Finite Element Technology

SS 2021

Prof. Dr.-Ing. Stefanie Reese

Robert Eggersmann, M.Sc.

1
2
Contents

0 Motivation 7

1 Basics of the low-order displacement approach 9

1.1 Basic equations (strong form) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.1.1 One-dimensional continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.1.2 Two-dimensional continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2 Basic equations (weak form) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.3 Interpolation of displacement and geometry . . . . . . . . . . . . . . . . . . . . . . . . 17

1.4 Split of N into linear and hourglass parts . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.5 Derivation of the strain-displacement operator . . . . . . . . . . . . . . . . . . . . . . . 22

1.6 Jacobi matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.6.1 General remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.6.2 Special cases of the Jacobi matrix . . . . . . . . . . . . . . . . . . . . . . . . . 25

1.6.3 Evaluation of the Jacobi matrix in the centre of the element . . . . . . . . . . . . 27

3
 1.6.4 Equivalent parallelogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

1.7 Finite element matrices and vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.7.1 Evaluation of K e by means of standard numerical integration . . . . . . . . . . 30

1.7.2 Evaluation of K e by means of analytical integration . . . . . . . . . . . . . . . 31

1.8 Assembly procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

1.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2 Locking 35

2.1 Typical examples for locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.1.1 Example for volumetric locking . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.1.2 Example for shear locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.2 Eigenvalue analysis for element stiffness matrix K e . . . . . . . . . . . . . . . . . . . . 38

2.2.1 Representation of K 0 and K hg . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.2.2 Intelligent formulation of the eigenvalue problem for K e . . . . . . . . . . . . . 41

2.2.3 Solution of the eigenvalue problem for K e . . . . . . . . . . . . . . . . . . . . 43

2.3 Comparison with analytical analysis at continuum mechanical level . . . . . . . . . . . 49

2.3.1 Homogeneous normal strain state . . . . . . . . . . . . . . . . . . . . . . . . . 50

2.3.2 Homogeneous shear strain state . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.3.3 Strain state for pure bending . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4
3 Methods against locking 55

3.1 General remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.2 Mixed variational formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.2.1 Repetition of the strong form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.2.2 Construction of alternative weak form – three-field variational functional . . . . 58

3.2.3 Concept for a possible discretization – reduction to two-field variational functional 58

3.3 Interpolation of the enhanced strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3.4 Element stiffness matrix K e in the context of Q1/E4 and Q1SP . . . . . . . . . . . . . . 62

3.4.1 General format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.4.2 Representation of K e by means of the reduced B matrix . . . . . . . . . . . . . 63

3.4.3 Evaluation of K e by means of standard numerical integration . . . . . . . . . . 64

3.4.4 Evaluation of K e by means of analytical integration . . . . . . . . . . . . . . . 64

3.5 Functional dependence of B red – is locking eliminated? . . . . . . . . . . . . . . . . . . 65

3.5.1 Limit of incompressibility – volumetric locking? . . . . . . . . . . . . . . . . . 66

3.5.2 Pure bending – shear locking? . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.6 Eigenvalue analysis at element level and comparison with analytical result . . . . . . . . 69

3.7 Pressure-displacement approach or selective reduced integration . . . . . . . . . . . . . 71

3.7.1 Repetition of the strong form - construction of the three-field variational functional 71

3.7.2 Concept for a possible discretization . . . . . . . . . . . . . . . . . . . . . . . . 74

5
 3.7.3 Element stiffness matrix K e in the context of Q1/P0 and Q1RV . . . . . . . . . 76

3.7.4 Functional dependence of (B red )4x8 – is locking eliminated? . . . . . . . . . . . 79

3.7.5 Eigenvalue analysis at element level and comparison with analytical result . . . . 80

4 Glossary 86

4.1 Notions in the text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

4.2 Mathematical symbols (only abbreviations) . . . . . . . . . . . . . . . . . . . . . . . . 88

5 Literature for further reading 89

6
Chapter 0

Motivation

The present text is written for readers who are already familiar with the basic finite element approach.
The notion finite element technology, however, is maybe less well-known and should be shortly ex-
plained.

In general, if one plans to investigate a mechanical system by means of the finite element method,
very many different finite element formulations could be used. A standard approach is the class of
isoparametric1 finite elements, where still the choice between low-order and high-order elements has to
be made. High-order elements are characterized by interpolation functions of high order. This leads to
the fact that more complex distributions of the physical quantities of interest, usually the displacement,
can be taken into account. Consequently, less elements are needed. On the contrary, low-order elements
include only very simple interpolation functions and require in general the use of more elements.

In the context of low-order elements, the focus is here on isoparametric quadrilaterals or bricks, respec-
tively, which are based on bilinear or trilinear shape functions. With respect to high-order elements, such
quadrilaterals or bricks are easier to mesh, more robust in large deformation situations and characterized
by a smaller bandwidth of the global stiffness matrix K. The consequence of the latter aspect is that
the equation system K U = P , where U is the global nodal displacement vector and P the global load
vector can be solved more efficiently.2
1
The class of isoparametric elements is characterized by using the same interpolation for the geometry and the field
variable of interest, here the displacement vector. This is indicated by the prefix “iso”.
2
Such a linear equation system is obtained for the case of linear elasticity. For non-linear problems, the equation to be
solved at global finite element level is a more complex function of U . Nevertheless, for the so-called tangential stiffness
matrix which shows up in the non-linear case, above statement about the bandwith holds in the same way.

7
For these reasons, two- or three-dimensional low order finite element formulations are usually preferred
to the corresponding higher order elements. Unfortunately, it has been found out that standard isopara-
metric elements of low order show the undesirable effect of locking. Locking is an artificial stiffening
which is caused by the simplicity of the shape functions. It leads, especially in bending-dominated situ-
ations and in the limit of incompressibility, to an overestimation of the stresses and an underestimation
of the displacements. The research field of finite element technology concentrates in the context of low-
order formulations on the elimination of the so-called locking effect. Typical strategies to achieve this
goal are reduced integration with hourglass stabilization and mixed methods.3 In the context of linear
elasticity, these approaches are now well established and offered in many commercial codes. The situa-
tion is not yet so comfortable in cases where large deformations or highly non-linear material behaviour
have to be considered. Thus, for more demanding applications, finite element technology is still a very
active field of international research.

In the present lecture, we restrict ourselves to the simplest case: linear elasticity.

Important notions and abbreviations of the following text are gathered in a glossary, see Chapter 4. The
glossary points to the number of the section or subsection, where the notion or abbreviation is introduced
and explained. Derivations in the field of finite element technology usually require many abbreviations.
However, mostly, these symbols refer to certain matrices or expressions within a derivation and do not
represent a physically important new quantity. Therefore, we abstain from indicating such definitions by
means of the otherwise common symbols =: or :=.

The text is often interrupted by remarks, additional mathematical derivations and summaries. The end
of these insertions will be indicated by a , if there is no new section directly placed behind it.

3
These notions will be explained in Chapter 3.

8
Chapter 1

Basics of the low-order displacement approach

1.1 Basic equations (strong form)

1.1.1 One-dimensional continuum

To develop a finite element formulation we first need to state some fundamental equations. We start
here with one-dimensional (1D) linear elasticity, i.e. a truss which behaves linearly elastically. Such a
system (see e.g. Fig. 1.1) has already been discussed in Technische Mechanik 2 (Strength of Materials).
The goal is to find the displacement function u (x). Having determined that we can compute the strain
ε = du/dx by means of the derivative of u with respect to the coordinate x (kinematical relation) and the
stress σ = E ε using Hooke’s law. However, there are three unknowns (u, ε, σ) and only two equations
(ε = du/dx, σ = E ε). We are missing the statement of equilibrium which can be easily derived at an
infinitesimal truss element.

Figure 1.1: Example for a truss system with length L, line load n = n (x) and extensional stiffness EA.

The equilibrium of forces yields −N + N + dN + n dx = 0, where N is the normal force, dN its

increment in x direction and n an external line load. Assuming a constant cross-section A, introducing
the volume force f = n/A and using the relation σ = N/A, we can finally state dσ/dx + f = 0. which

9
represents the third equation to determine the displacement u, the strain ε and the stress σ.

Let us summarize:
du
kinematics : ε=
dx
Hooke’s law : σ = E ε (1.1)
dσ
equilibrium : +f =0
dx

Inserting the first and second equation into the third yields with
d2 u
E +f =0 (1.2)
dx2
an ordinary differential equation of second order. The abbreviation
d(...)
= (...)0 (1.3)
dx
allows us to put Equation (1.2) into the well-known form

E u00 + f = 0. (1.4)

The latter equation can be solved analytically if also boundary conditions are provided. The boundary of
the truss system includes the two points x = 0 and x = L. Either the displacement u or the normal force
N = EA u0 must be described. Looking at the example shown in Fig. 1.1, the boundary conditions read
u (x = 0) = 0 and N (x = L) = EA u0 (L) = 0.

1.1.2 Two-dimensional continuum

In the case of a two-dimensional continuum (see Fig. 1.2) the situation becomes already more compli-
cated. Let us interpret the two-dimensional (2D) continuum as one, where the strain εzz in the direction
perpendicular to the considered xy plane vanishes (plane strain state). Please note that many equations
explained in the following hold also for the three-dimensional (3D) case. The differences between 2D
and 3D will be mentioned, where it becomes important.

Kinematics. In Technische Mechanik 2 (Strength of Materials) we have derived for the symmetric strain
matrix ε in 2D the expression
 
1

εxx εxy
 ux,x (ux,y + uy,x ) 
ε= = 1
 2  = sym (grad u), (1.5)
εxy εyy (ux,y + uy,x ) uy,y
2

10
where sym (grad u) denotes the symmetric part of the gradient of the displacement vector
 
ux
u= . (1.6)
uy

The notations
∂ux ∂ux ∂uy ∂uy
ux,x = , ux,y = , uy,x = , uy,y = (1.7)
∂x ∂y ∂x ∂y
are short hand notations for the partial derivative of ux or uy with respect to x or y, respectively. For
later purposes we introduce the so-called Voigt notation
   
 εxx   ux,x 
ε̂ = εyy = uy,y . (1.8)
2 εxy ux,y + uy,x
   

The Voigt notation is differentiated from the matrix notation ε by using a hat over the physical quantity.
In 3D, there are three displacement and six strain components to determine. The dimension of ε is then
3x3. The Voigt notation contains six entries.

Hooke’s law. The elasticity law, also called Hooke’s law, reads for the two-dimensional continuum

σ = 2 µ ε + Λ tr ε I 2 , (1.9)

where tr (...) denotes the trace of a quantity and I 2 the two-dimensional identy matrix. The quantity tr ε
is then given by tr ε = εxx + εyy . Please note again that εzz vanishes in the plane strain case considered
here. In the 3D situation, we need to take into account the corresponding 3x3 strain matrix. Further, the
trace of ε then contains εzz , and I 2 has to be replaced by the three-dimensional identity matrix I 3 .

In Voigt notation we obtain for 2D plane strain linear elasticity the relation
    
 σxx  Λ + 2µ Λ 0  εxx 
σ̂ = σyy = Λ Λ + 2µ 0  εyy = Ê ε̂, (1.10)
σxy 0 0 µ 2 εxy
   

where Ê represents the elasticity matrix. The parameters Λ and µ are two Lamé constants. Thereby,
Λ is often called simply Lamé constant, and µ is the shear modulus frequently also denoted by G. In
engineering, mostly the elasticity constants E (Young’s modulus) and ν (Poisson’s ratio) are used. They
can be easily computed from Λ and µ by using the relations
µ (3 Λ + 2 µ) Λ
E= , ν= . (1.11)
Λ+µ 2 (Λ + µ)
We end up with two unknown displacements ux and uy , three unknown strains εxx , εyy and εxy as well as
three unknown stresses σxx , σyy and σxy . Please note that both matrices ε and σ are symmetric. However,

11
Figure 1.2: Two-dimensional continuum (domain B) with parts of its boundary ∂B either displacement
constrained or loaded. Later, these two parts will be called ∂Bu and ∂Bt , respectively.

we have only six equations (three for the kinematics (1.5) and three for the elasticity law (1.9)). Thus,
two equations are missing which we obtain from stating equilibrium.

Equilibrium. The equilibrium equations can be quite well derived by considering equilibrium at an
infinitesimal volume element of the dimensions dx, dy and t (thickness). This is explained in the lecture
Continuum Mechanics and is – concerning the principal procedure – very similar to the derivation of the
relation dσ/dx + f = 0 in Section 1.1.1. We obtain the vector relation

div σ + f = 0, (1.12)

where div (...) indicates the divergence operator. The vector

 
fx
f= (1.13)
fy

includes the volume forces in x and y direction. Equation (1.12) includes the two scalar equations

σxx,x + σxy,y + fx = 0,
σyx,x + σyy,y + fy = 0, (1.14)

such that there are now eight equations for the eight unknowns available.

In 3D, we need to determine three displacements, six strains and six stresses. In total, we have in the
3D case 15 unknowns. The kinematics yields six equations. From Hooke’s law, six further equations
are available. Thus, three equations are still missing which come from formulating the equilibrium at an
infinitesimal volume element dx dy dz. The equations represented in the following box can be directly
taken over to the 3D situation. In order to retain generality, we replace here I 2 simply by I.

12
In summary, we have to solve the equations

kinematics : ε = sym (grad u)

Hooke’s law : σ = 2 µ ε + Λ (tr ε) I (1.15)
equilibrium : div σ + f = 0

Inserting the kinematical relation ε = sym (grad u) and Hooke’s law σ = 2 µ ε + Λ (tr ε) I 2 into
div σ + f = 0 leads in the 2D case to two scalar partial differential equations for the two components
of the displacement vector u.

In order to understand that, let us replace a vector component (...)x by (...)1 and a vector component
(...)y by (...)2 . In general, we can use the notation ui (i = 1, 2) in order to discuss the displacement
components ux and uy . Accordingly, we use σij (i, j = 1, 2) to talk about the stress components σxx ,
σxy , σyx and σyy . In this way, the kinematical relation (1.151 ) is now transferred into
1
εij =(ui,j + uj,i ), (1.16)
2
where we have additionally used the short hand notation
∂ui
ui,j = . (1.17)
∂xj
Hooke’s law reads in this so-called index notation

σij = 2 µ εij + Λ εkk (I2 )ij . (1.18)

The divergence div σ of a matrix σ (denoted by σij in index notation) is expressed by means of σij,j ,
such that the equilibrium relation becomes

σij,j + fi = 0. (1.19)

At this point, we would like to introduce Einstein’s summation convention which means that we sum
up about indices which occur twice. This means that in (1.18) the expression εkk can be replaced by
ε11 + ε22 + ε33 , where ε33 vanishes in the here assumed plane strain state.

In (1.19) we can write for σij,j the sum σ11,1 + σ12,2 . So, Einstein’s summation convention serves as short
hand notation which makes the resulting index notation very compact.

Inserting (1.16) and (1.18) into (1.19) yields after some steps

µ ui,jj + (µ + Λ) uj,ij + fi = 0. (1.20)

13
Together with the boundary conditions ui = ūi on ∂Bu and ti = σij nj = t̄i on ∂Bt the relation (1.20)
is considered as the so-called strong form in linear elasticity. Here the total boundary ∂B = ∂Bu ∪ ∂Bt
of a domain B has been split into a part ∂Bu where displacements ūi and the remaining part where the
coefficients t̄i are prescribed (see Fig. 1.2). The vector n with nT = {n1 , n2 } is the outward unit normal
vector on the boundary ∂Bt . Note that ∂Bu ∩ ∂Bt = ∅ holds, i.e. on a certain part of the boundary either
displacements or tractions are prescribed.

The analogy between the simple 1D equations in Section 1.1.1 and the more complex corresponding
equations for the 2D (or 3D) continuum derived in Section 1.1.2 should have become clear now. The
equations (1.1) can be directly compared to the ones in (1.15). The final 1D strong form (1.3) plus
boundary conditions is found for the 2D (or 3D) case in the 2D/3D strong form (1.20) plus boundary
conditions.

