . . . . . 16–4
§16.2. Advancing the Solution: Increment Control 16–5
§16.3. Advancing the Solution: Prediction 16–5
§16.4. Advancing the Solution: Correction 16–7
§16.5. Traversing Equilibrium Path in Positive Sense 16–7
§16.5.1. Positive External Work . . . . . . . . . . . . . . 16–8
§16.5.2. Angle Criterion . . . . . . . . . . . . . . . . 16–8
§16.6. Constraint Strategy 16–9
§16.6.1. λ Control . . . . . . . . . . . . . . . . . . . 16–9
§16.6.2. State Control . . . . . . . . . . . . . . . . . 16–9
§16.6.3. Arclength Control . . . . . . . . . . . . . . . . 16–11
§16.6.4. (Global) Hyperelliptic Control . . . . . . . . . . . 16–11
§16.6.5. Local Hyperelliptic Control . . . . . . . . . . . . . 16–12
§16.7. Practical Solution Requirements 16–12
§16.7.1. Tracing the Response . . . . . . . . . . . . . . 16–12
§16.7.2. Finding a Nonlinear Solution . . . . . . . . . . . . 16–13
§16.7.3. Stability Assessment . . . . . . . . . . . . . . 16–13
§16.7.4. Post-buckling and Snap-through . . . . . . . . . . . 16–13
§16.7.5. Multiple Load Parameters . . . . . . . . . . . . . 16–14
§16. Exercises . . . . . . . . . . . . . . . . . . . . . . 16–15

16–2
16–3 §16.1 INTRODUCTION

In previous Chapters we have covered the governing equations of geometrically nonlinear structural
analysis and the discretization of those equations by finite element methods. The result is a set of
parametrized nonlinear algebraic equations called residual force equations.
The solution of these equations as the control parameters are varied varied provides the equilibrium
response of the structure. In this Chapter we begin the coverage of solution methods suitable for
digital computation.
§16.1. Introduction
It was noted in Chapter 1 that all solution procedures of practical importance are strongly rooted in
the idea of “advancing the solution” by continuation. The basic idea is to follow the equilibrium
response of the structure as the control and state parameters vary by small amounts. The motivation
in terms of circumventing the “solution morass” is described in that Chapter.
This overarching framework gives rise to many variants called solution schemes. A common feature
is that continuation is a multilevel process, as illustrated in Figure 16.1. The process involves
a hierarchical breakdown into stages, incremental steps, and iterative steps. The middle level:
incrementation, is always present. Staging may be missing if there is only one control parameter.
Iteration may be missing if there if no correction process.
In the present Chapter multilevel continuation is described in general terms, with the goal of
maintaining independence from specific solution schemes. The final subsections describe how the
general procedure is adapted to the analysis of problems encountered in engineering practice.

§16.1.1. Stages, Increments and Iterations

As discussed in Chapter 4, processing a complex nonlinear problem generally involves performing
a series of analysis stages. Multiple control parameters are not varied independently in each
stage and may therefore be characterized by a single stage control parameter λ. Stages are only
weakly coupled in the sense that the end solution of one may provide the starting point for another.
Throughout this and following Chapters attention is focused on a generic stage and there is no need
to use an identifying index for it.
To advance the solution, the stage is broken down into incremental steps, or increments for short. If
necessary incremental steps will be identified by the subscript n; for example, the state vector after
the n th increment is un and the state vector before any increment (at stage start) is u0 . Over each
incremental step the state vector u and stage control parameter λ undergo finite changes denoted
by u and λ, respectively.
Incremental solution methods can be divided into two broad classes:
1. Purely incremental methods, also called predictor-only methods.
2. Corrective methods, also called predictor-corrector or incremental-iterative methods.
In purely incremental methods the iteration level is missing. In corrective methods a predictor step
is followed by one or more iteration steps. The set of iterations is called the corrective phase. Its
purpose is to eliminate or reduce the so called drifting error, discussed in §16.4, which plagues
purely incremental methods.
Iteration steps will be usually identified by the superscript k; for example {ukn , λkn } may denote the
solution after the k th iteration of the n th step, whereas {u0n , λ0n } is the predicted solution before starting

16–3
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–4

Stages
Increments
Iterations

Figure 16.1. Nested hierarchy in nonlinear solution methods:

stages, increments and iterations.

the corrective process. This superscript is enclosed in parentheses if there is potential confusion
with exponents. Iterative changes in u and λ are often shortened to d and η, respectively.

Remark 16.1. Solutions accepted after each increment following completion of the corrective process, are
often of interest to users because they represent approximations to equilibrium states. They are therefore saved
as they are computed. On the other hand, intermediate results of iterative processes are rarely of interest unless
one is studying the “insides” of solution processes. Hence most production programs discard them.

Remark 16.2. The terminology of nonlinear static analysis is far from standardized. Despite their practical
importance, few authors recognize the existence of stages. Many use the term step to mean incremental step
whereas the terms substep, subincrement and cycle are used for the iteration level. There is more uniformity
in dynamic analysis, possibly because there is only one advancing level: step is universally used to denote the
change over a time increment.

§16.1.2. Why Incrementation?

The use of increments may seem at first sight unnecessary if one is interested primarily in the final
solution. But breaking up a stage into increments may serve other purposes:
Helping convergence. Success in a corrective process done by a Newton-like method may hinge on
having a good initial guess supplied by the predictor, since such methods are notoriously finicky.
The quality of this guess can be improved by reducing the increment.
Sidestepping extraneous roots. Incrementation helps the solution procedure from falling into the
“root morass” discussed in Chapter 1.
Gaining insight into structural behavior. As noted in Remark 16.1, programs often save converged
solutions after each increment and for a good reason: a response plot can teach the engineer more
about the structural behavior than simply knowing the final solution.
Avoiding surprises. Critical points may occur before the stage ends. There are problems in which
such points, notably bifurcation, may be masked if coarse increments are taken.

16–4
16–5 §16.3 ADVANCING THE SOLUTION: PREDICTION

Alleviating path dependence. Although the focus of this course is on path-independent problems,
it should be noted that the presence of path-dependent effects severely restricts increment sizes
because of history-tracing constraints. For example, in plasticity analysis stress states must not be
allowed to stray too far outside the yield surface.