Remark. It is instructive to study, how the equations for the three-dimensional case reduce to the one-
dimensional case. For this purpose, we first look at Equation (1.18) and evaluate it for the three normal
stresses. This yields for σ11 the relation

σ11 = 2 µ ε11 + Λ (ε11 + ε22 + ε33 ). (1.21)

The results for σ22 and σ33 look analogously. The inversion of (1.11) yields
Eν E
Λ= , µ= . (1.22)
(1 + ν) (1 − 2 ν) 2 (1 + ν)
Together with the well-known relation between ε11 and ε22 = ε33 = −ν ε11 which holds for the one-
dimensional continuum, one arrives after some computation at

σ11 = E ε11 and σ22 = σ33 = 0. (1.23)

The shear stresses are zero, because all shear strains are zero. Obviously, Hooke’s law can then be
alternatively represented by σ = E ε as done in Section 1.1.1. Since there is only one non-zero stress,
the equilibrium condition (1.19) reduces to

σ11,1 + f1 = 0. (1.24)

The latter relation is identical to the equilibrium condition dσ/dx + f = 0 stated in Section 1.1.1. As
already mentioned, the strain state is characterized by the fact that all shear strains are zero. For the
computation of σ11 we need only ε11 which is then equal to u1,1 . This statement is the same as the
kinematical equation ε = du/dx in Section 1.1.1. To conclude, the purpose of this small excursion was
to show that the equations for the 3D continuum include the equations for the 1D continuum (Section
1.1.1) as a special case.

14
1.2 Basic equations (weak form)

Even in the present simple consideration of linear elasticity, the equations (1.20) in strong form are hardly
analytically solvable. Nevertheless, some analytical solutions for special cases are available. But this
does not help us in general. For this reason we strive to apply a numerical method to solve (1.20). One
very robust and frequently used method is the finite element method. In order to be able to apply the finite
element method, the equations (1.20) have to be transferred to another format. This is called the weak
form. Without introducing any concrete discretization, the equation in weak form as it stands below (see
(1.30)) is equivalent to the strong form (1.20). In the context of a numerical method as introduced later,
the new format allows to solve (1.20) only in a weak sense, i.e. in an averaged way. This means for the
finite element formulation described later in this chapter, that equilibrium is only solved approximately,
whereas Hooke’s law and the kinematical relation are fulfilled exactly, i.e. in every point of the structure.

It should be noted already at this point that in Chapter 3, more sophisticated finite element formulations
will be presented which go even beyond that. In particular, the kinematical relation (1.151 ) will be
additionally weakened. In other words, the kinematical relation will no longer be fulfilled pointwise but
only within the element domain. How this can be achieved, will be an important part of Chapter 3.

To make the step of transferring (1.20) derived from the equations (1.16), (1.18) and (1.19) into the
corresponding weak form more transparent, let us multiply (1.19) with a test function or variation δui .
This results into

(σij,j + fi ) δui = (σij δui ),j − σij δui,j + fi δui

1
= (σij δui ),j − σij (δui,j + δuj,i ) + fi δui = 0. (1.25)
2
Here, to rewrite the second term on the right hand side of the equation, the symmetry of σij has been
exploited. The variation δ(...) of a quantity (...) has the same meaning as in Technische Mechanik 1
(Statics), where so-called virtual displacements δu or virtual rotations δϕ were introduced in the context
of the principle of virtual work. Such virtual displacements or rotations are small, imaginary (in other
words virtual), compatible with the geometrical constraints of the system and otherwise arbitrary. The
geometrical compatibility is satisfied by requiring δui to fulfil the homogeneous boundary conditions
δui = 0 on ∂Bu .

The computation of a variation is done analogously to a time derivative. Let us abbreviate the time
derivative d(...)/dt by a dot over the quantity. So, using (1.16), the time derivative of the strain matrix

15
reads
1
ε̇ij = (u̇i,j + u̇j,i ). (1.26)
2
Accordingly, we represent the variation δεij by
1
δεij = (δui,j + δuj,i ). (1.27)
2
Exploiting that in (1.25) and integrating (1.25) over the domain B yields
Z Z Z
σij δεij dV − fi δui dV − (σij ui ),j dV = 0. (1.28)
B B B

Please note that the domain B is here introduced as a three-dimensional continuum, although the equa-
tions are specialized for the plane strain state. The infinitesimal volume element dV reads, as already
introduced in Section 1.1.2, dV = dx1 dx2 t, where t is the thickness of the 2D continuum. The next
step is to apply Gauss’ theorem on the last term in the latter equation. Further, we utilize the boundary
conditions ti = σij nj = t̄i on ∂Bt as well as δui = 0 on ∂Bu and come to the statement
Z Z Z
σij δεij dV − fi δui dV − t̄i δui dA = 0. (1.29)
B B ∂Bt

Additionally, the boundary condition ui = ūi on ∂Bu has to be considered.

In the context of the finite element method, the weak form (1.29) is often represented in Voigt notation.
Then, it reads
Z Z Z
T T
g= δε̂ σ̂ dV − δu f dV − δuT t̄ dA = 0. (1.30)
B B ∂Bt

At the end, we can replace σ̂ by Hooke’s law σ̂ = Ê ε̂. The letter g has been introduced here for later
purposes.

Remark. The weak form (1.30) can be alternatively derived by means of a potential. For this purpose
we set up the potential
Z Z Z
1 T T
π= ε̂ Ê ε̂ dV + (− u f dV − uT t̄ dA) (1.31)
B 2 B ∂Bt
| {z } | {z }
πinner πouter
which consists of the potential of inner and outer forces (represented by the symbols πinner and πouter ,
respectively). Assuming that equilibrium is characterized by a local extremum of the potential π, we
formulate the variation δπ of π and set it equal to zero:
Z Z Z
T T
δπ = δε̂ Ê ε̂ dV + (− δu f dV − δuT t̄ dA) = 0 = g. (1.32)
| B {z } | B {z ∂Bt }
δπinner δπouter

16
Again, the equation g = 0 (1.30) is obtained. Using the fact that π = −W holds, where the symbol W
denotes the mechanical work, we are able to state that (1.32) and then also (1.30) represents the principle
of virtual work. The latter, to be indicated by δW = −δπ, has to vanish.

1.3 Interpolation of displacement and geometry

The finite element method is based on the idea to discretize the weak form, here (1.30), by suitable in-
terpolation functions for unknown physical quantities as well as for the geometry. As already pointed
out, if kinematics and the elasticity relation are inserted into the equilibrium condition, only the dis-
placement vector u will remain as unknown quantity. The interpolation functions of the standard linear
isoparametric displacement formulation (2D) are often represented in the form
4
X
u(e) = NI uI , (1.33)
I=1

where the so-called shape functions (I = 1, ..., 4)

1
NI = NI (ξ, η) = (1 + ξ ξI ) (1 + η ηI ) (1.34)
4
depend on the local coordinates ξ and η defined on a quadratic reference element  = [−1, 1] × [−1, 1]
(see Fig. 1.3). The index (e) is used to indicate that the interpolation (1.34) refers to the element e. The
brackets mean that this index can be also omitted. The vector uI = u (x = xI ) (I = 1, ..., 4) refers to
the displacement vector at node I.

node ξI ηI
1 -1 -1
2 1 -1
3 1 1
4 -1 1

Figure 1.3: Quadratic reference element  = [−1, 1] × [−1, 1] and local coordinates of the nodes
I = 1, ..., 4.

The functions NI (I = 1, ..., 4) are accordingly expressed by

1 1
N1 = (1 − ξ − η + ξ η), N2 = (1 + ξ − η − ξ η),
4 4
17
 1 1
N3 = (1 + ξ + η + ξ η), N4 = (1 − ξ + η − ξ η). (1.35)
4 4
We will use the notation
         

 N 1 
 
 1 
 
 −1 
 
 −1 
 
 1 


N2
 1 
1
 1 
1
 1 
−1
 1  −1 
N= = +ξ +η +ξη

 N3  4 
 1 
 4 
 1 
 4 
 1 
 4 
 1 

N4 1 −1 1 −1
         
| {z } | {z } | {z } | {z }
r gξ gη h
= r + ξ gξ + η gη + ξ η h (1.36)

in the following. The 4x1 vector N can be termed shape function vector. The nodal displacements (uI )x
and (uI )y (I = 1, ..., 4) are included in the 8x1 vector
 

 (u 1 )x 

(u )
 
2 x

 

 
(u )
 
3 x 

 
 
  
(u4 )x (U e )x

Ue = = (1.37)

 (u1 )y  (U e )y
(u2 )y 

 


 
(u )
 
3 y

 

 
(u4 )y
 

which includes the two 4x1 vectors (U e )x and (U e )y . Let us call U e the element nodal diplacement
vector from now on. (U e )x and (U e )y contain the contributions of U e in x and y direction, respectively.
The vectors r, g ξ , g η and h play an important role in finite element technology. For later purposes, it
is helpful to visualize them as displacement modes, for instance in terms of (U e )x , see Fig. 1.4. The
descriptions translation, elongation and shear are self-explanatory. Surprising is the so-called hourglass1
mode h which is needed to depict e.g. bending.

We can finally write ux = N T (U e )x and uy = N T (U e )y . Further, we introduce

 " T #
oT4

ux N
u = u(e) = = U e. (1.38)
uy oT4 N T
In the latter relation, the vector o4 indicates a 4x1 vector with only zero entries. Characteristic for the so-
called isoparametric element concept which is used here (see e.g. the text books [Hug00], [ZTZ05]), is
the fact that element geometry and displacements are interpolated by means of the same ansatz functions.
Therefore, we state for the position vector
" #
N T oT4
 
x
x = x(e) = = X e, (1.39)
y oT4 N T
1
It got its name from an old-fashioned instrument to measure the time called sand glass or hourglass.

18
 a) b) c) d)

Figure 1.4: Visualization of the vectors r, g ξ , g η and h by means the nodal diplacement vector in x
direction, i.e. a) translation (U e )x = r, b) elongation (U e )x = g ξ , c) shear (U e )x = g η , d) hourglass
displacement (U e )x = h. The continuous line indicates the original element shape (here a square). The
dot-dashed line shows the displacement (U e )x .

a relation which is analogous to (1.38). The 8x1 vector X e includes at the first four positions the x
coordinates and at the last four positions the y coordinates of the nodes. Using (1.39), some domain B
is approximated by Bh = ∪ne=1
e
Bh e (e element index, ne number of elements). The index h indicates
an approximation. The approximation (1.39) leads to the fact that the element domain Bh e represents a
quadrilateral. Consequently, the boundary ∂Bh is not a smooth line but represents a continuous assembly
of straight lines with different slopes, see Fig. 1.5.

Figure 1.5: Domain B and its approximation Bh , discretization into finite elements Bh e . The boundaries
of Bh and Bh e are called ∂Bh and ∂Bh e , respectively.

1.4 Split of N into linear and hourglass parts

For the finite element technology to be introduced later, it is suitable to split the shape function vector
N into a part N lin which is linear in x and the remainder, the hourglass part N hg .2 The split N =
N lin + N hg can be seen as a representation of N which is alternative to (1.36). The thought behind
2
The notion N hg is used here, because this part of N = N lin + N hg will be responsible for the contribution of the
hourglass mode h in N (see also Eq. (1.36)).

19
that will become clearer in Section 1.5, where the strain-displacement operator B in ε̂ = B U e will be
derived. Splitting B accordingly, the element stiffness matrix K e to be computed in Section 1.7 can
be determined and analyzed analytically. This will enable us to determine its eigenvalues and judge the
quality of a finite element formulation, one of the major tasks in finite element technology.

To determine N lin and N hg , we carry out a Taylor expansion of N with respect to the centre of the
element given by x(e) = x0 . The vector x0 is obtained by evaluating the position vector x(e) at the
position ξ = o2 . The vector ξ is defined by ξ T = {ξ, η}. The vector o2 is a 2x1 vector with only zero
entries. The Taylor expansion with respect to the centre of the element characterized by x(e) = x0 or
equivalently ξ = o2 is represented by
∂N
N = N |ξ=o2 + (x − x0 ) + N hg
∂x ξ=o2
∂N ∂ξ
= N0 + (x − x0 ) + N hg
∂ξ ∂x ξ=o2

∂N −1
= N0 + J (x − x0 ) + N hg
∂ξ ξ =o2
 
∂N ∂N
= N0 + J −1
0 (x − x0 ) + N hg
∂ξ ∂η ξ =o2

= r + g ξ g η J −1
 
0 (x − x0 ) + N hg

= N lin + N hg . (1.40)

Some parts of these lines should be explained in more detail. We will always use index 0, if a quantity
is evaluated at ξ = o2 . This notation holds, for instance, for the quantities N 0 or J −1
0 . The matrix
J = ∂x/∂ξ is termed Jacobi matrix. Its inversion reads J −1 , and the evaluation of the latter matrix at
ξ = o2 yields (J −1 )0 . But it should be noted that inverting J 0 yields the same result as first inverting
J and then evaluating this matrix at ξ = o2 . Therefore, we can write (J 0 )−1 = (J −1 )0 = J −1
0 . The
derivative of the shape function vector N with respect to ξ yields a 4x2 matrix. This matrix includes the
two columns ∂N /∂ξ and ∂N /∂η, see the second last line of (1.40). Using (1.36), the two columns read
in detail
∂N ∂N
= g ξ + η h and = g η + ξ h. (1.41)
∂ξ ∂η
The evaluation of the latter expressions at ξ = o2 leads to the statement
 
∂N ∂N  
= gξ gη (1.42)
∂ξ ∂η ξ =o2
which has been used to come from the second last to the very last line of (1.40). Obviously, the part N lin
depends linearly on x. It remains to find a concrete expression for the second summand N hg .

20
For this purpose, Equation (1.39) is rewritten to read
  " #
x (X e )Tx
x= = N, (1.43)
y (X e )Ty
| {z }
xnode

where we have introduced the 2x4 matrix xnode . It is used to derive the equation

x − x0 = xnode (N − N 0 ) = xnode (ξ g ξ + η g η + ξ η h). (1.44)

Additionally, we use the relation

∂x
J −1 −1
= J −1
 
0 J0 = J0 0 xnode gξ gη = I 2. (1.45)
∂ξ ξ=o2

The matrix J 0 is expressed by taking the derivative of xnode N with respect to ξ and evaluating the result
at ξ = o2 . This leads to the expression J 0 = xnode [g ξ g η ] which is part of (1.45). The 2x2 identity
matrix I 2 includes two columns, first of all the cartesian basis vector ex represented by eTx = {1, 0},
secondly the cartesian basis vector ey given by eTy = {0, 1}. This enables to note the equations
   
1 0
J −1
0 xnode g ξ = , J −1
0 xnode g η = . (1.46)
0 1

From (1.40), (1.44) and (1.46) we obtain

J −1 J −1
   
N lin − r = gξ gη 0 (x − x0 ) = gξ gη 0 xnode (ξ g ξ + η g η + ξ η h)
   
  1   0
= gξ gη ξ + gξ gη η
0 1
  −1
+ g ξ g η J 0 xnode ξ η h

= ξ g ξ + η g η + g ξ g η J −1
 
0 xnode ξ η h. (1.47)

The latter result is now utilized to derive the equation

N hg = N − r − (N lin − r)
= ξ gξ + η gη + ξ η h
− (ξ g ξ + η g η + g ξ g η J −1
 
0 xnode ξ η h)

= (I 4 − g ξ g η J −1
 
0 xnode ) ξ η h = ξ η γ (1.48)

where the last line defines the so-called stabilization vector γ [BOL+84]. This vector has an important
meaning in finite element technology as will be shown in the following sections.

21
With the short hand notation
J −1
   
bx by = gξ gη 0 , (1.49)

the result of the previous derivation reads

 
N lin = r + bx by (x − x0 ),
(1.50)
N hg = ξ η γ .

1.5 Derivation of the strain-displacement operator

The weak form (1.30) contains several quantities which have to be expressed by means of the interpo-
lation introduced in Section 1.3. Among these the strain vector ε̂ is crucial which also enters the stress
vector σ̂ via the elasticity law σ̂ = Ê ε̂. Using (1.38), the following relation between ε̂ and the element
nodal displacement vector U e is given:
 T 
    N ,x oT4
 εxx   ux,x   T 
ε̂ = εyy = uy,y = 
 o4 N T,y  Ue

2 εxy ux,y + uy,x
   
N T,y N T,x

= B U e. (1.51)

At the same time, we have thus already introduced the strain-displacement operator B which is in the
2D case a 3x8 matrix. It is usually called B matrix or operator. Analogously to Section 1.4, where we
split N = N lin + N hg into a linear (N lin ) and an hourglass (N hg ) part, we additively decompose the
B operator into a constant and an hourglass part: B = B 0 + B hg . The first summand is computed from
N lin , the second from N hg . This is conveniently done by first exploiting Eq. (1.501 ) and rewriting it to

N lin = r + (x − x0 ) bx + (y − y0 ) by . (1.52)

It is now very simple to formulate the derivative of N lin with respect to x and y. Obviously, we obtain
the relations
∂N lin ∂N lin
= bx , = by (1.53)
∂x ∂y
which will be inserted into (1.51). To calculate the derivative of N hg = ξ η γ, it has to be understood
that the stabilization vector γ is a vector which is constant within the element. Thus, it does not depend
on x. Therefore, to determine B hg we only need to compute the derivative of ξ η with respect to x and

22
to y. This leads to
  
∂(ξ η) ∂ξ ∂η ∂ξ ∂η η
= η +ξ = ,
∂x ∂x ∂x ∂x ∂x ξ
  
∂(ξ η) ∂ξ ∂η ∂ξ ∂η η
= η +ξ = (1.54)
∂y ∂y ∂y ∂y ∂y ξ

and finally to      
∂N hg ∂ξ ∂η η ∂N hg ∂ξ ∂η η
= γ, = γ, (1.55)
∂x ∂x ∂x ξ ∂y ∂y ∂y ξ
| {z } | {z } | {z } | {z }
T
= cx = l hg = cTy = lhg
where we have introduced the vectors cx , cy and lhg as short hand notations. From (1.51), we finally
obtain the relation
 T   T   
N ,x oT4 bx oT4 cTx oT2   T
γ oT4
 
lhg o2
B =  oT4 N T,y  =  oT4 bTy T 
  T
 +  o2 cy 
  
o2 lhg oT γ T
N T,y N T,x bTy bTx cTy cTx | {z } | 4 {z }
| {z } | {z } = Lhg = M hg
= B0 =j

= B 0 + B hg . (1.56)

In particular, the construction of the 3x8 matrix B hg = j Lhg M hg is not trivial. It should be recapit-
ulated that j is a 3x4 matrix, Lhg a 4x2 matrix and M hg a 2x8 matrix. Please note again that the 1x2
vectors cTx , cTy and oT2 contain two entries in a row. The 2x1 vectors lhg and o2 represent two entries in
a column, and the 1x4 vectors γ T and oT4 show four entries in a row.

For the later chapters, it should be emphasized that the matrix B 0 is constant within the element. This is
due to the fact that also bx and by are constant within the element. Otherwise, N lin would not be linear
in x and y. The matrix Lhg is linear in ξ and η, whereas M hg is again constant within the element. The
matrix j is constant in special cases. However, in general, it is a non-linear function in ξ and η which we
have to investigate. This will be done in the next section. For now, it should be kept in mind that B 0 is in
any case constant within the element. The matrix B hg is in general a non-linear function which reduces
in special cases to a linear polynomial in ξ and η.

23
1.6 Jacobi matrix

1.6.1 General remarks

The Jacobi matrix J = ∂x/∂ξ was already introduced in Section 1.4. Using the knowledge (visible also
from (1.39)) that the position vector x is a function of the vector ξ, i.e. x = x (ξ) holds, we can also
write
∂x
dx = dξ = J dξ. (1.57)
∂ξ
Thus, in the context of the isoparametric mapping (see Fig. 1.6), it can be said that J maps a line element
dξ of the reference element  on a line element dx of the “real” element Bh e .

Let us take two line vectors dx1 and dx2 (see the magnification in Fig. 1.6) and formulate their cross
product dx1 × dx2 . The absolute value |dx1 × dx2 | of this cross product is the area of the parallelogram
spanned by these two vectors. From mathematics (see e.g. [Cha12]) we know that using (1.57) the
latter expression can be alternatively expressed by det J |dξ 1 × dξ 2 |. Without loss of generality we
may choose dξ 1 = dξ eξ and dξ 2 = dη eη , where eξ and eη are unit vectors which are orthogonal to
each other. With that, we obtain for the infinitesimal area the expression det J dξ dη and finally for the
infinitesimal volume the relation
dVe = det J dξ dη t, (1.58)

with t denoting the thickness of the element Bh e . Please note that this is already a specialized relation for
the present plane strain case. The product dξ dη t can be seen as an infinitesimal volume of the reference
element.

Analogy to continuum mechanics. Such a procedure is also known from continuum mechanics. Let
us define with xdef the position vector to a material point in the deformed configuration (see Fig. 1.7).
The function xdef = xdef (x) represents the deformation. This is usually time-dependent. We neglect
this aspect which is not important for the present derivation. The partial derivative F = ∂xdef /∂x is a
central notion in continuum mechanics and called deformation gradient. Analogously to (1.57), we can
formulate
∂xdef
dxdef = = F dx. (1.59)
∂x
The matrix F maps a line element dx of the undeformed configuration (often called reference configu-
ration) on a line element of the deformed configuration. The relation between the corresponding volume
elements dVdef and dV reads dVdef = det F dV . The analogy to the isoparametric mapping is obvious
and can be used to interpret the real element as a “deformation” of the reference element. Nevertheless,

24
by following this thought, it should be taken into account that the “deformation” from the reference to
the real element is not a physical deformation but occurs in an artificial way by choosing the shape of
the real element. 

Figure 1.6: Isoparameteric mapping expressed by J and J −1 between the reference (“fictive”) element
 and the “real” element Bh e . Magnification: infinitesimal area expressed by |dx1 × dx2 |.

Figure 1.7: Mapping expressed by F and F −1 between the undeformed and the deformed configuration
in continuum mechanics. The vectors x and xref are the position vectors to material points (see the
bullets) in the undeformed and the deformed configuration, respectively.

1.6.2 Special cases of the Jacobi matrix

For further derivations, it is important to understand, how the Jacobi matrix is constructed. Therefore,
we look in more detail at the representation (1.39) which can be, using N = r + ξ g ξ + η g η + ξ η h,
alternatively given by
x = (r T + ξ g Tξ + η g Tη + ξ η hT ) (X e )x , y = (r T + ξ g Tξ + η g Tη + ξ η hT ) (X e )y . (1.60)

25
The 2x2 Jacobi matrix is then written as
 " T T T
T
#
(g + η h ) (X ) (g + ξ h ) (X )

x,ξ x,η ξ e x η e x
J= = . (1.61)
y,ξ y,η (g ξ + η h ) (X e )y (g Tη + ξ hT ) (X e )y
T T

It is a linear function of ξ and η, because the 4x1 vectors (X e )x and (X e )y contain only the coordinates
of the element and are thus constant with respect to one element. However, the inversion
 −1  
−1 ∂x ∂ξ   ξ,x ξ,y
J = = = cx cy = (1.62)
∂ξ ∂x η,x η,y

of J is then certainly a highly non-linear function of ξ and η, since ξ and η show up in the denominator.
This issue is important for the computation of the matrix j which is needed in B hg (see Section 1.5). For
the definition of cx and cy see also Section 1.5.

Figure 1.8: Element with parallelogram shape.

Special case 1: element of parallelogram shape. Let us assume that the element Bh e has parallelogram
shape (see Fig. 1.8). Without loss of generality, the origin of the coordinate system x, y is put into the
left bottom node of the element. The vectors (X e )x and (X e )y then read

(X e )Tx = {0, 2 a, 2 a + 2 u, 2 u}, (X e )Ty = {0, 2 v, 2 b + 2 v, 2 b}. (1.63)

To evaluate the functions presented in (1.60), the quantities

r T (X e )x = a + u, g Tξ (X e )x = a, g Tη (X e )x = u, hT (X e )x = 0
r T (X e )y = b + v, g Tξ (X e )y = v, g Tη (X e )y = v, hT (X e )y = 0 (1.64)

need to be computed. Very interesting is the fact that both terms hT (X e )x and hT (X e )y vanish. Thus,
the linear dependence in the general format of J (see Equation (1.61)) is no longer present. For an
element of parallelogram shape, the Jacobi matrix reduces to
" T # 
g ξ (X e )x g Tη (X e )x

a u
J par = = . (1.65)
g Tξ (X e )y g Tη (X e )y v b

26
This is a matrix which is constant within the element. Obviously, then, also J −1
par and the matrix j derived
from J −1
par are constant. The matrix B hg becomes linear in ξ and η which is a very advantageous property
to be exploited later.

Special case 2: element of rectangular shape. For very simple structures, it might be even possible to
choose elements of rectangular shape. A rectangle is a parallelogram with u = v = 0. The Jacobi matrix
becomes a diagonal matrix and reads  
a 0
J rec = . (1.66)
0 b
This result is also easily understood by interpreting the shape of the element as a pure stretch (or con-
traction) of the reference element. The real element is now a times longer (or shorter) in x direction and
b times longer (or shorter) in y direction than the reference element. Due to the fact that a and b could
take on all positive values, both stretch and contraction is possible in both directions.

1.6.3 Evaluation of the Jacobi matrix in the centre of the element

Even if the element has a general shape, it is of value to extract a physically reasonable constant part of
J . This is suitably done by evaluating (1.61) for ξ = o2 which leads with
" T #
g ξ (X e )x g Tη (X e )x
J0 = (1.67)
g Tξ (X e )y g Tη (X e )y

to the already introduced Jacobi matrix in the centre of the element (see Section 1.4). Obviously, J 0 is
constant and equal to the Jacobi matrix of an element with parallelogram shape. Let us call that element
equivalent parallelogram.

Figure 1.9: Example for equivalent parallelogram. The solid line shows the real (original) element.
Its Jacobi matrix evaluated in the centre of this element yields J 0 . The dot-dashed line depicts the
corresponding equivalent parallelogram. Its Jacobi matrix J par is equal to J 0 .

27
1.6.4 Equivalent parallelogram

The next section is mathematically quite challenging. Since it is not necessary for the understanding of
the remainder of the lecture, it can be easily postponed to later reading. Nevertheless, it gives interesting
insights for later use in finite element technology.

It is instructive to compute the equivalent parallelogram mentioned in Section 1.6.3. Obviously, the
equivalent parallelogram will be different from the real element, if the real element is not a parallelo-
gram. So, the question is, how we can come to the coordinates (X e )eqp
x and (X e )eqp
y of the equivalent
parallelogram whose Jacobi matrix J par is equal to J 0 computed for the real element. In other words:
we seek to find a parallelogram shaped element with the Jacobi matrix J par = J 0 , where J 0 is the Jacobi
matrix evaluated in the centre of the real element. See Fig. 1.9 for illustration.

Please remember that the relation (1.61) represents the Jacobi matrix of an element with general shape.
If we evaluate (1.61) in the centre of the element, we will come to the equation (1.67). The difference lies
in the contributions including hT (X e )x and hT (X e )y . So, for the equivalent parallelogram, the terms
involving g ξ and g η have to be retained, whereas the terms involving h shall vanish. The following
conditions must hold for (X e )eqp eqp
x and (X e )y in order to satisfy that:

g Tξ (X e )eqp
x = g Tξ (X e )x , g Tη (X e )eqp
x = g Tη (X e )x , hT (X e )eqp
x =0
g Tξ (X e )eqp
y = g Tξ (X e )y , g Tη (X e )eqp
y = g Tη (X e )y , hT (X e )eqp
y = 0. (1.68)

Mathematically spoken, the projections of (X e )x and (X e )y in the direction of g ξ and g η have to be

retained, the ones in the direction of h have to be eliminated. To achieve that, we construct an orthogonal
basis by transforming r, g ξ , g η and h into four unit vectors:

r̄ = 2 r, ḡ ξ = 2 g ξ , ḡ η = 2 g η , h̄ = 2 h. (1.69)

These four vectors can be put into a matrix R which reads

 
R= r̄ ḡ ξ ḡ η h̄ . (1.70)

It can be easily verified that this matrix is orthogonal (RRT = I) which is due to the fact that the four
basis vectors are unit vectors and orthogonal to each other. So, any vector v = I v can be alternatively
represented by
T
v = (RRT ) v = R (RT v) = (r̄ T v) r̄ + (ḡ Tξ v) ḡ ξ + (ḡ Tη v) ḡ η + (h̄ v) h̄. (1.71)

28
The application of this transaction on (X e )x and (X e )y allows us to represent the two latter vectors by
means of the relations
       
T
(X e )x = r̄ T (X e )x r̄ + ḡ Tξ (X e )x ḡ ξ + ḡ Tη (X e )x ḡ η + h̄ (X e )x h̄,
       
T T T T
(X e )y = r̄ (X e )y r̄ + ḡ ξ (X e )y ḡ ξ + ḡ η (X e )y ḡ η + h̄ (X e )y h̄. (1.72)

Finally, we simply keep the first three terms in each line to arrive at
   
T T
(X e )eqp
x = (X e )x − h̄ (X e )x h̄ and (X e )eqp
y = (X e ) y − h̄ (X e )y h̄. (1.73)

It can be easily verified that in this way, the conditions (1.68) are fulfilled.

Example. In order to understand the derivation in a better way, an example is investigated. See Fig. 1.9.
The coordinates of the real element are given by

(X e )Tx = { 0 8 4.5 1 }, (X e )Ty = { 0 1.5 4.5 6 }. (1.74)

Using (1.67), the Jacobi matrix in the centre of the element is computed to read
 
(23/8) −(5/8)
J0 = . (1.75)
0 (9/4)
T T
The formulas (1.73) yield with h̄ (X e )x = −(9/4) and h̄ (X e )y = −(3/2) the result

1 1
((X e )eqp T eqp T
x ) = { 9 55 45 −1 } , ((X e )y ) = { 3 3 21 21 } . (1.76)
8 4
It is quickly checked that Eq. (1.67) yields again (1.75).

1.7 Finite element matrices and vectors

The previous results are now inserted into the weak form (1.30). For this purpose, we rewrite (1.30) in
the format ne Z Z Z 
X
T T T
δε̂ Ê ε̂ dVe − δu f dVe − δu t̄ dAe = 0. (1.77)
e=1 Bh e Bh e ∂Bt h e

The only difference with respect to (1.30) is the split of the integral over the whole domain Bh into ne
integrals over the element domains Bh e . The index h indicates, as already stated in Section 1.3 that due
to the interpolation of the geometry, the domain B might be represented in an approximated way. Since
this approximation is not a critical point in finite element technology, we omit this index from now on.

29
To prepare the derivation of the stiffness matrix and the load vector, the short hand notation
" #
N T oT4
S= (1.78)
oT4 N T

is introduced. The displacement vector u can then be very compactly written as u = S U e (see Eq.
(1.38)). This holds analogously for the virtual displacement vector δu = S δU e . Further important
for the derivation of the finite element matrices and vectors are the relations ε̂ = (B 0 + B hg ) U e ,
δε̂ = (B 0 + B hg ) δU e (see Section 1.5) and dVe = det J dξ dη t (see Section 1.6.1). The infinitesimal
area dAe refers to the boundary of the element. As such, it can be expressed as dAe = ds t, where ds
represents an infinitesimal line increment along the boundary ∂Bt e . The interval for the integration over
s is denoted by st e . The computation of the load vector is certainly important for the use of the finite
element method. However, the derivation is standard and not crucial for the topic of finite element
technology. Therefore, we refer to relevant text books such as [BLM+14], [Hug00], [OBC71] and
[ZTZ05] and restrict ourselves to the expression for the 8x1 load vector P e at element level given below.

We insert the afore-mentioned relations into (1.77) and summarize the loading terms (2nd and 3rd term)
in the expression δU Te P e with
Z Z
T
Pe = S f det J dξ dη t + S T t̄ ds t. (1.79)
 st e
R
The integral 
(...) is a short hand notation for the double integral over the domain of the reference
element  = [−1, 1] × [−1, 1]. Equation (1.77) finally reads

Xne h Z i
T
δU e (B 0 + B hg )T Ê (B 0 + B hg ) det J dξ dη t U e − δU Te P e = 0. (1.80)
e=1 | {z }
Ke

The 8x8 matrix K e represents the element stiffness matrix which plays a crucial role in finite element
technology.

1.7.1 Evaluation of K e by means of standard numerical integration

At this point, it is important to concentrate on the integration to be carried out for the computation of
K e . As already pointed out, in general, B hg is a highly non-linear function in ξ and η. Therefore, the
integration cannot be performed analytically. It is standard to use Gauss integration for this purpose.
More information about such a procedure can be found in the already mentioned general text books on

30
the finite element method. The element stiffness matrix is finally approximated by
np
X
Ke ≈ (B 0 + B hg )Tp Ê (B 0 + B hg )p det J p t wp , (1.81)
p=1

where np denotes the total number of Gauss points, p the number of one Gauss point and wp the weight
for the contribution of Gauss point p. Very frequent is the choice of 2x2 Gauss point integration with the
Gauss point coordinates and weights plotted in the following table:

p=1 p=2 p=3 p=4

√ √ √ √
ξp −1/√3 1/ √3 1/√3 −1/√ 3
ηp −1/ 3 −1/ 3 1/ 3 1/ 3
wp 1 1 1 1

Table 1.1: Coordinates and weights for 2x2 Gauss point integration.

The error by applying this kind of numerical integration is well investigated and very small.