§16.2. Advancing the Solution: Increment Control

A nonlinear analysis program is “marching” along a stage. Assume that n incremental steps have
been completed. The last accepted solution is un , which corresponds to λn . Performing the (n+1)th
step entails the calculation of the increments

un = un+1 − un , λn = λn+1 − λn , (16.1)

that satisfy the residual equilibrium equations r(u, λ) = 0 to requested accuracy. As stated the task
is not fully defined because there are less equations than unknowns, which makes the increment
sizes indeterminate. The problem is closed by adopting an increment control strategy. The strategy
may be expressed in general form as a constraint condition:

c(un , λn ) = 0, (16.2)

which equalizes the number of equations to the number of unknowns.

A rate form of the constraint equation (16.2) is obtained by differentiating with respect to the pseudo
time t:

aT u̇ + g λ̇ = 0, (16.3)
where
∂c ∂c
aT = , g= . (16.4)
∂u ∂λ

Remark 16.3. The addition of the constraint equation serves two purposes: it makes the algebraic problem
determinate, and it can be used to control the increment size directly or indirectly to enhance robustness and
convergence.

Remark 16.4. Note that the constraint c = 0 is expressed in terms of the increments {un , λn } from the last
solution and not in terms of the total values. This localization condition is essential to maintaining invariance
with respect to the origin chosen for u and λ. Once a step is finished, the constraint is reset for the next one
by moving its origin.

Remark 16.5. Specific choices for (16.4) are discussed in §164 below but for some developments it is possible
to keep c arbitrary. Furthermore, it is also possible to specify the constraint directly in the rate form (16.3)
without an explicit integral. An example is Fried’s orthogonal trajectory accession method.1

1 I. Fried, Orthogonal Trajectory Accession to the Nonlinear Equilibrium Curve, Comp. Meth. Appl. Mech. Engrg., 47,
283–297, (1984).

16–5
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–6

§16.3. Advancing the Solution: Prediction

Having decided upon an increment control strategy, to start up the (n+1)th incremental step, an
initial approximation
u0n , λ0n , (16.5)
to the increments (16.1) is calculated by a prediction step. These values are called the predicted
increments and the formula used is called a predictor or extrapolator.
Most predictors are based on the first-order path equation derived in Chapter 4 and repeated here
for convenience:
ṙ = 0, or Ku̇ = q λ̇, (16.6)
Assuming K to be nonsingular, the forward Euler method furnishes the simplest predictor:

u0n = K−1
n qn λn = vn λn ,
0 0
(16.7)

in which v is the incremental velocity vector defined in Chapter 4. The process is completed by
selecting an increment control strategy through the constraint (16.2). Two examples follow.

Example 16.1. For the prescribed-load-value strategy in which λn is specified to be

n (positive or negative),
the constraint is
c(u n , λn ) = λn −
n = 0. (16.8)
Then the increments are directly given by (16.6), i.e.

u0n = vn
n , λn =
n . (16.9)

This formula obviously fails when Kn is singular, i.e. at critical points, because there vn becomes either infinite
(at limit points) or nonunique (at bifurcation points). This suggests that the solution process will break down
at those points.

Example 16.2. For the arclength strategy in which the absolute value of the distance (4.23) is specified to be

n > 0, the constraint is


1 T


c(u n , λn ) = |sn | −
n = v un + λn  −
n = 0, (16.10)
fn n

where f n = + 1 + vnT vn . Substitution into (16.6) yields

n f n
n
n vn
n
λ0n = =  =± , u0n = ± . (16.11)
±(vn vn +
T 1) ± vnT vn + 1 fn fn

In this case two signs for the increment are obtained. The proper one is obtained by applying one of the “path
advancing” criteria discussed below.
Note also that (16.10) does not fail at isolated limit points if one properly passes to the limit v/|v| → z, as per
Remark 4.2. This limit process yields

λ0n = 0, u0n = ±
n z (16.12)

The normalized v near the limit point serves as a good approximation for z. It should be noted, however, that
the formula fails at multiple limit points and at bifurcation points; thus the arclength strategy is no panacea.

16–6
16–7 §16.5 TRAVERSING EQUILIBRIUM PATH IN POSITIVE SENSE

Computed solutions
λ

Drift error

Actual Equilibrium Path

v
Figure 16.2. Drift error in purely incremental solution procedure.

Both of the foregoing examples above contain a specified length

n . For the first step,
0 is normally
chosen by the user. If the predictor is followed by a corrective process, in subsequent steps
n may
be roughly adjusted according to the “last iteration count” rule of Crisfield,2 which works well in
practice. If no corrective phase follows, the proper selection of
n is discussed later in the section
dealing with purely incremental methods.
§16.4. Advancing the Solution: Correction
If the predicted increments (16.5) are inserted in the residual equation r(u, λ) = 0, there will
generally be a departure from equilibrium:
 
r0n = r un + u0n , λn + λ0n = 0. (16.13)
This departure is called drift error. A corrective process is an iterative scheme that eliminates, or
at least reduces, the drift error by producing a sequence of values
ukn , λkn , (16.14)
that as k → ∞ hopefully tend to the increments (16.2) that satisfy equilibrium and meet increment
control specifications. Popular corrective methods are studied in subsequent Chapters.
As previously noted, there are purely incremental methods that omit the corrective phase. They are
covered in following Chapters. See Figure 16.2 for an illustration of the drift error phenomenon
that occurs when a corrector is not applied.
Remark 16.6. An even simpler predictor consists of setting u0n = 0, λ0n = 0. The corrective process then
starts from the previous solution. This overcautious approach is rarely used in practice.

2 M. A. Crisfield, An Incremental-Iterative Algorithm that Handles Snap-Through, Computer & Structures, 16, 55–62
(1981)
M. A. Crisfield, An Arc-Length Method Including Line Searches and Accelerations, Int. J. Num. Meth. Engrg., 19,
1269–1289 (1983).

16–7
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–8

§16.5. Traversing Equilibrium Path in Positive Sense

In Example 16.2 two signs were obtained for the predicted λ0n and u0n . This is typical of
constraints that are reversible about the last solution point; that is, reversing the signs of both u
and λ satisfies c = 0.3 In that case the resulting algebraic system usually provides two solutions:

±λ0n , ±u0n . (16.15)

Even in Example 16.1 there is an ambiguity because the specified

n may be positive or negative.
The sign ambiguity arises because, as explained in Chapter 4, the tangent at regular points of an
equilibrium path has two possible directions, which generally intersect the constraint hypersurface
in at least two points. Thus it becomes necessary to chose the direction corresponding to a positive
path traversal. Two rules for chosing the proper sign are described below.