1.7.2 Evaluation of K e by means of analytical integration

In Section 1.6.2, it was already discussed that, if the element had parallelogram shape, the Jacobi matrix
J would be constant and then be certainly equal to J 0 . The hourglass part of the B operator would
reduce to a linear function B hg = j 0 Lhg M hg in ξ and η. Also det J = det J 0 would be a constant.
The integrands B T0 Ê B hg and B Thg Ê B 0 (see below) would include linear functions of ξ and η which
are symmetric with respect to the ξ or η axis. This has the consequence that the integrals
Z Z
T
B 0 Ê B hg det J 0 dξ dη, B Thg Ê B 0 det J 0 dξ dη, (1.82)
 

would vanish, and the remaining two terms in K e could be integrated analytically.

In Section 1.6.3, the equivalent parallelogram was introduced. The difference between the original
(general) shape of the element and the equivalent parallelogram is a good measure for the error made by
working with J 0 instead of with J . It is a valid suggestion to work always with J 0 , even if the shape
of the element deviates from a parallelogram. Such a step would allow to work always with analytical
integration which is on the one hand a significant advantage. This becomes particularly evident in 3D
simulations with non-linear behaviour (which is not part of the present course).

31
On the other hand, we have to be aware of the fact that replacing J by J 0 introduces an error into the
finite element modeling which has to be quantified in some way. Mathematical work on this topic is for
instance found in [AR95a] and [AR95b]. By engineering intuition, an error measure such as
(Jij − (J0 )ij ) (Jij − (J0 )ij )
errJ −J 0 = (1.83)
Jkl Jkl
could be a reasonable choice. It can be quite easily shown that even if the shape of an element deviates
significantly from a parallelogram, further refinement reduces the error (1.83) very quickly. In other
words, using J = J 0 might influence the results for coarse meshes. Nevertheless, for sufficiently fine
meshes which are anyway needed to reach convergence to a solution, the result obtained by J 0 agrees
with the one obtained by J .

Replacing J by J 0 at all locations, where components of J or J −1 show up, finally leads to

Z
Ke ≈ 4 B T0 Ê B 0 det J 0 t + B Thg Ê B hg det J 0 dξ dη t. (1.84)


Again, it should be emphasized that the integral in the second summand, based on B hg = j 0 Lhg M hg ,
can be evaluated analytically because it contains polynomials only.

1.8 Assembly procedure

The equation to be finally solved is obtained by the standard assembly procedure described in many
books. See for example [BLM+14], [Hug00], [OBC71] and [ZTZ05].

The procedure consists of the following steps. At first, all displacements are assembled in the vector U
e.
Equation (1.80) then reads
e T (K
δU fUe −P
e ) = 0, (1.85)
where the element stiffness matrices K e are assembled into K f and the element load vectors P e into
P
e . In the latter relation, the boundary conditions u = ū on ∂Bu are not yet considered. This has the
consequence that U
e still contains known nodal displacements. Further, The matrix K
f is not invertible.
After taking the displacement boundary conditions into account, the equation δU T (K U − P ) = 0 is
obtained. Here, the global nodal displacement vector U includes only the unknown nodal displacements.
K is termed global stiffness matrix and is now invertible. P is the global load vector. The latter equation
has to hold for arbitrary δU which finally leads to the equation system

KU =P (1.86)

32
which has to be solved to determine the unknown nodal displacements stored in U . Using the interpola-
tion at the element level, the strain vector ε̂ can be computed and based on that the stress vector σ̂. The
stress vector is usually one of the important quantities of interest.

1.9 Exercises

... general information ...

1.1 Introduction to MATLAB

(a) first set of basic commands: vectors, arrays, matrices

(b) programming statements: FOR loop, WHILE loop, IF-ELSEIF-ELSE statement, SWITCH-
CASE statement
(c) plotting: multiple plots, contour plots
(d) functions: ...
(e) performance: matrix multiplication, inverse of a matrix
(f) second set of basic commands: create symbolic variables and functions (Syms), Jacobian ma-
trix (Jacobin), equations and systems solver (Solve), (Expand), symbolic substitution (Subs),
differences and approximate derivatives (Diff), algebraic simplification (Simplify), definite
and indefinite integrals (Int), set assumption on symbolic object (Assume), 2-D and 3-D grids
(Meshgrid), create sparse matrix (Sparse), reshape array (Reshape)

1.2 In continuum mechanics the basic equations are derived from kinematics, equilibrium, and the
material law. To illustrate how the corresponding equations can be analyzed, let us imagine a
two-dimensional continuum in a plane strain state (εzz = 0). Further, we assume to know the
displacement field as a function depending on x and y. These displacements can be stated as
   
ux sin (x) y
u (x, y) = = 0.001
uy x2 + y exp (x/2)

where the considered continuum is defined in the range of x ∈ [0; 2] and y ∈ [0; 2].

(a) Compute the symmetric strain matrix ε. Plot the components of ε.

33
(b) The elasticity constants Λ and µ can be represented in terms of the Young’s modulus E and
the Poisson’s ratio ν (see Eq. 1.22)). The Young’s modulus and the Poisson’s ration are given:
E = 100000 MPa, ν = 0.3. Compute Λ and µ using (1.22). Calculate the stress components
by means of Eq. (1.9) and plot them.
(c) Further, the equilibrium equation given in vector form by (1.12) has to hold true. Compute
the volume force vector f and show the corresponding contour plots.

34
Chapter 2

Locking

The locking effect is characterized by an overestimation of the stresses and an underestimation of the
deformation. It occurs even for relatively fine meshes and is typical for isoparametric displacement
formulations of low order as the one presented in Chapter 1. Let us denote this element formulation as
Q1. The letter Q stands for quadrilateral, the number 1 for first order (the shape functions contain only
linear terms in ξ and η). The version, where J is replaced by J 0 will be termed Q1J0 from now on.
An increase of the number of elements leads to an only slight improvement of the solution. One usually
denotes that as “bad convergence behaviour” (the element “locks”).

2.1 Typical examples for locking

2.1.1 Example for volumetric locking

In Fig. 2.1, geometry and boundary conditions of an almost incompressible two-dimensional block are
depicted. Additionally, the displacements at the top of the structure are constrained in the horizontal
direction. “Almost incompressible” means that the material parameter ratio Λ/µ is very large. The
material is non-linearly elastic, i.e. it is able to undergo large deformations. A soft polymer is a good
example for a material which behaves in this way. Certainly, this goes beyond the present course, and
therefore no further details are given. But due to the significant deformation, the locking is optically
well recognizable. In Fig. 2.2, the deformed systems computed by means of (a) a locking-free element
formulation and (b) an element formulation which exhibits locking are depicted. In both pictures, the
same loading was chosen. The dramatic influence of the locking effect becomes evident. It leads to

35
a clear underestimation and wrong representation of the deformation state, whereas the stress is far
overestimated (not visible in the figures). In short, locking makes the structure artificially stiff.

Unfortunately, the behaviour does not majorly improve with increasing mesh density. This is shown in
the study of convergence, see Fig. 2.3. The element formulation Q1 (standard displacement approach,
see Chapter 1) which exhibits locking – to be shown in the present chapter – will not reach the converged
solution, even if the number of elements is increased up to 300 elements. Certainly, 300 elements are
not many and one could argue that such a mesh is still not fine enough. However, using so-called
locking-free element formulations such as Q1SP and Q1/P0 – these abbreviations will be explained later
– convergence is already achieved for 128 elements. Convergence means here that the solution is hardly
altered by further mesh refinement.

It should be already mentioned at this point that we see here an example for so-called volumetric locking
which appears for large material parameter ratios Λ/µ.

Figure 2.1: Geometry and boundary conditions of an almost incompressible block made of non-linearly
elastic material. Horizontal displacements at top are additionally constrained (not shown).

(a) (b)

Figure 2.2: Deformed mesh for a given loading state, (a) result of a locking-free element formulation,
(b) result of an element formulation which exhibits locking.

36
 Study of convergence (almost incompressible block)
70

compression [%] in point A

Q1SP: nu=20
60 Q1SP: nu=40
Q1SP: nu=60
50 Q1/P0: nu=20
Q1/P0: nu=40
40 Q1/P0: nu=60
30 Q1: nu=20
Q1: nu=40
20 Q1: nu=60
10
0
0 50 100 150 200 250 300
number of elements
Figure 2.3: Study of convergence for element formulations Q1SP, Q1/P0 and Q1 and different load
factors ν = p/p0 (“nu”). The formulation Q1 is the standard displacement approach described in Chapter
1. It exhibits locking which will be discussed in this chapter. Q1SP and Q1/P0 are more sophisticated
element formulations which are locking-free in this example.

2.1.2 Example for shear locking

Another group of applications which is prone to locking are bending-dominated problems. See e.g. the
thick shell under a line load plotted in Fig. 2.4. We also see the deformed structure computed by a
locking-free formulation. The study of convergence shown in Fig. 2.5 looks similar to the one of the
almost incompressible block in Fig. 2.3. Q1 does not reach the correct result whereas the two locking-
free formulations Q1SP and Q1/E12 – these abbreviations will be explained later – need only 16 elements
in the circumferential direction to obtain a relatively good result.

In this example, the locking is due to the fact that the deformation is dominated by bending and shear.
We observe the so-called shear locking.

Concluding remark. It should be emphasized again that the shown examples include large deformations
and non-linearly elastic material behaviour. The locking is, however, also visible in small deformation
problems, even in linear elasticity which is the topic of the present course. To understand this phe-
nomenon at the theoretical level, we will analyze the eigenvalues of the element stiffness matrix K e in
what follows.

37
 p=1800 kN/cm
A

r=3m
t = 0.3 m

L=5m

(a) (b)

Figure 2.4: Geometry and boundary conditions of a thick shell made of non-linearly elastic material, (a)
undeformed structure, (b) deformed structure computed by a locking-free element formulation.

2.2 Eigenvalue analysis for element stiffness matrix K e

The element stiffness matrix K e includes two summands. The first one is constant within the element. It
will be denoted as K 0 . The integrand in the second term depends on ξ and η. Carrying out the integration
yields the so-called hourglass part K hg of the element stiffness matrix K e . We obtain

K e = K 0 + K hg (2.1)

with Z
K0 = 4 B T0 Ĉ B 0 det J 0 t and K hg = B Thg Ĉ B hg det J 0 dξ dη t. (2.2)


2.2.1 Representation of K 0 and K hg

For the following analysis, we introduce the 8x8 matrix

 
T r̄ b̄x b̄y γ̄ o4 o4 o4 o4
M = . (2.3)
o4 o4 o4 o4 r̄ b̄x b̄y γ̄

As in Section 1.6.4, a bar over a vector v indicates that the vector has been normalized to |v| = 1. So,
in addition to r̄ = 2 r, we have b̄x = bx /|bx |, b̄y = by /|by | and γ̄ = γ/|γ|. At first we look at the part

38
 Study of convergence (shell)
6

displacement in point A [m]

5

2 Q1SP
Q1/E12
1 Q1

0
0 10 20 30 40 50 60 70
number of elements in circumferential direction [-]

Figure 2.5: Study of convergence for element formulations Q1SP, Q1/E12 and Q1. The formulation Q1
is the standard displacement approach described in Chapter 1. It exhibits locking which will be discussed
in this chapter. Q1SP and Q1/E12 are more sophisticated element formulations which are locking-free
in this example.

K 0 . This part includes the matrix B 0 which can be alternatively to (1.56) expressed by the product

B 0 = L0 M , (2.4)

where the 3x8 matrix L0 is defined by

 
0 |bx | 0 0 0 0 0 0
L0 =  0 0 0 0 0 0 |by | 0  . (2.5)
0 0 |by | 0 0 |bx | 0 0

It should be mentioned that bx and by have the dimension (length)−1 , because B 0 is a mapping which
transfers a displacement (quantity with dimension length) into a strain (dimensionless quantity). Here,
the matrix L0 includes the magnitudes of bx and by which carry the dimension of bx and by , respectively.
As such, also L0 has the dimension (length)−1 . The matrix M is dimensionless. In a further step, the
elasticity matrix Ê is rewritten to read
 
r+2 r 0
Ê = µ  r r + 2 0 , (2.6)
0 0 1

where the material parameter ratio r = Λ/µ has been introduced. We are now in the position to describe
K 0 by means of the relation
 
T k
K0 = M LT0 Ê L0 M
µ

39
  
0 0 0 0 0 0 0 0

 0 b2x (r + 2) 0 0 0 0 bx by r 0 


 0 0 b2y 0 0 bx by 0 0 


T  0 0 0 0 0 0 0 0 
= M   k M. (2.7)
 0 0 0 0 0 0 0 0 

2

 0 0 by by 0 0 bx 0 0 

 0 by by r 0 0 0 0 b2y (r + 2) 0 
0 0 0 0 0 0 0 0
| {z }
K
c0

The quantity k = 4 µ det J 0 t is a short hand notation. Further, we have abbreviated the magnitude
|bx | of the vector bx by bx . The quantity by is defined analogously. The matrix K
c0 is important for the
eigenvalue analysis. It will be used in Section 2.2.2.

In a next step we have to reformulate K hg . For this purpose we need to compute the matrix B hg which
can be – alternatively to (1.56) – given by

B hg = j 0 L4x8
hg M (2.8)

The matrix j was introduced in Section 1.5. The index 0 indicates that we evaluate it here in the centre
of the element described by ξ = o2 . In contrast to the 4x2 matrix Lhg defined in (1.56), the matrix L4x8
hg
has the dimension 4x8 with the format
 
4x8 o2 o2 o2 lhg /2 o2 o2 o2 o2
Lhg = . (2.9)
o2 o2 o2 o2 o2 o2 o2 lhg /2

Using the latter relation in the equation for K hg leads to

Z 
T 4x8 T T 4x8
K hg = M (Lhg ) j 0 Ê j 0 Lhg det J 0 dξ dη t M

| {z }
K
chg

0 0 0 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 K
b 44 0 0 0 K
b 48 
= M   M. (2.10)
 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 K
b 48 0 0 0 K
b 88
| {z }
K
chg

40
The coefficients Kb 44 , K
b 48 and K
b 88 are complicated expressions. They are not given here. In Section
2.2.2, they will be computed for an element of rectangular shape. For now, only the structure of K
chg is
of interest.

We are now able to describe the element stiffness matrix by means of

K e = M T (K
c0 + K
chg ) M , (2.11)

where both K
c0 and K
chg have a very sparse structure. Let us introduce the notation K
c=K
c0 + K
chg for
the sum of K
c0 and K
chg .

2.2.2 Intelligent formulation of the eigenvalue problem for K e

Preliminary mathematical consideration. In order to understand the following, it is helpful to make

a small excursion to mathematics. We introduce the symmetric N xN matrix H = RT H c R, where R
is an orthogonal matrix. We are interested in the eigenvalues and eigenvectors of H and thus formulate
the eigenvalue problem
(RT H
c R − ωJ I N ) ϕJ = oN . (2.12)

Here, ωJ are the eigenvalues and ϕJ the eigenvectors of H = RT H

c R. Please note that there is no
summation over J. The index J runs from 1 to N , where N is the dimension of H, H,
c R and the
identity matrix I N . Now, we exploit the orthogonality of R expressed by RT R = I N . By multiplying
with R from the left, the eigenvalue problem (2.12) can be rewritten to read

R (RT H
c − ωJ RT ) R ϕJ = (H
c − ωJ I N ) R ϕJ = oN . (2.13)

There are two important conclusions from this short analysis. First of all, H and H
c have the same
eigenvalues. Secondly, the eigenvectors ϕJ of H can be computed by determining the eigenvectors
ϕ c and multiplying them by RT from the left, since ϕJ = RT ϕ
b J = R ϕJ of H b J holds.

Why is this knowledge important? This is because H c might have a much simpler structure than H. It
might be possible to compute eigenvalues and eigenvectors of H
c still analytically. Using the transactions
just described, it is then also possible to compute the eigenvalues and eigenvectors of H, although this
matrix might be fully occupied. 

One can imagine that this situation just arises in the present context. The element stiffness matrix K e =
MTK c M is fully occupied. The matrix K c has a very sparse structure which might be simple enough

41
to carry out an analytical solution of its eigenvalue problem. Nevertheless, we have still the problem that
M is not orthogonal in general.

Under which circumstances is M orthogonal? In fact, this is the case, when the element has rectangular
shape. Then b̄x reduces to ḡ ξ , b̄y to ḡ η and γ̄ to h̄. As already mentioned in Section 1.6.3, the four
vectors r̄, ḡ ξ , ḡ η and h̄ are orthogonal to each other which yields the orthogonality of the 8x8 matrix
M . So, the following eigenvalue analysis is restricted to rectangular element shapes. One could ask
whether this is not too restrictive. Actually, the restriction is no problem, because the locking is already
observed for rectangular element shapes. It is therefore even recommendable to study first such a simple
element geometry.

It is assumed for the remainder of Section 2.2 that the element has a rectangular shape with the side-
lengths 2 a and 2 b in x and y direction, respectively. The Jacobi matrix J 0 and its inverse J −1
0 then
read
a−1 0
   
a 0
J0 = , J −1 = , (2.14)
0 b 0
0 b−1
respectively. The matrix j 0 is given by

a−1 0
 
0 0
j0 =  0 0 0 b−1  . (2.15)
0 b−1 a−1 0

The matrix K
c reduces after several calculation steps (which can be still carried out by hand) to
 
0 0 0 0 0 0 0 0
 0 s−1 (r + 2) 0 0 0 0 r 0 
 
 0 0 s 0 0 1 0 0 
 
 0 0 0 Kb 44 1 0 0 0 0 
µt
Kc=  µ t, (2.16)
 
 0 0 0 0 0 0 0 0 
−1
0 0 1 0 0 s 0 0
 
 
 0 r 0 0 0 0 s (r + 2) 0
 

1
0 0 0 0 0 0 0 K88 µ t
b

where K
b 44 and K
b 88 are given by

b 44 = 1 µ t ((r + 2) s−1 + s),

K b 88 = 1 µ t (s−1 + (r + 2) s).
K (2.17)
3 3
The quantity s denotes the ratio a/b.

42
2.2.3 Solution of the eigenvalue problem for K e

It has been found out in Section 2.2.2 that the eigenvalues of K c are the same as the ones of K e . The
eigenvectors ϕb J (J = 1, ..., 8) of K
c are related to the ones (ϕJ ) of K e by the relation ϕ
b J = M ϕJ .

The eigenvectors ϕJ (J = 1, ..., 8) can be seen as possible nodal displacement vectors of an element. One
speaks of eigenmodes or also of secondary displacement modes. Eigenvectors can be only determined up
to a scalar factor. They will be normed to |ϕJ | = 1. So, not the quantity of these secondary displacements
matters but their quality (the mode in mathematical terms). Below, it will be seen that we obtain eight
eigenmodes: three rigid body movements (two rigid body translations and one rigid body rotation),
one mode for volumetric expansion, one mode for elongation/contraction (stretch mode), one mode for
shearing (shear mode) and two bending modes. For an isoparametric 2D element with four nodes, it is
well-known that these modes must be present.

Important are the eigenvalues associated with these modes. The eigenvalues can be seen as stiffnesses
(and also have the corresponding dimension force divided by length). It is interesting to investigate,
whether these stiffnesses are physically reasonable. After having computed the eigenvalues, we will
check by means of a comparison with a continuum mechanical analysis (see Section 2.3), whether this
is the case.

Preliminary mathematical consideration. It is instructive to look at a 8x8 matrix with the sparse
structure
 
• • 0 0 0 0 0 0

 • • 0 0 0 0 0 0  
 
0 0 • • 0 0 0 0 H 0 0 0
  sub 1
  c

0 0 • • 0 0 0 0   0 H sub 2 0 0
 c 
H= = , (2.18)
c 

 0 0 0 0 • • 0 0   0
 0 H sub 3
c 0 

 0 0 0 0 • • 0 0 
 0 0 0 H
c sub 4
 0 0 0 0 0 0 • • 
0 0 0 0 0 0 • •

where the symbol • indicates a component which is not zero. It is easy to identify four 2x2 submatrices
H
c sub K (K = 1, 2, 3, 4). The 8 eigenvalues of H
c are found by setting the determinant det (Hc − ωJ I 8 )
(J = 1, ..., 8) equal to zero. This determinant is alternatively presented by
  
det (H − ωJ I 8 ) = det (H sub 1 − ωM I 2 ) det (H sub 2 − ωM +2 I 2 )
c c c
  
c sub 3 − ωM +4 I 2 ) det (H
det (H c sub 4 − ωM +6 I 2 ) = 0. (2.19)

43
The index M runs from 1 to 2. Obvisously, the determinant det (H c − ωJ I 8 ) will be zero, if one of the
c sub K − ωM +G I 2 ) vanishes. The case differentiation
four (K = 1, ..., 4) factors det (H


 0 for K = 1
2 for K = 2

G= (2.20)

 4 for K = 3
6 for K = 4


is considered for G. This means that the eigenvalues ωM +G (G determined by (2.20) and M = 1, 2) can
be obtained by solving the 2x2 eigenvalue problems for H
c sub K with K = 1, ..., 4. Fortunately, it is still
possible to solve a 2x2 eigenvalue problem analytically. So, the computation of the eigenvalues of the
8x8 matrix Hc is well feasible, if the matrix has the sparse structure indicated in (2.18). 

The question is now, how this insight can be applied for the present situation, i.e. to the sparse matrix
K.
c This is actually no problem at all. It is certainly possible to change the order of the unknowns. Up
to now, we use the order prescribed by the order in the vector M U e , because K c is multiplied with this
vector. This is the standard order 1, 2, 3, 4, 5, 6, 7, 8 of unknowns. The matrix K
c has presently the sparse
structure  
K
b 11 0 0 0 K b 15 0 0 0
0 K b 22 0 0 0 0 K b 27 0
 
 
 

 0 0 K b 33 0 0 K b 36 0 0 

 0 0 0 K b 44 0 0 0 K b 48 
K=
c , (2.21)

 K
b 51 0 0 0 K b 55 0 0 0 

0 0 K b 63 0 0 K b 66 0 0
 
 
0 K b 72 0 0 0 0 K b 77 0
 
 
0 0 0 K b 84 0 0 0 K b 88

where the values K

b 11 , K
b 15 , K
b 51 , K
b 55 , K
b 48 and K
b 84 are also zero. This has not been written out, because
in this way, the structure of the matrix is recognizable in an easier way. Now we come to the important
point that we change the order of the unknowns to 1, 5, 2, 7, 3, 6, 4, 8 which leads to the matrix
 
Kb 11 Kb 15 0 0 0 0 0 0
 K51 K b 55 0 0 0 0 0 0 
 b 
 
 0 0 K b 22 Kb 27 0 0 0 0 
 
 0 0 K
b 72 Kb 77 0 0 0 0 
K
creord = 
 0
. (2.22)
 0 0 0 K b 33 K b 36 0 0  
 0 0 0 0 K63 K66 0 0 
 b b 
 0 0 0 0 0 0 K b 44 K
 b 48 

0 0 0 0 0 0 K84 K88
b b

The eigenvalues of K
creord are the same as the ones of K.
c But the reordering has to be considered in the
computation of the eigenvectors which will be discussed below. It is now clearly visible that the solution

44
to the eigenvalue problem for the matrix K creord can be found by solving the four smaller eigenvalue
problems for the four submatrices Kcsub K (K = 1, ..., 4). The four submatrices have the form
" #  
Kb 11 Kb 15 0 0 µt
K sub 1 =
c = ,
Kb 51 Kb 55 0 0 s
" #  
Kb 22 Kb 27 r+2 sr µt
K sub 2 =
c = 2 ,
Kb 72 Kb 77 s r s (r + 2) s
" #  2  (2.23)
Kb 33 Kb 36 s s µ t
K
csub 3 = = ,
Kb 63 Kb 66 s 1 s
" #  
Kb 44 Kb 48 r + 2 + s2 0 µt
K sub 4 =
c = 2 .
K84 K88
b b 0 1 + s (r + 2) 3 s

For simplicity, we abbreviate the factor µ t/s in the remainder of this section by bk. It should be noted
that the ratios r and s are dimensionless. The dimension of K csub K (K = 1, ..., 4) is carried by the
k with the dimension force divided by length. This is correct because the element stiffness matrix
factor b
Ke = M T K
c M must also have the dimension force divided by length. The orthogonal matrix M is
dimensionless.

Eigenvalue problem for K

csub 1 . The eigenvalue problem reads
 
−ω1,5 0
k
b φ1,5 = o2 (2.24)
0 −ω1,5
and yields the eigenvalues
ω1 = ω5 = 0 b
k. (2.25)
The letter φ with an index such as 1, 5 has been introduced to denote the eigenvectors of the four 2x2
eigenvalue problems. The indices 1, 5 refer to the original ordering (the one of K).
c Certainly, the
eigenvalue problem (2.24) is a very special eigenvalue problem, because the matrix is equal to the zero
matrix. Both eigenvalues are equal. Thus, we have here a so-called double eigenvalue. In this case,
the eigenvectors are not uniquely determined. The vectors φT1 = {1, 0} and φT5 = {0, 1} are possible
choices. Expanding back to the original 8x8 matrix (before reordering was carried out), we come to the
eigenvectors
b T1 = {1, 0, 0, 0, 0, 0, 0, 0}, ϕ
ϕ b T5 = {0, 0, 0, 0, 1, 0, 0, 0}. (2.26)
Using finally the relation ϕJ = M T ϕ
b J (see the beginning of Section 2.2.3), the eigenvectors

ϕT1 = {r̄ T , oT4 }, ϕT5 = {oT4 , r̄ T } (2.27)

are determined as two eigenvectors of K e . Please see the visualization of these two eigenmodes in Fig.
2.6. It is well-known that K e has three zero eigenvalues. Two refer to rigid body translations which are
visible in Fig. 2.6.

45
Figure 2.6: Eigenmodes ϕ1 and ϕ5 (rigid body translations). The original element shape is represented
by the solid line. The dot-dashed line indicates the eigenmode which can be visualized as a secondary
displacement U e = ϕ1,5 .

How are these eigenmodes visualized? To do that, one may recall that the eigenvectors ϕJ (J = 1, ..., 8)
play the role of secondary element nodal displacement vectors. As such, they can be interpreted as a
solution for U e . Thus, we write for instance U Te = ϕT1 = {1, 1, 1, 1, 0, 0, 0, 0}/2 and move the nodes
accordingly. In this example, we see a simple rigid body translation (see Fig. 2.6).

The third zero eigenvalue will be an eigenvalue of K

csub 3 (see below). It will be associated with a rigid
body rotation. The three zero eigenvalues are physically correct, because in the element stiffness matrix
no boundary conditions are considered. So, the element has the possibility to translate and to rotate. The
stiffness (eigenvalue) corresponding to these three modes is zero.

Eigenvalue problem for K

csub 2 . The eigenvalue problem for K
csub 2 reads
 
r + 2 − ω2,7 s r
k
b φ2,7 = o2 . (2.28)
sr s2 (r + 2) − ω2,7

The expressions for ω2 and ω7 are rather complicated for arbitrary s. Therefore we present these eigen-
values here only for s = 1:
ω2 = (2 r + 2) b
k, ω7 = 2 b
k. (2.29)
√ √
csub 2 are given for s = 1 by φT = {1, 1}/ 2 and φT = {−1, 1}/ 2. Following
The eigenvectors of K 2 7
the same procedure as for K sub 1 , we arrive at
c
√ √
b T2 = {0, 1, 0, 0, 0, 0, 1, 0}/ 2, ϕ
ϕ b T7 = {0, −1, 0, 0, 0, 0, 1, 0}/ 2 (2.30)

and
√ √
ϕT2 = {ḡ Tξ , ḡ Tη }/ 2, ϕT7 = {−ḡ Tξ , ḡ Tη }/ 2. (2.31)

These modes are plotted in Fig. 2.7. The eigenmode ϕ2 is a volumetric expansion, the eigenmode ϕ7
represents elongation/contraction (stretch). An important question is, whether the eigenvalues associated
with these modes are physically reasonable. The eigenvalue ω2 increases with r, thus approaches ∞ for
r → ∞ which characterizes an almost incompressible material. This is physically reasonable, because
the resistance of the material to withstand a volumetric expansion must go to ∞, when the material

46
becomes almost incompressible. So, this eigenvalue can become very large. However, this should not
be attributed to locking which is an artificial, non-physical stiffening phenomenon. The eigenvalue ω7 is
constant and therefore independent of the material parameter ratio r. Please see Section 2.3 for a further
discussion of this issue.

Figure 2.7: Eigenmodes ϕ2 and ϕ7 (volumetric expansion and stretch). The original element shape is
represented by the solid line. The dot-dashed line indicates the eigenmode which can be visualized as a
secondary displacement U e = ϕ2,7 .

Eigenvalue problem for K

csub 3 . The third eigenvalue problem to be formulated is
 2 
s − ω3,6 s
bk φ3,6 = o2 . (2.32)
s 1 − ω3,6
The results are
ω3 = 0 bk, ω6 = (1 + s2 ) b
k (2.33)
√ √
as well as φT3 = {−s−1 , 1}/ 1 + s−2 and φT6 = {s, 1}/ 1 + s2 . In the same way as before, we arrive
at
√ √
b T3 = {0, 0, −s−1 , 0, 0, 1, 0, 0}/ 1 + s−2 , ϕ
ϕ b T6 = {0, 0, s, 0, 0, 1, 0, 0}/ 1 + s2 (2.34)
and
√ √
ϕT3 = {−s−1 ḡ Tη , ḡ Tξ }/ 1 + s−2 , ϕT6 = {s ḡ Tη , ḡ Tξ }/ 1 + s2 . (2.35)
The modes are plotted in Fig. 2.8. It is easily recognizable that ϕ3 characterizes the rigid body rotation
discussed earlier. The corresponding eigenvalue ω3 is as expected zero. The eigenmode ϕ6 is a shear
mode. The associated eigenvalue is independent of the material parameter ratio r.

Having solved the first three 2x2 eigenvalue problems, we have determined eigenvalues and eigenvectors
of K
c0 . Up to this point, we have not discovered any source for artificial stiffening effects. The latter are
exhibited by unphysically high eigenvalues of the stiffness matrix K e . But it remains to investigate the
last term, the matrix K
chg which has diagonal structure. See further below.

Eigenvalue problem for K

csub 4 . The eigenvalue problem reads
" #
Kb 44 − ω4,8 0
0 b 88 − ω4,8 φ4,8 = o2 .
K
(2.36)

47
Figure 2.8: Eigenmodes ϕ3 and ϕ6 (rigid body rotation and shear). The original element shape is
represented by the solid line. The dot-dashed line indicates the eigenmode which can be visualized as a
displacement U e = ϕ3,6 .

The results are

b 44 = k (r + 2 + s2 ), ω8 = K
b 88 = k (1 + (r + 2) s2 )
b b
ω4 = K (2.37)
3 3
and φT4 = {1, 0} and φT8 = {0, 1}. This leads to

b T4 = {0, 0, 0, 1, 0, 0, 0, 0}, ϕ
ϕ b T8 = {0, 0, 0, 0, 0, 0, 0, 1} (2.38)

and
T T
ϕT4 = {h̄ , oT4 }, ϕT8 = {oT , h̄ }. (2.39)

The eigenmodes are plotted in Fig. 2.9.

Figure 2.9: Eigenmodes ϕ4 and ϕ8 (bending modes). The original element shape is represented by
the solid line. The dot-dashed line indicates the eigenmode which can be visualized as a displacement
U e = ϕ4,8 . Obviously, these are the hourglass modes.

Important conclusions for the eigenvalues of K

csub 4 . Evidently, the so-called hourglass eigenvalues ω4
and ω8 increase for constant s linearly with r. This is a clearly non-physical result because the hourglass
mode conserves the volume per definition. In the limit r → ∞ the hourglass eigenvalues should not be
significantly altered. Here, we detect the volumetric locking (see the example in Section 2.1.1).

Furthermore, increasing the aspect ratio s = a/b, the element shape becomes more and more beam-like.
The eigenvalue ω4 associated with the bending deformation displayed by the eigenmode ϕ4 should not
go to infinity with increasing s. Unfortunately, this is the case. The preceding remark holds analogously
for ω8 . This observation reveals another kind of locking which is the shear locking (see the example in
Section 2.1.2).

48
2.3 Comparison with analytical analysis at continuum mechanical
level

We have now detected problems of the element formulation by observing that two eigenvalues are too
high and lead to an artificially high stiffness of the element. This is the reason, why the element for-
mulation shows very bad convergence behaviour in examples such as the ones investigated in Section
2.1.

To remedy the element formulation, it is important to know the physically reasonable eigenvalues of an
element formulation of this class. They can be determined in the way explained in the following.

Let us use the strain energy for this purpose. In continuum mechanics, the strain energy per reference
volume (dimension energy divided by volume) is given by
1 T
ψ cont = ε̂ Ê ε̂. (2.40)
2
The strain energy for the domain of one element reads
Z
cont
ψe = ψ cont det J 0 dξ dη t, (2.41)


where we continue to assume that we deal here with rectangular elements.