§16.5.1. Positive External Work

The simplest rule requires that the external work expenditure over the predictor step be positive:

W = qT u0n = qT vn λn > 0. (16.16)

That is, λ should have the sign of qT v = qT K−1 q.

This condition works well when “the structure follows the load” and is particularly effective at limit
points. It fails if q and v are orthogonal:

qT v = 0, (16.17)

because then the condition (16.16) is vacuous. This happens in the following cases.
Bifurcation points. As a bifurcation point B is approached, v/|v| → z, achieving equality at B.
Since qT z = 0, it follows that (16.16) fails at B.
Incremental velocity reversal. If the structure becomes “infinitely stiff” at a point in the equilibrium
path v vanishes. This case is rarer than the previous one, but may arise in the vicinity of turning
points.
Bifurcation points demand special treatment and cannot be easily passed through simple predictor
methods. One way out is to insert artificial purtuebations that transformperturbations are inserted.
However, the case v → 0 can be overcome by a modification of the previous rule.

§16.5.2. Angle Criterion

There are problems in which the structure gains suddenly stiffness, as for example in the vicinity of
a turning point T . If the positive work criterion is used eventually the solution process “turns back”
and begins retracing the equilibrium path. When it reaches the high stiffness point again it does
another U-turn and so on. The net result of this “ping pong” effect is that the solution process gets
stuck. Physically a positive work rule is incorrect because the structure needs to release external
work to continue along the equilibrium path.

3 Some authors call such constraints symmetric.

16–8
16–9 §16.6 CONSTRAINT STRATEGY

To get over this difficulty a condition on the angle of the prediction vector is more effective. Let
tn−1 be the tangent at the previous solution. Then chose the positive sense so that

tnT tn−1 > 0. (16.18)

Once the “ping-pong” region is crossed, the work criterion should be reversed so the external work
is negative.

Remark 16.7. Other geometric criteria are given by Crisfield (loc. cit. in footnote 2) and Skeie and Felippa4

§16.6. Constraint Strategy

So far the form of the constraint equation, (16.2) or (16.3), has been left arbitrary. In the sequel we
list, roughly in order of ascending complexity, instances that are either important in the applications
or have historical interest. In what follows
is always a dimensionless scalar that characterizes the
size of the increment. Six constraints are pictured in Figure 16.3. In this figure, c is the constraint
curve, S is the last solution point, P the predicted point and C the converged solution.

§16.6.1. λ Control
At each step λn =
n , where
is a dimensionless scalar. The constraint equation is (16.8) listed
in Example 16.1. This is generally called λ-control. Often the parameter λ is associated with a
loading amplitude, in which case this is called load control. The physical analogy would be a test
machine in which the operator increases the load to specific values.
The differential form (16.3) has
a = 0, g = 1. (16.19)
As noted in Example 16.1, this constraint form fails as critical points are approached.

§16.6.2. State Control

This consists of specifying a norm of un , for example the Euclidean norm:

c(un ) ≡ (unT un )2 −

n u 2 = 0, (16.20)

where u is a reference value with dimensions of displacement, which is introduce for scaling
purposes. An alternative way of doing that consists of using the scaled increment of §4.6:


unT 
un −
2n = 0. (16.21)

(See also Remark below.) The differential form (16.3) has

aT = 2un , g = 0. (16.22)

4 G. Skeie and C. A. Felippa, A Local Hyperelliptic Constraint for Nonlinear Analysis, Proceedings of NUMETA’90
Conference, Swansea, Wales, Elsevier Sci. Pubs, 1990.

16–9
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–10

c=0
c=0 P P P
C C C
S S S
c=0

(a) (b) (c)

P P P
C C C
S S S
c=0

c=0
c=0

(d) (e) (f)

Figure 16.3. Geometric representation of constraint equations for a one-dof problem, with
state u and control parameter λ plotted horizontally and vertically, respectively.
(a) load control, (b) state control, (c) arclength control, (d) hyperspherical control,
(e) global hyperelliptical control, and (f) local hyperelliptic control.

Remark 16.8. In the finite element literature the term displacement control has been traditionally associated
with the case in which the magnitude of only one of the components of u, say u i , is specified, which is
tantamount to choosing a special infinity norm of u. This old technique was used in the mid-1960s by Argyris
and Felippa.5 There is a generalization of single displacement control in which several reference displacements
are used. This multiple dimensional hyperplane control has been investigated by Powell, Bergan and others.6

5 J. H. Argyris, Continua and Discontinua, in Proceedings Conference on Matrix Methods in Structural Engineering,
AFFDL-TR-66-80, Wright-Patterson AFB, Dayton, Ohio, 11–189 (1966).
C. A. Felippa, Refined Finite Element Analysis of Linear and Nonlinear Two-dimensional Structures, Ph.D. thesis, Dept.
of Civil Engrg, University of California, Berkeley (1966).
6 G. H. Powell and J. Simons, Improved Iteration Strategy for Nonlinear Structures, Int. J. Num. Meth. Engrg., 17, 1655–
1667 (1981)
P. G. Bergan, G. Horrigmoe, B. Krakeland and T. H. Søreide, Solution Techniques for Nonlinear Finite Element Problems,
Int. J. Num. Meth. Engrg., 12, 1677–1696 (1978)
P. G. Bergan, Solution Algorithms for Nonlinear Structural Problems, Computers & Structures, 12 497–509 (1980)
P. G. Bergan and J. Simons, Hyperplane Displacement Control Methods in Nonlinear Analysis, in Innovative Methods
for Nonlinear Problems, ed. by W. K. Liu, T. Belytschko and K. C. Park, Pineridge Press, Swansea, U.K., 345–364
(1984)

16–10
16–11 §16.6 CONSTRAINT STRATEGY

§16.6.3. Arclength Control

Arclength control consists of specifying a distance |s| =

along the path tangent. The constraint
equation is (16.10) in Example 16.2. This form has scaling problems since it intermixes u and λ.
It is generally preferable to work with the scaled quantities of §4.6 in which case the constraint
becomes
1  T 
sn −
n = vn un + λn  −
n = 0, (16.23)

fn
The differential form (16.3) for the unscaled form (16.9) is

a T = vn / f n , g = 1/ f n . (16.24)

and for (16.24)

aT = vn S2 / 
fn , g = 1/ 
fn . (16.25)
Without the scaling this becomes the constraint of Riks and Wempner,7 also called arclength control.
Geometrically the unscaled equation represents a hyperplane normal to t, located a distance
n
from the last solution point S(un , λn ) in the state-control space. The scaled form admits a similar
interpretation in the scaled state-control space space (Su, λ).