This strain energy must be correctly displayed by means of our finite element formulation. How can this
be found out? The strain energy at element level is in general (for linear elasticity) represented by
1 T
ψeFEM = U K e U e. (2.42)
2 e
Let us use the spectral decomposition for K e which reads
8
X
Ke = ωJ ϕJ ϕTJ . (2.43)
J=1

Applying (2.43) in the equation (2.42) for the strain energy of the element leads to the quite simple
relation
8
1 X
ψeFEM = ωJ (U Te ϕJ )2 . (2.44)
2 J=1
The index J shows up three times, therefore the summation convention cannot be applied. Instead, the
summation over J is written out explicitly. The question to be answered is whether the eigenvalues
determined in Section 2.2.3 are such that ψeFEM = ψecont holds for the strain states of interest. Below,

49
we will investigate three strain states. We start with a strain state, where only normal strains are present
which are constant within the element (Section 2.3.1). The second strain state includes only the shear
strain (Section 2.3.2). Finally, in Section 2.3.3, we investigate the strain state which is obtained for
bending of a structure.

The instrument of spectral decomposition (2.43) might not be familiar to all readers. It is a standard
mathematical tool1 which we would like to demonstrate at the example of the matrix K
csub 3 .

Example for spectral decomposition. We repeat the results of Section 2.2.3 for K
csub 3 :

k, ω6 = (1 + s2 ) b
ω3 = 0 b k,
√ √
φT3 = {−s−1 , 1}/ 1 + s−2 , φT6 = {s, 1}/ 1 + s2

Since K
csub 3 is a 2x2 matrix, the spectral decomposition includes only two terms:

csub 3 = ω3 φ3 φT + ω6 φ6 φT .
K (2.45)
3 6

Additionally, the eigenvalue ω3 is zero. Therefore, we obtain from (2.45) only the second term which
leads with  
s2 s
K
csub 3 = (1 + s ) b
k2
(1 + s2 )−1 (2.46)
| {z } s 1
ω6 | {z }
φ6 φT6
again to the expression (2.233 ).

2.3.1 Homogeneous normal strain state

Let us start with a strain state, where the shear strain is zero. The normal strains are assumed to be
constant within the element. Thus, we have a homogeneous strain state with

ε̂T = {εxx , εyy , 0}. (2.47)

The strain energy per reference volume (2.40) can be represented by means of (2.6) as
1
ψ cont = µ (r + 2) (ε2xx + ε2yy ) + µ r εxx εyy . (2.48)
2
1
Sufficient information about the spectral decomposition can for example be found in Wikipedia under the notion “Eigen-
decomposition of a matrix”.

50
This yields for the strain energy of one element (rectangular shape with sidelengths 2 a, 2 b and thickness
t) the result 1 
ψecont = (r + 2) (ε2xx + ε2yy ) + r εxx εyy 4 µ a b t. (2.49)
2
The next step is to express such a homogeneous normal strain state by means of U e . If the element
undergoes only a strain εxx in x direction, i.e. εyy = 0, the nodes will move only in horizonal direction.
From the displacement state (U e )x = ḡ ξ , we would obtain the strain εxx = 1/(2 a). Prescribing εxx = 1
would lead to (U e )x = 2 a ḡ ξ . This consideration is analogously possible for the y direction. A general
homogeneous normal strain state (2.47) leads to

(U e )x = εxx 2 a ḡ ξ and (U e )y = εyy 2 b ḡ η . (2.50)

Please note again that εxx and εyy are constant within the element. We exploit (2.44) and see that
only the eigenvectors ϕ2 and ϕ7 (2.31) yield a contribution, because for the other eigenvectors (J =
1, 3, 4, 5, 6, 8) U Te ϕJ vanishes. Further, it is important to remember that ϕ2 and ϕ7 were computed
based on the assumption s = a/b = 1. Therefore, we insert b = a for this strain state. With
√ √
U Te ϕ2 = (εxx + εyy ) a 2 and U Te ϕ7 = (−εxx + εyy ) a 2 (2.51)

we obtain from (2.44) the result

k a2 (ε + ε )2 + 2 b
ψeFEM = (2 r + 2) b k a2 (−εxx + εyy )2
| {z } | xx{z yy } |{z} |
ω7
{z }
ω2 (U Te ϕ2 )2 /2 (U Te ϕ7 )2 /2
 
2 2 2
= k a 2 (r + 2) (εxx + εyy ) + 4 r εxx εyy
b (2.52)

Exploiting the special case s = 1 which leads to b

k = µ t in the latter equation and to b = a in (2.49),
it is now evident that the equality ψecont = ψeFEM is obtained. This means that the element formulation
presented in Chapter 1 is well suitable to display strain states with homogeneous normal strains. The
eigenvalues ω2 and ω7 are physically reasonable.

2.3.2 Homogeneous shear strain state

The next situation to be investigated is a homogeneous strain state, where the normal strains are zero:

ε̂T = {0, 0, 2 εxy }. (2.53)

The energy per reference volume (2.40) is now given by ψ cont = 2 µ ε2xy . This yields for the strain energy
of one element (rectangular shape with sidelengths 2 a, 2 b and thickness t) the result

ψecont = 8 µ a b t ε2xy . (2.54)

51
Simple shear in x direction can be expressed by (U e )x = px ḡ η . For simple shear in y direction, we can
write analogously (U e )y = py ḡ ξ . See the illustration in Fig. 2.10. If we superpose these two simple
shear strain states, we will come to a shear strain of
1 1 px py 1
εxy = (ux,y + uy,x ) = ( + )= (px s + py ). (2.55)
2 2 2b 2a 4a
Many choices of px and py are possible to generate the same εxy . Let us leave this point open and
continue to work with px and py . It should be also noted that the assumption s = 1 is no longer valid.
We work with general aspect ratios s = a/b.

Figure 2.10: Simple shear states in x and y direction.

The next step is to evaluate (2.44). Only the eigenvectors ϕ3 and ϕ6 yield a contribution, because for
the other eigenvectors (J = 1, 2, 4, 5, 7, 8) U Te ϕJ = 0 holds. Using (2.35), we obtain
−px s−1 + py px s + py 4 a εxy
U Te ϕ3 = √ , U Te ϕ6 = √ =√ , (2.56)
1+s −2 1+s 2 1 + s2
where interestingly in the expression for U Te ϕ6 the shear strain εxy can be identified. The expression for
U Te ϕ3 is here mentioned only for completeness. Since the corresponding eigenvalue ω3 is zero, it will
not yield any contribution to ψeFEM . The energy ψeFEM finally reads

1 16 a2 ε2xy
ψeFEM = (1 + s2 ) b
k = 8 µ a b t ε2xy . (2.57)
| {z } 2 1 + s2
ω6 | {z }
(U Te ϕ6 )2 /2
The result is identical to (2.54). Thus, the element formulation presented in Chapter 1 yields for a
homogeneous shear strain state a physically reasonable eigenvalue ω6 .

2.3.3 Strain state for pure bending

Let us now investigate the situation of pure bending (see Fig. 2.11) characterized by the strain state
η
ε̂ = {− px , •, 0}T . (2.58)
a
52
The strain εyy is still unknown (indicated here by the symbol •). Exploiting the knowledge that the
normal stress in the y direction is zero in combination with the elasticity law σ̂ = Ê ε̂ and (2.6), we
obtain εyy by the relation
r r η
σyy = µ r εxx + µ (r + 2) εyy = 0 ⇒ εyy = − εxx = px . (2.59)
r+2 r+2 a
The corresponding normal stress σxx is given by
r r+1
σxx = µ (r + 2) εxx + µ r (− εxx ) = 4 µ εxx . (2.60)
r+2 r+2
The pure bending strain state is no longer homogeneous. At the bottom of the element, εxx is positive
and εyy negative. At the top, the situation is just in the other way round. There is no shear strain. Such
a strain state could be caused by a normal stress σxx which also varies linearly over η (see Equation
(2.60)).

The next step is to compute (2.40) and then (2.41). In fact, ψ cont can be represented simply by (2.48) in
Section 2.3.1, certainly then with different entries for εxx and εyy . Using (2.58) and (2.59) and subse-
quently carrying out the integration over , one arrives at
8 µt r + 1 2
ψecont = p . (2.61)
3 s r+2 x
This expression remains finite for large aspect ratios s as well as extreme material parameter ratios r,
whereas it takes on very large values for extreme aspect ratios s−1 . All three results are physically
reasonable. For the latter it has to be considered that the bending of an infinitely high structure (s−1 →
∞) requires infinitely large strain energy.

How does that compare to the result of the finite element formulation presented in Chapter 1? For this
purpose, we first need to formulate the element nodal displacement vector U e for the pure bending state.
It is not surprising that we see now the bending (or hourglass) mode involved (see also Fig. 2.11):
T
U Te = {−px , px , −px , px , 0, 0, 0, 0} = 2 px {−h̄ , oT4 }. (2.62)

Again, we seek to exploit (2.44). The factor U Te ϕ4 = −2 px is the only non-zero one among the eight.
Thus, we need to consider only one contribution in ψeFEM :

k
b 2 µt
ψeFEM = (r + 2 + s2 ) 2 p2x = (r + 2 + s2 ) p2x . (2.63)
|3 {z } |{z} 3 s
ω4 (U Te ϕ4 )2 /2

First of all, this result deviates from ψecont . Secondly, it is obvious that we see here the already described
artificial stiffening, i.e. locking. The strain energy goes to infinity for r → ∞ and s → ∞ which is a

53
 Figure 2.11: Pure bending state.

clearly non-physical result. This is all the more clear by the comparison with ψecont which yields a quite
different result.

Another observation should be mentioned here. It will be further explained in Chapter 3. In the analytical
solution, εyy depends linearly on η. Such a dependence is not included in the finite element formula-
tion worked out up to this point. Further, the shear strain has to vanish. In the context of the present
formulation, due to the simplicity of the interpolation, this is only possible, if also the displacement
vanishes. These effects are well-known under the notions Poisson thickness locking and shear locking,
respectively.

The analysis has been done here for the dominating bending deformation in x direction. One could carry
out the derivation analogously for the y direction. Then the eigenvalue ω8 would come into play. The
result would be, however, just the same.

This gives us now the main incentive of the field of finite element technology which is to eliminate the
undesirable locking phenomena by a more intelligent choice of the interpolation functions. However,
we wish to keep the simplicity of the element, i.e. we still strive to continue to work with isoparametric
bilinear element formulations. How this can be achieved, will be discussed in Chapter 3.

54
Chapter 3

Methods against locking

3.1 General remarks

The eigenvalue analysis discussed in Section 2.2 and the comparison with continuum mechanical results
in Section 2.3 have shown that among the eight eigenvalues of the element stiffness matrix K e , only the
so-called hourglass eigenvalues ω4 and ω8 do not come out to be physically reasonable. In other words,
the eigenvalues of the part K 0 of the element stiffness matrix K e agree with continuum mechanical
findings, whereas the eigenvalues of the part K hg do not. So, an obvious conclusion is to change K hg .

The presently used interpolation functions NI = NI (ξ, η) = (1 + ξ ξI + η ηI + ξ η ξI ηI )/4 with I =

1, ..., 4 are complete up to an order of one. This means that all polynomials ξ k η l with integer exponents
k, l and k + l ≤ 1 are represented. The completeness up to the order two (k + l ≤ 2) is, however, not
fulfilled. See e.g. [ZTZ05] for more information about this issue. Let us denote the order of completeness
by m. If completeness of NI up to order m is guaranteed, “a constant value of the mth derivative” will
be “attainable in the element domain ... when the size of any element tends to zero” [ZTZ05]. So, the
order m = 1 is sufficient to obtain constant values of ui,j (i, j = 1, 2). In other words, in the limit of an
infinitely dense mesh, a constant strain state can be displayed which would be sufficient to display the
physics of a problem correctly.

This is an important result and makes clear that a first approach to get rid of the locking would be to
work with NI = (1 + ξ ξI + η ηI )/4. Such a choice results into B hg = 0 and, finally, to K hg = 0.
Consequently, the hourglass eigenvalues ω4 and ω8 vanish, and the stiffness associated with the bending
modes shown in Fig. 2.9 is no longer overestimated. However, this approach brings at least two major

55
disadvantages. First of all, to work with zero or an underestimated stiffness for the bending modes
activates hourglass-like deformation patterns, see Fig. 3.1. This is an undesirable result. Secondly,
since the element stiffness matrix has too many zero eigenvalues, the invertibility of the assembled
global stiffness matrix K can no longer be guaranteed. If K is singular, the final equation system
K U = P will not be solvable. Whether this happens, depends on the concrete boundary conditions of
the mechanical system under investigation.

Figure 3.1: Hourglass-like deformation pattern in a deformed structure. The system was already inves-
tigated in Section 2.1.1. Fig. 2.2a shows the correct result obtained for an irregular mesh. Here, one
sees the result obtained for a regular mesh by means of a finite element formulation with underestimated
hourglass eigenvalues.

Reduced integration. The above-mentioned approach is often termed “reduced integration” because
the cancellation of K hg is easily obtained by evaluating the integral in (1.80) only in the centre of the
element. This is then a one Gauss point integration. In other words, compared to the standard choice of
two by two Gauss points, one uses a reduced number of Gauss points.
1
0 11
00 1
0 1
0
0
1 00
11 0
1 0
1
0
1 00
11 0
1 0
1

11
00
00
11 11
00
00
11 11
00
00
11 1
0
0
1
00
11 00
11 00
11 0
1

Figure 3.2: Schematic visualization of incompatible modes at the neighboured edges of two elements.

Method of incompatible modes. Another approach is the method of incompatible modes. Instead of
omitting the bilinear part in the shape function vector it is also possible to enrich the interpolation by the
quadratic terms ξ 2 and η 2 . These so-called bubble modes (see Fig. 3.2) are constructed in such a way
that the additional displacements are zero at the nodes but non-zero between the nodes. This renders
the interpolation in some way complete up to order two and makes the element softer. For instance,
the element is then able to exhibit curved edges during the deformation. However, if one proceeds in
this way, one will not be able to require that the edges of one element coincide with the ones of the

56
neighbouring element. The additional “enhanced” deformation is not compatible.

Summary. In summary, it is very important that the hourglass eigenvalues are correctly computed.
If they are overestimated, locking will be observed, i.e. the stress will be over- and the deformation
far underestimated (in particular in the limit of incompressibility and for extreme aspect ratios). The
element will turn out to be too stiff. If the hourglass eigenvalues are too small (e.g. zero), one will
observe hourglass-like deformation patterns and the stress will be underestimated. The element will
behave too softly.

The methods of reduced integration and incompatible modes were derived in the 19seventies, but do
not represent the last conclusion to eliminate locking. Nevertheless, they can be seen as the ground
for further developments in finite element technology which finally lead to a very efficient, robust and
locking-free finite element formulation. This will be discussed in the remainder of Chapter 3.

3.2 Mixed variational formulation

3.2.1 Repetition of the strong form

A very important step is to find a variational formulation which allows us to include additional terms in
the interpolation. For this purpose we go back to Section 1.1 and repeat the basic equations (1.15) in the
format
kinematics : ε − εc = 0,
Hooke’s law : σ − σ ε = 0, (3.1)
equilibrium : div σ + f = 0,

where the short hand notations εc = sym (grad u) and σ ε = 2 µ ε + Λ (tr ε) I have been introduced.
Note that the Voigt notation of σ ε is given by σ̂ ε = Ê ε̂.

Before, in Section 1.1, the equations (3.11 ) and (3.12 ) were inserted into (3.13 ). In Section 1.2, we derived
the corresponding weak form (1.30) which incorporates the displacement vector u as the only unknown
field. The equation (1.30) is therefore also called “one-field variational functional”. The corresponding
finite element method discussed in Chapter 1 yields the global displacement vector U as the solution,
from which the element nodal displacement vector U e and the nodal contributions uI with I = 1, ..., 4
are extracted. Using the interpolation (1.34) with (1.33) provides the displacement vector u and then,

57
by applying (1.51) and (3.12 ) the strain vector ε̂ and the stress vector σ̂, respectively. This procedure
implies that the equations (3.11 ) and (3.12 ) are fulfilled pointwise, i.e. in every point of the element.

In order to arrive at a new finite element formulation, we proceed differently in the way that every
equation in (3.1) is treated in the same way (see the next subsection).

3.2.2 Construction of alternative weak form – three-field variational functional

So, the idea is to multiply each equation in (3.1) by an appropriate test function. Please note that for
(3.13 ), this has already been accomplished in Section 1.2. The derivation laid down in Section 1.2 can
be directly taken over. Thereby, it has to be taken into account that, first of all, δε̂c cannot be replaced
by δε̂ and, secondly, σ̂ cannot be replaced by σ̂ ε = Ê ε̂, because (3.11 ) and (3.12 ) are not inserted into
(3.13 ). The weak form of (3.1) reads
Z
δ σ̂ T (ε̂ − ε̂c ) dV = 0, (3.2)
Z B

δε̂T (σ̂ − σ̂ ε ) dV = 0, (3.3)

Z Z B Z
T T
δε̂c σ̂ dV − δu f dV − δuT t̄ dA = 0. (3.4)
B B ∂Bt

Important statements. It should be emphasized that in contrast to the one-field variational functional
derived in Section 1.2, the equations listed in (3.2) to (3.4) include three independent unknown fields.
These are the displacement vector u, the strain vector ε̂ and the stress vector σ̂. Please note again that
ε̂c and σ̂ ε are not independent, since they depend on u and ε̂, respectively. Before introducing any
discretization, the three-field variational functional (see the equations (3.2) to (3.4)) is equivalent to the
one-field variational functional (1.30).