Remark 16.9. The “orthogonal trajectory” constraint discussed by Fried (see footnote 1) may be regarded as
a generalization of the arclength constraint in which a traversal orthogonality condition is applied throughout
the corrective phase. This differential constraint is interesting in that it does not fit the form (16.2) and may
in fact be followed independently of the the predictor and past solution. But following the trajectory depends
on v = K−1 q being frequently updated and is practical only with a true Newton corrector.

§16.6.4. (Global) Hyperelliptic Control

There is a wide family of constraints that combine the magnitude of λn and a norm of un . A
frequently used combination is the hyperelliptic constraint

an2 unT un + bn2 (λn )2 =

2n , (16.26)

where scalar coefficients a and b may not be simultaneously zero.

More effective in practice is the scaled form of the above, namely

an2 
unT 
un + bn2 (λn )2 =
2n , (16.27)

where all quantities are now dimensionless.

Geometrically these constraints corresponds to an hyperellipse that has the last solution as center,
and includes other constraints as degenerate cases. The scaling parameters a and b were introduced

7 E. Riks, The Application of Newton’s Method to the Problem of Elastic Stability, Trans. ASME, J. Appl. Mech., 39,
1060–1065 (1972)
G. A. Wempner, Discrete Approximations Related to Nonlinear Theories of Solids, Int. J. Solids Structures, 7, 1581–1599
(1971).

16–11
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–12

by Padovan and Park.8 The expression was rendered dimensionless by Felippa9 who introduced
scaling parameters and and discussed appropriate choices. If a = b = 1 in the unscaled form
(16.27) we recover the hyperspherical constraint proposed (but not used) by Crisfield (loc. cit. in
footnote 4).
The constraint gradients are
a = 2a 2 u, g = 2b2 λ. (16.28)

§16.6.5. Local Hyperelliptic Control

This is a variation of the previous one in which we take a combination of λ and a norm of u,
where λ and u are to be determined according to a local coordinate system at S(un , λn ):

c(u, λ) = a 2 (u − un )T S(u − un ) + b2 (λ − λn )2 −
2n = 0, (16.29)

where a and b are scalar coefficients and

n is prescribed. Geometrically this is a hyperellipse with
principal axes in a coordinate system defined by λ and u. An attractive choice for the local
system is provided by the path tangent vector tn and the normal hyperplane at point S(un , λn ).
These are given by by (4.16) and (4.20) respectively, with v ≡ vn .
Near critical points, v → ∞. In such a case we would like to recover the global system to avoid
numerical difficulties. This is achieved by defining the new variables λ and u according to

λ = vT (u − vλ), u = v(vT u + λ). (16.30)

Scaling of this constraint to achieve consistency is discussed by Skeie and Felippa (work cited in
footnote 4), where additional computational details may be found. It turns out that this constraint
can include all ones previously discussed as special regular or limit cases.
10
Remark 16.10. Another interesting strategy: the work constraint of Bathe and Dvorkin limits the total
external work spent during the corrective phase.

Remark 16.11. In path-independent problems that involve only geometric or conservative boundary-condition
nonlinearities, it is generally best to maximize step lengths subject to stability and equilibrium accuracy
constraints. Stability depends on the curvature of the response path, presence of critical points, and solution
method used. Equilibrium accuracy depends chiefly on whether a corrective process is applied.

8 J. Padovan and S. Tovichakchaikul, Self-Adaptive Predictor-Corrector Algorithm for Static Nonlinear Structural Analysis,
Computers & Structures, 15, 365–377 (1982).
K. C. Park, A Family of Solution Algorithms for Nonlinear Structural Analysis Based on the Relaxation Equations, Int.
J. Num. Meth. Engrg., 18, 1637–1647 (1982).
9 C. A. Felippa, Dynamic Relaxation under General Increment Control, in Innovative Methods for Nonlinear Problems,
ed. by W. K. Liu, T. Belytschko and K. C. Park, Pineridge Press, Swansea, U.K., 103–163 (1984).
10 K. J. Bathe and E. Dvorkin, On the Automatic Solution of Nonlinear Finite Element Equations, Computers & Structures,
17, 871–879 (1983).

16–12
16–13 §16.7 PRACTICAL SOLUTION REQUIREMENTS

§16.7. Practical Solution Requirements

The remaining subsections describe various types of nonlinear structural analyses encountered in
engineering practice, and the requirements they pose on solution procedures.

§16.7.1. Tracing the Response

“Tracing the response” is of interest for many nonlinear problems. For a typical stage, perform a
sequence of incremental steps to find equilibrium states

un , λn , n = 1, 2, . . .

in sufficient number to ascertain the response u = u(λ) of the structure within engineering require-
ments.
If the control parameter is associated with a fundamental load system, the response path is known
as the fundamental equilibrium path, as it pertains to the service range in which the structure is
supposed to operate.
One class of problems that fit this requirement is that in which structural deflections, rather than
strength, are of primary importance in the design. For example, some large flexible space structures
must meet rigorous “dimensional stability” tolerances while in service.

§16.7.2. Finding a Nonlinear Solution

A variant of the foregoing occurs if the primary objective of the analysis is to find a solution u
corresponding to a given λ (for example, λ = 1), whereas tracing of the response path is in itself
of little interest.
Very flexible structures that must operate in the nonlinear regime during service fit this problem
class. The example of the suspension bridge under its own weight, discussed in §3.4, provides
a good illustration. The undeflected “base” configuration u = 0 is of little interest as it has no
physical reality and the bridge never assumes it. It is merely a reference point for measuring
deflections.
Under such circumstances, the chief consideration is that the accuracy with which the response path
is traced is of little concern. Getting the final answer is the important thing. Once this reference
configuration is obtained, “excursions” due to live loads, temperature variations, wind effects and
the like may be the subject of further analysis staging.

§16.7.3. Stability Assessment

This is perhaps the most important application of nonlinear static analysis. The analyst is concerned
with the value (or values) of λ closest to 0 at which the structure behavior is not uniquely determined
by λ. These are the critical points discussed in Chapter 5. In physical terms, the system becomes
uncontrollable and may “take off” dynamically.
Problem of this nature arise in stability design. The determination of limit points is called collapse
or snapping analysis. The determination of bifurcation points is called buckling analysis.

16–13
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–14

§16.7.4. Post-buckling and Snap-through

Occassionally it is of interest to continue the nonlinear analysis beyond a limit or bifurcation
point. Continuation past a limit point is post-collapse or snap-through analysis; continuation past
a bifurcation point is post-buckling analysis.
Post-critical analyses are less commonly encountered in practice than the previous two types. They
are of interest to ascertain imperfection sensitivity of primary structural components, or to assess
strength reserve in fail-safe analysis under abnormal conditions such as construction, deployment
or accidents.
Conventional load control is not generally sufficient to trace snap-through. This may be achieved,
however, with the aid of the more general increment control strategies discussed above. Traversing
bifurcation points is notoriously more difficult; a technique applicable to well isolated bifurcation
points is discussed later in the context of augmented equations and auxiliary systems.

§16.7.5. Multiple Load Parameters

As discussed in Chapter 3, the case of multiple control parameters is reduced to a sequence of
one-parameter analyses. The previous classification apply to individual stages, and not all stages
necessarily fit the same type of analysis requirements.
The systematic determination of a complete equilibrium surface as the envelope of all response paths
is rarely pursued in practice aside from academic examples. For practical structures, an investigation
of this type would put enormous demands on human and computer time and is doubtful whether
the additional insight would justify such expenditures.
There is, however, a special case of multiparameter investigation that is gaining popularity for
designing lightweight structures: stability interaction curves as envelopes of critical points.

16–14
16–15 Exercises

Homework Exercises for Chapter 16

Overview of Solution Methods

EXERCISE 16.1
[C:20] Consider the following residual equilibrium equation:
  
r (ψ, λ) = sec(α−ψ) sec α (2+λ sin ψ) − 2 sec(α−ψ) tan(α−ψ) − λ cos ψ sec α = 0, (E16.1)

in which α is a problem parameter, λ the control parameter, and ψ the only degree of freedom. This r comes
from the 2-bar arch problem already studied in Exercise 6.2. Here α is the initial arch rise angle whereas
ψ = α − θ is the angle change from the reference state, at which ψ0 = 0 and λ0 = 0. A plot of the exact
λ(ψ) for ψ = [0, 60◦ ] = [0, π/3] is shown in Figure E16.1(a); the fundamental path ends at limit point L.

λ
(a) λ (b) 0.08 B
0.06 Exact response for Exact response for
L 0.06
Exercise 16.1 Exercise 16.2
0.04 0.04
0.02 ψ (rad)
ψ (rad) 0.02

0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

−0.02 −0.02

−0.04 −0.04
−0.06
−0.06

Figure E16.1. Exact responses for Exercises E16.1 and E16.2.

(a) Derive the first-order rate form K ψ̇ = q λ̇ by taking t ≡ λ, and convert to u̇ = du/dλ = v. (Recall that
K = ∂r/∂ψ, q = −∂r/∂λ, and v = K −1 q.)
(b) Integrate numerically the rate equation u̇ = v found in (a) by the purely incremental, forward-Euler
method with load control over λ = [0, 0.1]. Start from λ0 = 0 and ψ0 = 0. (All angles should be in
radians.) Use α = 30◦ = π/6 as arch rise angle and take 10 load increments of
n = 0.01 (same for all
steps). Are you able to detect and traverse the limit point L?
(c) Repeat the run twice, each time cutting
n by 1/4 and quadrupling the number of steps. Is limit point
detection and traversal improved?
Hints. If using Mathematica the 10-step forward Euler script followed by the response plot could be imple-
mented as

v=q/K; Eulersol={{0,0}}; psin=0; lambdan=0; ns=10; ln=0.01;

For [i=1,i<=ns,i++,
vn=N[v/.{lambda->lambdan,psi->psin}];
lambdanp1=lambdan+ln; psinp1=psin+vn*ln;
Eulersol=AppendTo[Eulersol,N[{psinp1,lambdanp1}]];
lambdan=lambdanp1; psin=psinp1];
ListPlot[Eulersol,PlotJoined->True];
EXERCISE 16.2
[C:20] Repeat (a)–(c) of the foregoing Exercise for the residual
 
r = r̂ 1/4 − 16λ + cos(4ψ) . (E16.2)

in which r̂ is the residual (E16.1). The fundamental path now ends at a bifurcation point B, as pictured in
Figure E16.1(b). Use α = 30◦ and same solution method. Are you able to detect the bifurcation point?

16–15
 15
.

The CR Description:
C1 Plane Beam

15–1
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–2

TABLE OF CONTENTS

Page
§15.1. Introduction 15–3
§15.2. CR Beam Kinematics 15–3
§15.2.1. Coordinate Systems . . . . . . . . . . . . . . . 15–3
§15.2.2. Degrees of Freedom . . . . . . . . . . . . . . . 15–3
§15.2.3. Partial Derivatives . . . . . . . . . . . . . . . . 15–5
§15.2.4. Arbitrary Initial Configuration . . . . . . . . . . . 15–7
§15.2.5. Stress Resultants . . . . . . . . . . . . . . . . 15–8
§15.3. The Deformational Strain Energy 15–8
§15.4. Internal Force Vector and Tangent Stiffness Matrix 15–9
§15.4.1. Internal Force Vector . . . . . . . . . . . . . . 15–9
§15.4.2. Material Stiffness Matrix . . . . . . . . . . . . . 15–9
§15.4.3. Geometric Stiffness Matrix . . . . . . . . . . . . 15–10
§15. Exercises . . . . . . . . . . . . . . . . . . . . . . 15–11

15–2
15–3 §15.2 CR BEAM KINEMATICS

§15.1. Introduction

In this Chapter we use the CR description to construct a geometrically nonlinear, 2-node Bernoulli-
Euler plane beam. Unlike Chapter 9 we will do a C 1 (Hermitian) beam from the start, since with CR
it is as easy to do C 1 or C 0 , and the former has a much better geometric stiffness matrix.

§15.2. CR Beam Kinematics

The CR formulation of the beam motion is quite similar to that of the bar element in many respects,
and much of the development can be reused. Only the major differences will be noted here.