3.2.3 Concept for a possible discretization – reduction to two-field variational

functional

The fact that all three equations in (3.1) have been transferred into weak form offers many possibilities,
when we think about discretization. It has already been stated that the bilinear interpolation of u has
many advantages. One crucial advantage is that the resulting element formulation has only four nodes
for the evaluation of the displacements. So, it would be favourable to keep the interpolation (1.34).

58
Further, we would like to integrate the terms ξ 2 and η 2 in some way into the interpolation. The method
of incompatible modes was mentioned (see [TBW76]) which has its merits. How can the bubble modes
on the edges of the element included in a variationally consistent manner? The answer to this question
is given in the pioneering work of [SR90]. We follow this line and define the so-called enhanced strain
vector
ε̂e = ε̂ − ε̂c (3.5)

which can be interpreted as the strain induced by the incompatible displacements schematically visual-
ized in Fig. 3.2. Let us consider Equation (3.2) which is, using (3.5) and the discretization of B into ne
element domains Be , rewritten as
ne Z
X
δ σ̂ T ε̂e det J dξ dη t = 0. (3.6)
e=1 

It should be emphasized that the summation index e which is the index for the elements shall not be
confused with the index e in ε̂e , where the letter e refers to the enhanced strain. At this point, we make
an assumption and choose σ̂ to be constant within each element. Remember that σ̂ is an independent
field quantity. Therefore, we have to define a discretization of it. This is in contrast to the stress σ̂ ε which
is a function of ε̂ and, therefore, will not be independently discretized. Using σ̂ = const. and then also
δ σ̂ = const. on Be , the relation (3.6) can be transferred into
Z
T
δ σ̂ ε̂e det J dξ dη t = 0. (3.7)


Thus, to fulfill (3.7), we have to choose the interpolation of ε̂e det J in such a way that its integral over 
vanishes. We will show in the next section that a suitable interpolation can be found. One might already
have a look at the relation (3.18) and its derivation.

It is elucidating to think about the physical meaning of (3.7). By introducing the enhanced strain ε̂e
which violates compatibility between the edges of the elements, additional strain energy would be put
into the system if no further care was taken. Requiring
Z
(δ)σ̂ T (δ)ε̂e det J dξ dη t = 0 (3.8)


ensures that this does not happen. Please note that this so-called orthogonality condition holds for the
variations of σ̂ and ε̂e in the same way. Therefore, the symbol δ has been put into brackets.

With this knowledge, we go to the next equation (3.3) and reformulate it to read
Z Z
((δε̂c + δε̂e ) σ̂ − δε̂ σ̂ ε ) dV = (δε̂Tc σ̂ − δε̂T σ̂ ε ) dV = 0,
T T
(3.9)
B B

59
where (3.8) has been exploited. Finally, one takes (3.4) and subtracts the latter relation (3.9). This leads
to the statement
Z Z Z
T T T T
(δε̂c σ̂ − δε̂c σ̂ +(δε̂c + δε̂e ) σ̂ ε ) dV − δu f dV − δuT t̄ dA = 0. (3.10)
B | {z } B ∂Bt
=0
Interestingly, the independent field quantity σ̂ cancels out of the formulation. Before, it was stated that
we choose it to be constant within one element. This still holds, but σ̂ will never be computed. The role
of the stress σ̂ is taken over by σ̂ ε = Ê (ε̂c + ε̂e ) which depends not only on the so-called compatible
strain ε̂c but also on the enhanced strain ε̂e . The discretization of (δ)ε̂c is given by the one of (δ)u.
The discretization of (δ)ε̂e was just discussed. The two parts ε̂c and ε̂e of the total strain ε̂, as well as
their variations δε̂c and δε̂e , are independent of each other. Therefore, Equation (3.10) has to hold for
independent variations δε̂c and δε̂e . We arrive at the two statements
Z Z Z
g1 = δε̂Tc σ̂ ε dV − T
δu f dV − δuT t̄ dA = 0,
ZB B ∂Bt
(3.11)
g2 = δε̂Te σ̂ ε dV = 0.
B

The first equation resembles the original one-field variational functional (1.30). The difference is that the
stress σ̂ ε also depends on the enhanced strain ε̂e . To determine it, we have the second equation available.
As already said, the stress σ̂ has dropped out as an independent quantity. Thus, the original three-field
variational functional derived in Section 3.2.2 has reduced to the two-field variational functional (3.11)
with the independent field quantities u and ε̂e .

Remark. In the presentation chosen for these notes, we abstain from associating the field quantities
with their mathematical spaces. In the opinion of the author, this information is not absolutely needed
to develop a basic understanding of finite element technology. Readers who are interested in a more
mathematically-oriented representation, are referred to e.g. [SR90], [RS95] or [KT00].

3.3 Interpolation of the enhanced strain

The derivation in Section 3.2 has shown that the interpolation of the enhanced strain vector ε̂e has to ful-
fill the condition (3.7). In order to come to a suitable interpolation, let us first deal with the incompatible
displacements which can be expressed by
" #
P T oT2
 
wx
w = w(e) = = W e. (3.12)
wy oT2 P T

60
The analogy with (1.38) is obvious. It remains to define the 2x1 vector P and the 4x1 vector W e given
by
1 2
PT = {ξ − 1, η 2 − 1} and W Te = {(W e )Tx , (W e )Ty }, (3.13)
2
where the 2x1 vectors (W e )x and (W e )y contain the parameters of the functions
1 2 1
wx = (ξ − 1) (We )x 1 + (η 2 − 1) (We )x 2 ,
2 2
1 2 1 2
wy = (ξ − 1) (We )y 1 + (η − 1) (We )y 2 (3.14)
2 2
for the incompatible displacements wx and wy in x and y direction, respectively. As desired, wx and wy
vanish at the nodes I = 1, ..., 4. So, in addition to the eight nodal displacements included in U e , we
have four additional degrees-of-freedom in W e . In the context of the assembly procedure (see Section
1.8), we consider the fact that there are continuity requirements for the nodal displacements, because
one node usually belongs to several elements. Such continuity requirements do not exist for W e , since
the incompatible displacements are defined separately for each element.

The next task is to formulate the derivative

 

wx,x
 P T,x oT2
 
 T 
T 
ε̂e = wy,y = o P ,y  W e = G̃ W e , (3.15)
  2
wx,y + wy,x

P T,y P T,x

where the 3x4 matrix G̃ takes the same role for ε̂e as is taken by B for ε̂c (see Section 1.5). Using

∂P T ∂P T
     
∂ξ ∂η ξ 0 ∂ξ ∂η ξ 0
= , = , (3.16)
∂x ∂x ∂x 0 η ∂y ∂y ∂y 0 η
| {z } | {z }
T
= cx = cTy

the matrix G̃ is represented by

   ξ 0 0 0

cTx oT2
  0 η 0 0 
G̃ =  oT2 cTy   , (3.17)

 0 0 ξ 0 
cTy cTx
| {z } | 0 0{z0 η
}
=j = Le
where the 4x4 matrix Le has been introduced. Note that the 2x1 vectors cx and cy as well as the 3x4
matrix j were already introduced in Section 1.5.

Inserting ε̂e = j Le W e into (3.7) shows that the integral vanishes only for parallelogram shaped ele-
ments. In this case, j and det J reduce to j 0 and det J 0 , respectively, and ε̂ becomes with ε̂e = j 0 Le a

61
linear function of ξ and η. The integral of ε̂e over  vanishes. Otherwise, the matrix G̃ shows through j
a more complex dependence of ξ and η. The orthogonality condition (3.7) should be, however, fulfilled
for arbitrary element geometries. Therefore we modify the interpolation of ε̂e according to
det J 0
ε̂e = G W e with G = j 0 Le . (3.18)
det J
Remark. In fact, the latter modification means that, in general, εe can no longer be represented by the
relation εe = sym (grad w). This is no problem, since in the mixed variational formulation, presented in
Section 3.2, this statement was not used. The strain ε̂e is considered as an independent variable directly.
The fact that ε̂e could be seen as the gradient of w was helpful to find a suitable interpolation, in this
case, the bubble functions included in P . This step is, however, not necessary, and it is well possible to
state another interpolation of ε̂e which is not related to any incompatible displacement. 

It is very important to understand that the present choice of the interpolations for σ̂, ε̂e and u is one
of many possibilities. Nevertheless, this formulation [SR90] is well-established and has been proven
to be a very good choice. It is also called “enhanced strain method” and will be abbreviated by Q1/E4
from now on. The short hand notation E4 stands for the four additional (enhanced) degrees-of-freedom
contained in W e . The enhanced strain method has also been extended to three dimensions. One of the
first versions (see [SA92]) is called Q1/E12 (for twelve enhanced degrees-of-freedom in the 3D case), an
improved version is called QM1/E12 [SAT93]. Please see Section 2.1 for the use of Q1/E12 in the shell
example. The abbreviation Q1SP used in the heading of the next section will be introduced in Section
3.4.4.

3.4 Element stiffness matrix K e in the context of Q1/E4 and Q1SP

3.4.1 General format

P4
Inserting u = I=1 NI uI with (1.33), ε̂c = B U e with (1.56) and the just discussed interpolation
(3.18) for ε̂e into the two equations of the two-field variational functional (3.11) yields
X ne Z Z 
T T
g1 = δU e B Ê B det J dξ dη t U e + B T Ê G det J dξ dη t W e − P e = 0,
e=1 | {z } | {z }
K uu K uw
X ne Z Z 
T T T
g2 = δW e G Ê B det J dξ dη t U e + G Ê G det J dξ dη t W e = 0. (3.19)
e=1 | {z } | {z }
K wu K ww

62
The relation (3.192 ) must be fulfilled elementwise because – as already discussed in Section 3.3 – there is
no continuity requirement for ε̂e . In other words, the additional degrees-of-freedom stored in W e do not
have any reference to the additional degrees-of-freedom of another element. Taking this into account,
we can derive from (3.192 ) the relation

W e = −K −1
ww K wu U e . (3.20)

This means that the vector W e can be expressed in terms of U e and does not have to be determined
explicitly. The procedure to eliminate part of the unknowns at the element level is called static conden-
sation. The latter expression is inserted into (3.191 ) which leads to the relation
ne
X  
g1 = δU Te (K uu − K uw K −1 K wu ) U e − P e = 0. (3.21)
| {z ww }
e=1
Ke

We see that the element stiffness matrix K e = K uu of the standard linear isoparametric element for-
mulation (Q1) discussed in Chapter 1 is reduced by the part K uw K −1
ww K wu (note that K uw = (K wu )
T

holds). This will change in any case the eigenvalues of K e . Whether this cures the locking, will be
discussed later.

3.4.2 Representation of K e by means of the reduced B matrix

It is inspiring to express K e by means of a so-called reduced B matrix which reads

B̃ red = B − G K −1 −1
ww K wu = B 0 + B hg − j 0 Le K ww K wu d. (3.22)

For simplicity, we have introduced the short hand notation d = det J 0 /det J . It can be easily verified
that the reduced B matrix B̃ red enables the following alternative representation of the element stiffness
matrix: Z
T
Ke = B̃ red Ê B̃ red det J dξ dη t. (3.23)

It is also interesting to investigate the expression

B̃ red U e = B U e − G K −1
ww K wu U e = ε̂c + ε̂e = ε̂ (3.24)

which turns out to just yield the total strain vector. Here, the relation (3.20) in combination with the
interpolation ε̂e = G W e which leads to ε̂e = −G K −1
ww K wu U e has been used.

63
3.4.3 Evaluation of K e by means of standard numerical integration

There is not much more to say than in Chapter 1. Since the matrix j is, in general, a complex non-linear
function of ξ and η, the integrals embedded in the matrices K uu , K uw and K ww cannot be evaluated
analytically. Standard Gauss point integration is usually applied. No further simplifications are possible.

3.4.4 Evaluation of K e by means of analytical integration

In Section 1.7.2, we followed the idea to replace J by J 0 (and then also j by j 0 ), independently of
the actual shape of the element. This replacement can be done because the error (1.83) decreases with
mesh refinement. Even for coarse meshes, similar solutions are obtained (see [RKR99], [RWR00]). By
doing this, we obtain another element formulation which we term Q1SP. For elements with parallelogram
shape, Q1/E4 and Q1SP yield identical results.

The matrix B̃ red simplies with J = J 0 to

e red = B 0 + j 0 (Lhg − Le K −1 K
B fwu ) M hg . (3.25)
ww

Please note that K wu = K

fwu M hg holds. The matrices K ww and K wu are now computed by means of
J 0 . Lhg and Le are linear functions of ξ and η which vanish for ξ = 0 or η = 0. Therefore, the integral
in Z
Ke = e T Ê B
B e red det J 0 dξ dη t (3.26)
red

finally contains – as before in Section 1.7.2 – only two terms:
Z
Ke = K0 + B Tred Ê B red det J 0 dξ dη t = K 0 + K stab . (3.27)


It should be emphasized that

e red − B 0 = j 0 (Lhg − Le K −1 K
B red = B fwu ) M hg (3.28)
ww

includes only the terms of B

e red which are linear in ξ and η. The part B 0 is considered in K 0 . In the
context of the formulation Q1 (Chapter 1), the element stiffness matrix K e incorporated besides K 0
the summand K hg which was formulated by means of B hg . In the new formulation Q1SP derived in
the present chapter, this second summand is different. Instead of using B hg , we insert B red and obtain
K stab .

64
The role of K stab should be explained in more detail. At the beginning of Chapter 3, we discussed
the idea of simply setting K hg equal to the zero matrix. Such a procedure would lead to K e = K 0 and
two additional zero eigenvalues of K e . Consequences of that are hourglass-like deformation patterns and
eventually non-physical zero eigenvalues of the global stiffness matrix K. By working with K e = K 0 +
K stab , this situation is completely cured. We used the method of incompatible modes as motivation to
introduce the enhanced strain and finally obtained two new element formulations. The first one (Q1/E4)
does not make use of the simplification J = J 0 . Therefore, in general, the element stiffness matrix for
Q1/E4 cannot be represented in the format K 0 + K stab . For the second element formulation Q1SP, this
is always possible.

One can interpret K stab also as a so-called hourglass stabilization matrix which prevents K e from ha-
ving non-physical zero eigenvalues. This is the reunion of two big branches of finite element technology.
On the one hand, we see reduced integration methods, where various choices for the hourglass stabiliza-
tion matrix K stab are made. On the other hand, there are the mixed methods, where interpolations for
additional field quantities – such as the enhanced strain vector – come into play. The present hourglass
stabilization matrix K stab does not drop from the sky but is consistently derived on the basis of a well
established mixed method, here the enhanced strain formulation Q1/E4.

Again, the method based on replacing J by J 0 is called Q1SP [RKR99]. The letter S stands for sta-
bilization, the letter P for (equivalent) parallelogram. Since J is everywhere replaced by J 0 , one can
imagine that every element of general shape is – certainly only virtually – replaced by its equivalent
parallelogram. Due to the fact that the error introduced by making this approximation approaches zero
with increasing mesh density [AR95a, AR95b], Q1SP converges to the same result as Q1/E4. This can
be mathematically proven for the presently investigated case of linear elasticity.

Remark. In the context of geometrical and material non-linearity, however, Q1/E4 is known to show
numerical instabilities for certain loading scenarios ([WR96], [DPH+95], [Arm00]). Q1SP includes
further amendments to cure this problem (see e.g. [RW00]). So there are cases where Q1/E4 fails to
solve the problem, whereas Q1SP is successful.