§15.2.1. Coordinate Systems

As in Chapter 9, we consider a plane, straight, prismatic beam element with two nodes. The element
is initially aligned with the global X axis in the initial configuration C0 , with the origin O0 located
at the element midpoint. This configuration is assumed to be straight and undeformed although it
may be under initial uniform axial stress with resultant N 0 . The bar properties include the elastic
modules, E, the cross section area A0 and the moment of inertia I0 about the neutral axis. The length
in C0 is L 0 .
The motion on the {X, Y } plane carries it to the current configuration C. The corotated configuration
C R is selected as depicted in Figure 15.1:
1. The longitudinal axis passes through the current position of the end nodes. This defines the
local axis x e . The origin of {x e , y e } is placed halfway between the nodes. This forms an angle
ψ with X .
2. The C R nodes are placed at an equal distance from the C nodes. Hence the corotated axes
{x Re , y Re }, including origin, coincide with {x e , y e }.
The new ingredient is the rotation angle θ about Z or z. With C R chosen as indicated, the deformation
part of these rotations is easily extracted: θ̄ = θ − ψ.
Other possibilities for selecting C R are possible. The foregoing choice has the advantage of being
compatible with that of the bar element discussed in the previous Chapter.

§15.2.2. Degrees of Freedom

The beam element has six degrees of freedom, which are placed in the vectors
     1 
u X1 ū eX 1 −2d
 uY 1   u eY 1   0 
     
 θ   θ̄ e   θ̄ 
 1   1   1 
u= , ūe =  e  =  1  . (15.1)
 u X2   ū X 2   2 d 
   e   
 uY 2   ū Y 2   0 
θ2 θ̄2e θ̄2

See Figure 15.2 for a picture of the global displacements and Figure 15.3 for the deformation dis-
placements.

15–3
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–4

_
// Y
φ = ψ+ϕ_
θ2 ψ
// X
ϕ
2(x2 ,y2) // X
Current C
_
// Y
uY2
θ1
Corotated CR_
1(x1 ,y1) X
Base or initial configuration ϕ
C0 uY1
2(X2, Y2 )
Y, y _ uX2
Yγ
X, x
uX1
1(X1, Y1)

Figure 15.1. Kinematics of corotational C 1 plane beam element.

θ2
2 //X, x

C
θ1
//X, x u Y2
u X1
1

uY1

C0 uX2
10 20

Figure 15.2. Global element displacements upon aligning X and X̄ .

ye
_
−d/2 θ1 d/2 _
CR θ2
1 xe
2

L0 C
L

Figure 15.3. Deformational displacements in element system.

15–4
15–5 §15.2 CR BEAM KINEMATICS

Proceeding as in the general formulation specialized to the 2D case, we can obtain the following
relation:
 ū e      1 L (1 − c ) 
x1 cψ sψ 0 0 0 0 u X1 − u X0 2 0 ψ
 ū y1   −sψ cψ 0
e
0 0    
  uY 1 − uY 0  
1
   0 L s
2 0 ψ 
 θ̄ e   0    
 1   0 1 0 0 0  θ   −ψ 
ū =  e  = 
e

1
+1  (15.2)
 ū x2   0 0 0 cψ sψ 0   u X 2 − u X 0   2 L 0 (cψ − 1) 

 e      
 ū y2   0 0 0 −sψ cψ 0   uY 2 − uY 0   − 2 L 0 sψ 
1

θ̄ e 0 0 0 0 0 1 θ2 −ψ
2

Here cψ and sψ and the angle ψ are implicitly defined by the displacements through the trigonometric
relations
Ly Lx Ly
sψ = sin ψ = , cψ = cos ψ = , ψ = arctan (15.3)
L L Lx
where L x = L 0 + u X 2 − u X 1 , L y = u Y 2 − u Y 1 , and

L = L 2x + L 2y (15.4)

is the bar length in the current configuration, ignoring the bending deformation.
We note the following relations
∂L ∂L ∂L ∂L ∂L ∂L
=− = cψ , =− = sψ , = = 0,
∂u X 2 ∂u X 1 ∂u Y 2 ∂u Y 1 ∂θ1 ∂θ2
∂cψ ∂cψ sψ2 ∂cψ ∂cψ sψ cψ ∂cψ ∂cψ
=− = , =− =− , = = 0,
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
(15.5)
∂sψ ∂sψ sψ cψ ∂sψ ∂sψ cψ2 ∂sψ ∂sψ
=− =− , =− = , = = 0,
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
∂ψ ∂ψ sψ ∂ψ ∂ψ cψ ∂ψ ∂ψ
=− = , =− =− , = = 0.
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
which are useful in the calculations that follow.

§15.2.3. Partial Derivatives

The first and second partial derivatives of the deformations d, θ̄1 and θ̄2 with respect to the node
displacements are necessary for the computations of internal forces and stiffness matrices.
Using (15.5) and Mathematica, one obtains for the first derivatives:

 δ ū e    
x1
1
c
2 ψ
1
s
2 ψ
0 − 12 cψ − 12 sψ 0 δu X 1
 δ ū y1
e   −sψ cψ L 0 /L cψ2 L 0 /L 0 sψ cψ L 0 /L −cψ2 L 0 /L 0  
     δu Y 1 
 δ θ̄ e   −s /L cψ /L sψ /L −cψ /L 0   δθ1 
 
 1   ψ 1 
 e =   (15.6)
 δ ū x2   − 12 cψ − 12 sψ 0 1
c
2 ψ
1
s
2 ψ
0   δu 
 e     X2 
 δ ū y2   sψ cψ L 0 /L −cψ2 L 0 /L 0 −sψ cψ L 0 /L cψ L 0 /L
2
0   δu Y 2 
δ θ̄2e −sψ /L cψ /L 0 sψ /L −cψ /L 1 δθ2

15–5
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–6

_
// Y
φ = ψ+ϕ _
ψ
θ2 // X
ϕ
2(x2 ,y2 ) // X
_
// Y
C
uY 2
θ1

1(x1 ,y1) _
X
ϕ
2(X2, Y2 )
uY 1
Y, y _ uX 2
Y C0
X, x
uX1
1(X1, Y1 )

Figure 15.4. Beam element with arbitrarily oriented initial configuration C0 ,

forming an angle ϕ with X . Corotated configuration not shown
to reduce clutter.

Since −u ex1 = u ex2 = 12 d, θ̄1e = θ̄1 , and θ̄2e = θ̄2 we get

 
δu X 1
     δu Y 1 
δd −cψ −sψ 0 cψ sψ 0  δθ 

 δ θ̄1  =  −sψ /L  1 
cψ /L 1 sψ /L −cψ /L 0  (15.7)
 δu X 2 
δ θ̄2 −sψ /L cψ /L 0 sψ /L −cψ /L 1  
 δu Y 2 
δθ2
The second derivatives of deformation variables are
 
sψ2 −sψ cψ 0 −sψ2 sψ cψ 0
 −sψ cψ cψ2 0 sψ cψ −cψ2 0
 
∂ 2d 1 0 0 0 0 0 0
=  
 (15.8)
∂u ∂u L −s 2
ψ sψ cψ 0 sψ2 −sψ cψ 0
 s c −cψ2 0 −sψ cψ cψ2 0
ψ ψ
0 0 0 0 0 0

 
−2sψ cψ cψ2 − sψ2 0 2sψ cψ sψ2 − cψ2 0
 c2 − s 2 2sψ cψ 0 sψ2 − cψ2 −2sψ cψ 0
 ψ ψ 
∂ 2 θ̄1 1 
 0 0 0 0 0 0
= 2 (15.9)
∂u ∂u L  2sψ cψ sψ − cψ2
2
0 −2sψ cψ cψ − sψ2
2
0
 s 2 − c2 −2sψ cψ 0 cψ2 − sψ2 2sψ cψ 0
ψ ψ
0 0 0 0 0 0

15–6
15–7 §15.2 CR BEAM KINEMATICS
 
−2sψ cψ cψ2 − sψ2 0 2sψ cψ sψ2 − cψ2 0
 c2 − s 2 2sψ cψ 0 sψ2 − cψ2 −2sψ cψ 0
 ψ ψ 
∂ 2 θ̄2 1  0 0 0 0 0 0
= 2  (15.10)
∂u ∂u L  2sψ cψ sψ − cψ2
2
0 −2sψ cψ cψ − sψ2
2
0
 s 2 − c2 −2sψ cψ 0 cψ2 − sψ2 2sψ cψ 0
ψ ψ
0 0 0 0 0 0

§15.2.4. Arbitrary Initial Configuration

The foregoing relations can be generalized to the case of a initial configuration C0 not aligned
with the X axis as shown in Figure 15.4. Given the node coordinates and displacements shown
in the figure, it is easily shown (Section §9.4) that cos ϕ = X 21 /L 0 , sin ϕ = Y21 /L 0 , cos φ =
cos(ψ + ϕ) = x21 /L, sin φ = sin(ψ + ϕ) = y21 /L, cos ψ = (X 21 x21 + Y21 y21 )/(L L 0 ) and
sin ψ = (X 21 y21 − Y21 x21 )/(L L 0 ).
The preceding transformation rules remain correct if ψ is replaced by φ = ϕ + ψ, except for the
deformation angle computation, which remain θ̄1 = θ1 − ψ and θ̄2 = θ2 − ψ because the θ s are
measured from X̄ .
The relation between deformational and global displacements become
d = L − L 0 = u X 21 cφ + u Y 21 sφ + L 0 (1 − cφ )
θ̄1 = θ1 − ψ (15.11)
θ̄2 = θ2 − ψ
The first derivatives of d=eformation variables are
 
δu X 1
     δu Y 1 
δd −cφ −sφ 0 cφ sφ 0  δθ 

 δ θ̄1  =  −sφ /L  
0
1
cφ /L 1 sφ /L −cφ /L  (15.12)
 δu X 2 
δ θ̄2 −sφ /L cφ /L 0 sφ /L −cφ /L 1  
 δu Y 2 
δθ2
The second derivatives of deformation variables are
 2 
sφ −sφ cφ 0 −sφ2 sφ cφ 0
 −sφ cφ cφ2 0 sφ cφ −cφ2 0
 
∂ d
2
1 0 0 0 0 0 0
=   (15.13)
∂u ∂u L −s 2
φ sφ cφ 0 sφ2 −sφ cφ 0
 s c −cφ2 0 −sφ cφ cφ2 0
φ φ
0 0 0 0 0 0
 
−2sφ cφ cφ2 − sφ2 0 2sφ cφ sφ2 − cφ2 0
 c2 − s 2 2sφ cφ 0 sφ2 − cφ2 −2sφ cφ 0
 φ φ 
∂ θ̄1
2
1  0 0 0 0 0 0
= 2 (15.14)
∂u ∂u L  2sφ cφ sφ − cφ2
2
0 −2sφ cφ cφ − sφ2
2
0
 s 2 − c2 −2s c 0 cφ2 − sφ2 2sφ cφ 0
φ φ φ φ
0 0 0 0 0 0

15–7
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–8

N
M2
V

M1
V
M20
N N0
V0
C0

V0
N0
M10

Figure 15.5. Beam stress resultants depicting positive sign conventions. Axial forces N
and transverse shear forces V are constant along the length, but the bending
moments M vary linearly. Hence two nodal values of M are required.
 
−2sφ cφ cφ2 − sφ2 0 2sφ cφ sφ2 − cφ2 0
 c2 − s 2 2sφ cφ 0 sφ2 − cφ2 −2sφ cφ 0
 φ φ 
∂ 2 θ̄2 1  0 0 0 0 0 0
= 2  (15.15)
∂u ∂u L  2sφ cφ sφ − cφ2
2
0 −2sφ cφ cφ − sφ2
2
0
 s 2 − c2 −2sφ cφ 0 cφ2 − sφ2 2sφ cφ 0
φ φ
0 0 0 0 0 0

§15.2.5. Stress Resultants

The stress resultants in the reference configuration (either C0 or C R ) are N 0 , M10 and M20 . The initial
shear force is V 0 = (M10 − M20 )/L 0 . See Figure 15.5 for sign conventions.
Denote by N , V and M the stress resultants in the current configuration. Whereas N and V are
constant along the element, M = M(x e ) varies linearly along the length because this is a Hermitian
or model, which relies on cubic transverse displacements. Consequently we will define its variation
by the two node values M1 and M2 . The shear V is recovered from equilibrium as V = (M1 − M2 )/L,
which is also constant. The stress resultants can be obtained from the deformations as
E A0 2E I0
N = N0 + d, M1 = M10 − (2θ̄1 + θ̄2 ),
L0 L0
(15.16)
2E I0 M1 − M2 L0 2E I
M2 = M20 + (θ̄1 + 2θ̄2 ), V = = V0 + (θ̄1 − θ̄2 ).
L0 L L L L0

§15.3. The Deformational Strain Energy

The next step in the CR formulation is to work out the deformational strain energy of the beam. The
basic choices are:

15–8
15–9 §15.4 INTERNAL FORCE VECTOR AND TANGENT STIFFNESS MATRIX

1. A linear beam
2. A nonlinear TL beam
The strain energy of the beam for small strains can be written

U = Ua + Ub + Ug (15.17)

where U a , U b and U g are the energy taken by axial (bar) deformation, bending deformation, and
initial-stress geometric effects, respectively. We adopt the following energy expressions:

U a = N 0 d + 12 (N − N 0 )d 2 = N 0 L 0 e + 12 E A0 L 0 e2 ,

T

1 θ̄1 E I0 4 2 θ̄1
U = M2 θ̄2 − M1 θ̄1 + 2
b 0 0
,
θ̄2 L0 2 4 θ̄ 2 (15.18)

T 0

1 θ̄1 N L 0 4 −1 θ̄1
U =2
g
.
θ̄2 30 −1 4 θ̄2

The 2 × 2 matrices appearing in U b and U g may be derived from those given in Chapters 5 and 15,
respectively, of Przemieniecki’s book.1 This book, howevr, omits the initial stress terms.
§15.4. Internal Force Vector and Tangent Stiffness Matrix

The internal force vector and tangent stiffness matrix of the corrotational element are then obtained
by the usual formulas:
∂U ∂p
p= , K= = K M + KG (15.19)
∂u ∂u
To develop these quantities it is necessary to find the first and second partial derivatives of d, θ̄1 and
θ̄2 in terms of the node displacements.

§15.4.1. Internal Force Vector

Using the partial derivatives compiled above and Mathematica, one obtains the following expression
for the internal forces.
p = pa + pb + pg (15.20)
where
∂U a
pa = = N [ −cφ −sφ 0 cφ sφ 0 ]T
∂u
∂U b
pb = = [ V sφ −V cφ −M1 −V sφ V cφ M2 ]T
∂u (15.21)
∂U g N 0 L0
pg = = [ −3sφ (θ̄1 + θ̄2 )/L 3cφ (θ̄1 + θ̄2 )/L 4θ̄1 − θ̄2
∂u 30
3sφ (θ̄1 + θ̄2 )/L −3cφ (θ̄1 + θ̄2 )/L 4θ̄2 − θ̄1 ]T

1 J. S. Przemieniecki, Theory of Matrix Structural Analysis, Dover, New York, 1985.

15–9
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–10

§15.4.2. Material Stiffness Matrix

Carrying out the computations one obtains the following compact expression for the material stiffness:
K M = TT K M0 T (15.22)
where  EA 
L 0 0 − ELA 0 0
 12E I 6E I 6E I 
 0 0 − 12E3 I 
 L 3
L 2
L L2 
 
 0 6E I 4E I − 6E I 2E I 
 L2 L 0
L2 L 
K M0=
 EA

 (15.23)
− 0 0 E A 0 0 
 L L 
 6E I 
 0 − 12E3 I − 6E2I 0 12E I − 
 L L L3 L2 
0 6E I 2E I 0 − 6E2I 4E I
L 2 L L L
is the stiffness matrix of the linear beam element, and T is the transformation matrix
 
cφ sφ 0 0 0 0
 −sφ cφ 0 0 0 0
 
 0 0 1 0 0 0
T=  (15.24)
 0 0 0 cφ sφ 0 
 
0 0 0 −sφ cφ 0
0 0 0 0 0 1
which introduces the effect of finite rigid body motions.
§15.4.3. Geometric Stiffness Matrix
The expression for the geometric stiffness is a bit more complicated. It can be presented in a compact
form as follows:
KG = TT KGN T + KGV (15.25)
N
where T is the transformation matrix (15.24), KG is the well known geometric stiffness for a Hermitian
beam element under axial force:
 
0 0 0 0 0 0
 0 36 3L 0 −36 3L 
 
N  0 3L 4L 2
0 −3L −L 2 
KGN =   (15.26)
30L  0 0 0 0 0 0 
 
0 −36 −3L 0 36 −3L
0 3L −L 2 0 −3L 4L 2
and the remaining term introduces the effect of varying moments through the transverse shear force
in C:  
sin 2φ − cos 2φ 0 − sin 2φ cos 2φ 0
 − cos 2φ − sin 2φ 0 cos 2φ sin 2φ 0
V  
0 0 0 0 0 0
KG = 
V
 (15.27)
L  − sin 2φ cos 2φ 0 sin 2φ − cos 2φ 0 
 
cos 2φ sin 2φ 0 − cos 2φ − sin 2φ 0
0 0 0 0 0 0
in which sin 2φ = 2sφ cφ and cos 2φ = cφ − sφ .
2 2

15–10
15–11 Exercises

Homework Exercises for Chapter 15

The Corotational Description: 2D C1 Beam

EXERCISE 15.1 Complete the derivation of p for the 2-node C 1 beam element and implement in Mathematica,
using the same inputs as in Chapter 9 Addendum. (Implemented and posted on Web)

EXERCISE 15.2 Complete the derivation of K for the 2-node C 1 beam element and implement in Mathematica,
using the same inputs as in Chapter 9 Addendum. (Implemented and posted on Web)

EXERCISE 15.3 A plane 2-node C 1 beam element has properties L 0 = 6, E = 3000, A0 = 2, I0 = 12, N 0 = 5
◦
in the initial state C0 along X , with node 1 at (0,0) and node 2 at (L 0 , 0). The beam rotates by 45
√ about the origin
√
so that at the current configuration C node 1 stays at {0, 0} while node 2 moves to {(L 0 + d)/ 2, (L 0 + d)/ 2},
where d = L 0 /1000. The rotational freedoms at C are θ1 = θ2 = 45◦ = π/4 radians. Compute p, K M and
KG at the current configuration, and compare those quantities with those of the C 0 beam element presented in
Chapter 9, using RBF for the latter.
Note: A Mathematica implementation of this C 1 element has been posted on the Web as a Mathematica 4.1
Notebook PlaneBeamC1.nb. The element checks out when moving about the reference configuration C0 . It
gives excellent buckling values for the problem of Exercise 9.3. More tests are needed, however, for an arbitrary
configuration to make sure the internal force vector and the tangent stiffness are consistent.

EXERCISE 15.4 Confirm the previous statement by repeating the buckling calculations of Exercise 9.3 using the
CR beam element provided in the Mathematica Notebook mentioned above (extract the material and stiffness
matrices, ignore the rest). Compare the speed of convergence of the CR and TL element for the cantilever
buckling problem.

15–11