3.5 Functional dependence of B red – is locking eliminated?

In the following, we intend to study the structure of the reduced B matrix B red (see Eq. (3.28)) further.
The matrix M hg stands behind the bracket. So, we leave this matrix out in the first step. For a rectan-

65
gular element shape with sidelengths 2 a by 2 b, the 3x2 matrices j 0 Lhg and −j 0 Le K −1
ww K wu can be
f
summarized to read
 r 
0 −b−1 ξ
a−1 η
 
0 r+2
 r 
j 0 Lhg = 0 −1
b ξ , −j 0 Le K −1 K wu =  −a−1 η 0 . (3.29)
 f  
ww
r+2
b ξ a−1 η
−1  
−b−1 ξ −a−1 η

The matrix Lhg is given in (1.56). The calculation of j 0 Le K −1

ww K wu requires some simple com-
f
putation which can be done by e.g. Mathematica or Matlab. Using ε̂ = B e red U e = (B 0 + B hg +
(B red − B hg )) U e , we calculate the three components of the total strain vector ε̂ to read

η T ξ r
εxx = bTx (U e )x + h (U e )x − hT (U e )y ,
a b r+2
ξ η r
εyy = bTy (U e )y + hT (U e )y − hT (U e )x , (3.30)
b a r+2
ξ T η ξ η
2 εxy = bTy (U e )x + bTx (U e )y + h (U e )x + hT (U e )y − hT (U e )x − hT (U e )y .
b a b a
Please note that the stabilization vector γ, which is part of M hg , will reduce to h if the element has
parallelogram shape. The blue parts come from the contribution B 0 U e , the red parts from B hg U e and
the green parts from (B red − B hg ) U e . For the shear strain εxy , the red and the green part cancel each
other.

The relations (3.30) can be suitably used to investigate the locking in a way alternative to the eigenvalue
analysis performed in Section 2.2.

3.5.1 Limit of incompressibility – volumetric locking?

In the limit situation r → ∞ (this is Λ/µ → ∞), it must be possible to satisfy the incompressibility
constraint εxx + εyy → 0 without enforcing the deformation to vanish. The limit case r → ∞ means that
the ratio r/(r + 2) approaches one. Adding up (3.301 ) and (3.302 ) yields the relation

εxx + εyy = bTx (U e )x + bTy (U e )y = 0, (3.31)

where the red and the green term cancel each other.

Let us consider a specific system (see Fig. 3.3) in order to understand this result. The displacements at
the nodes 1, 2 and 4 are assumed to be constrained such that they vanish. The element has the sidelengths

66
Figure 3.3: Example for volumetric locking. The displacements of the nodes 1, 2 and 4 are constrained.
The material parameters are chosen such that r → ∞ holds. The value of u3 x is negative. Therefore, the
distance |u3 x | is expressed by means of the absolute value of u3 x .

2 by 2, so that bx boils down to g ξ and by becomes g η (see Eq. (1.49)). Using this together with the fact
that u1 = u2 = u4 = o2 holds, lets us write
1 1
u3 x + u3 y = 0 ⇒ u3 x = −u3 y . (3.32)
4 4
Obviously, this is just the expected result also visualized in Fig. 3.3a which must come out to retain the
volume of the element. So, Q1/E4 and Q1SP (which are identical for parallelogram shaped elements)
are free of volumetric locking in the present case.

What is the situation for Q1? To investigate that, we simply leave out the green terms in (3.30). Instead
of (3.31), we obtain the relation
η T ξ
εxx + εyy = bTx (U e )x + bTy (U e )y + h (U e )x + hT (U e )y = 0. (3.33)
a b
Investigating again the system shown in Fig. 3.3, one arrives at the statement
1 1 1 1
(u3 )x + (u3 )y + η (u3 )x + ξ (u3 )y = 0. (3.34)
4 4 4 4
In general, ξ 6= η holds within the element, the latter equation requires the solution (u3 )x = (u3 )y = 0
which represents volumetric locking.

3.5.2 Pure bending – shear locking?

It was discussed that for pure bending, the shear strain must have the possibility to vanish. So we have
the constraint εxy = 0. In the relation (3.303 ) the red and green terms cancel each other. We finally

67
obtain
2 εxy = bTy (U e )x + bTx (U e )y . (3.35)

Let us investigate the system depicted in Fig. 3.4. All displacements in y direction are zero. Further, we
can state u1 x = −u2 x = u3 x = −u4 x = −u or, equivalently, (U e )x = 4 h u1 x . Since by is proportional
to g η and g Tη h = 0 holds, the latter result for (U e )x is consistent with (3.35). The term bTy (U e )x
vanishes independently of u1 x . The constraint (3.35) does not enforce the vanishing of the displacement
(u1 )x . So, Q1/E4 and Q1SP are free of shear locking in the present case.

Again, we carry out the analogous investigation for Q1. The green terms are not present, and we obtain
ξ T η
2 εxy = bTy (U e )x + bTx (U e )y + h (U e )x + hT (U e )y . (3.36)
b a
Working with (U e )x = 4 h u1 x and (U e )y = o4 leads to the constraint

ξ
u1 x = 0. (3.37)
b
Due to the fact that ξ = 0 does not hold everywhere within the element, the relation enforces u1 x to
vanish. This is shear locking.

Summary. In both investigated examples, the part connected to B red − B hg (indicated by green colour)
is needed to get rid of the disturbing linear terms. If the linear terms are not cancelled, locking will be
observed.

Figure 3.4: Example for shear locking. The distance u is equal to the displacement of the nodes 2 and 4.
The nodes 1 and 3 undergo the displacement −u.

68
3.6 Eigenvalue analysis at element level and comparison with ana-
lytical result

The next task is to carry out the eigenvalue analysis of the element stiffness matrix K e = K 0 + K stab .
As in Section 2.2 and Section 2.3, this analysis will finally be based on a rectangular-shaped element
(sidelengths 2 a and 2 b). For such an element shape, the element stiffness matrices of Q1SP and Q1/E4
are identical. The part K 0 is included in the element stiffness matrix of all three formulations Q1, Q1/E4
and Q1SP. The investigation in Sections 2.2 and 2.3 showed that the eigenvalues of K 0 are physically
reasonable and do not contribute to the locking. Therefore, we are in the good situation that we do not
have to investigate eigenvalues and eigenvectors of K 0 again. This situation is different for K stab .

The eigenvalue analysis of K hg was simple because the corresponding submatrix K csub 4 had diagonal
structure (see Eq. (2.234 )). Fortunately, this will also hold if K
csub 4 is computed on the basis of K stab .
A prerequisite is that the element has rectangular shape.

Independently of the element shape, we may state

Z
B red = j 0 (L4x8
hg − Le K −1
ww
cwu ) M with K
K cwu = LTe j T0 Ê j 0 L4x8
hg det J 0 dξ dη t. (3.38)
| {z } 
B red
b

This lets us, analogously to (2.10), write

Z 
T T
K stab = M (B red ) Ê B red det J 0 dξ dη t M
b b

| {z }
K stab
c

0 0 0 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 K
b 44 0 0 0 K
b 48 
= M   M. (3.39)
 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 K
b 48 0 0 0 K
b 88
| {z }
K
cstab

Again, this result is – as in Section 2.2.1 – completely general, i.e. independent of the shape of the
element.

69
However, in order to obtain a diagonal structure for K
cstab as well as simple expressions for K
b 44 and K
b 88 ,
an element of rectangular shape with sidelengths 2 a by 2 b is assumed. In the same way as discussed in
Section 2.2.3, we set up the submatrix
" # 
s−1 0 4

K
b 44 K
b 48 r+1
K sub 4 =
c = µt (3.40)
K84 K88
b b 0 s 3 r+2

which contains now quite different entries. The eigenvalue problem for K
csub 4 is easily solvable. We
obtain the eigenvalues

ω4 = Kb 44 = 4 µ t s−1 r + 1 , ω8 = K
b 88 = 4 µ t s r + 1 (3.41)
3 r+2 3 r+2
together with the same eigenvectors as in Section 2.2.3.

In Section 2.3, a comparison of the strain energy ψeFEM computed for one Q1 element and the corre-
sponding analytical strain energy ψecont was carried out. The task here is to determine ψeFEM for Q1SP
and Q1/E4 (which yield identical results for rectangular elements). In the same way as discussed in
Section 2.3.3, the following result is obtained:
4 r+1 8 µt r + 1 2
ψeFEM = µ t s−1 2 p2x = p (3.42)
|3 {z r + 2}
|{z} 3 s r+2 x
ω4 (U Te ϕ4 )2 /2
The finite element-based strain energy ψeFEM is equal to the continuum mechanical result (2.61) which
shows that the strategy taken in the new element formulations Q1/E4 and Q1SP to eliminate the locking
is successful. One could do the analogous analysis for ω8 , with the same positive result.

The volumetric locking is eliminated because the hourglass eigenvalues ω4 and ω8 do not show a non-
physical increase for r → ∞. Further, the term s = a/b is no longer contained in ω4 . This term causes
shear locking for the standard formulation Q1. The fact that ω4 approaches infinity for s−1 = b/a → ∞
is a physically reasonable result, because the element is then extremely high. In this case, a large stiffness
must be associated with the hourglass mode (U e )x = h̄. These considerations apply analogously to the
second hourglass eigenvalue ω8 .

Summary. The element formulations Q1/E4 and in particular Q1SP are in their final application hardly
more elaborate than Q1. Even for general element shapes, Q1SP allows the analytical computation of
K e . Certainly, the derivation is more complicated. In order to apply Q1/E4 or Q1SP, however, it is not
necessary to understand the theoretical background. To summarize, it can be stated that Q1/E4 and Q1SP
fulfil all requirements which were mentioned before. These are extremely efficient element formulations
with only four nodes which do not show any artificial stiffening. As such, they are able to provide good
solutions with rather coarse meshes. This was already shown in Section 2.1.

70
3.7 Pressure-displacement approach or selective reduced integra-
tion

In Section 3.4, the unification of the two concepts mixed formulation and reduced integration with hour-
glass stabilization was discussed. The formulation Q1SP is the result of this beneficial interplay, Q1/E4
the mixed method upon which it is based. In the 19seventies, it was already recognized that reduced
integration alone (without hourglass stabilization) is not sufficient. So, one came to the suggestion to
“underintegrate” only some part of the stress response. This method was called selective reduced in-
tegration and showed up at first in the context of plates (see e.g. [ZTT71]). More or less at the same
time, so-called pressure-displacement formulations were developed, where the pressure (the volumet-
ric part of the stress) was differently interpolated than the displacement. A very common strategy was
to use a constant interpolation for the pressure and a linear interpolation for the displacement. It was
recognized already in 1978 [MH78] that a unification of concepts is possible, was recognized already
in 1978 [MH78]. This branch of finite element technology developed somewhat independently of the
enhanced strain method. Interestingly, one can see it as a special case of the mixed method laid down at
the beginning of this chapter.

3.7.1 Repetition of the strong form - construction of the three-field variational

functional

Again, we repeat the basic equations (1.15). However, the stress matrix σ and then also Hooke’s law are
split into volumetric (σ V ) and deviatoric (σ D ) parts. We obtain

σ = σV + σD (3.43)

1 1
with σV = (tr σ) I = K tr ε I, σ D = σ − (tr σ) I = 2 µ εD . (3.44)
|3 {z } 3
p
The deviatoric part εD of the strain ε is formulated analogously to σ D . Hooke’s law is now represented
in terms of the bulk modulus K = Λ + (2/3) µ and the shear modulus µ. The expression (tr σ)/3 is
alternatively denoted as pressure p.

71
Using the volumetric-deviatoric split, we present the basic equations (1.15) in the format
kinematics : εV − εVc = 0,
ε D − εD
c = 0,

Hooke’s law : σ V − σ Vε = 0, (3.45)

σD − σD
ε = 0,

equilibrium : div (σ V + σ D ) + f = 0,
where the short hand notations εVc = tr (sym (grad u)) I/3, εD D V
c = (sym (grad u)) as well as σ ε =
K (tr ε) I and σ D D
ε = 2 µ ε are used.

In contrast to Section 3.2, we will insert the equations for the deviatoric parts, these are (3.452 ) and
(3.454 ), into (3.455 ). The equations (3.451 ) and (3.453 ) are treated independently. In the finite element
formulation to be developed on this ground, (3.452 ) and (3.454 ) will be fulfilled pointwise. So, the
variables εD and σ D are dependent quantities. The former depends on u, the latter on εD and then
also on u. The volumetric parts, however, these are the quantities εV = (tr ε) I/3 = εV I and σ V =
(tr σ) I/3 = p I, remain as independent variables in the system. Please note that we have introduced
here – to ease notation – analogously to the very common pressure p, the scalar quantity εV . To conclude,
we have now again three independent variables. Besides the displacement vector u, these are the scalar
quantities εV and p.

Accordingly, each of the remaining equations of the strong form has to be multiplied with a test function.
One arrives, analogously to (3.2) to (3.4), at the three equations
Z
(δ σ̂ V )T (ε̂V − ε̂Vc ) dV = 0, (3.46)
ZB
(δε̂V )T (σ̂ V − σ̂ Vε ) dV = 0, (3.47)
Z Z B Z
V T V D T D T
(δε̂c ) σ̂ + (δε̂c ) σ̂ ε ) dV − δu f dV − δuT t̄ dA = 0. (3.48)
B B ∂Bt

In contrast to Section 3.2, the equations (3.46) and (3.47) include only the volumetric parts. In the
last equation (3.48), the deviatoric part σ̂ D
ε is expressed in terms of ε̂
D
= ε̂D
c and is therefore not
independent. For this finite element formulation also, we intend to limit ourselves to a plane strain state.
Nevertheless, using the volumetric-deviatoric split has the disadvantage that we need to consider the
stress σzz in the third direction, which is not zero.

Voigt notation of volumetric-deviatoric split. For the last and the following transactions, the Voigt
notation of volumetric and deviatoric quantities has to be clarified. First, we will do that for the stress

72
vector σ̂ and will then go over to the strain vector ε̂. An intermediate step is to write down the normal
stress components
1 1
σxx = σxx − (σxx + σyy + σzz ) + (σxx + σyy + σzz )
| 3 {z } |3 {z }
D
σxx p
1 1
σyy = σyy − (σxx + σyy + σzz ) + (σxx + σyy + σzz )
| 3 {z } |3 {z }
σ D p
yy
1 1
σzz = σzz − (σxx + σyy + σzz ) + (σxx + σyy + σzz ) (3.49)
| 3 {z } |3 {z }
D
σzz p

D
The volumetric part of the shear stress components is zero, such that σxy = σxy holds. So, we come up
with the extended Voigt notation
     

 σxx 
  2 σxx − σyy − σzz
 
 
 1 
 

σyy
 1  2 σyy − σzz − σxx 
1

σ̂ 4x1 = = +p

 σzz 
 3 
 2 σzz − σxx − σyy 
 
 1 
 
σxy 3 σxy 0
    

= σ̂ D + σ̂ V (3.50)

which implies the additive split of the 4x1 vector σ̂ 4x1 into σ̂ D and σ̂ V = p i with iT = {1, 1, 1, 0}.

The analogous notation for the strain vector ε̂(c) is given by

     

 ε(c) xx  
 2 ε(c) xx − ε(c) yy − ε(c) zz 
 
 1 
1  
ε(c) yy 2 ε(c) yy − ε(c) zz − ε(c) xx 1
    
ε̂4x1
(c) = = + εV(c)
ε 3  2 ε(c) zz − ε(c) xx − ε(c) yy 1 
 (c) zz 

   
 

 
2 ε(c) xy 6 ε(c) xy 0
   

= ε̂D V
(c) + ε̂(c) . (3.51)

Here, the index (c) is added to indicate that the notation does not only hold for the total strain vector
ε̂4x1 but also for the compatible part ε̂4x1
c . The interpolation of the 3x1 vector ε̂c by means of the 3x8 B
matrix derived in Section 1.5 takes into account that εc zz vanishes. For this reason, the Voigt notation of
the compatible strain normally includes only three entries. In the context of the present formulation, the
situation will become a little more complex. In any case, it can be stated that the standard interpolation
of u will not be changed. Therefore, εc zz will definitely be zero.

ˆ V and a devia-
Finally, it is now easy to understand that scalar products between a volumetric vector (...)
ˆ D are zero. This has been exploited in (3.48), where mixed terms such as the transpose
toric vector (...)

73
 ˆ V multiplied with (...)
of (...) ˆ D do not show up. It should be noted that the vectors (...)
ˆ V ˆ D
(..) and (...)(..)
are defined in such a way that they have always the dimension 4x1.

3.7.2 Concept for a possible discretization

Let us first consider Eq. (3.46). Using δ σ̂ V = δp i and ε̂V(c) = εV(c) i, this equation reads
Z
3 δp (εV − εVc ) dV = 0. (3.52)
B

Please note that 3 εVc is a short hand notation for tr (sym grad u)). We represent it, as already done in
Section 3.4.1, by means of ε̂c = B U e . Denoting each line of B by B i (with i = 1, 2, 3), adding up the
first two lines of B and dividing by 3 yields the interpolation εVc = ((B 1 + B 2 )/3) U e = B V U e , where
B V is a 1x8 matrix. For the next step, it is helpful to proceed analogously to Section 3.2.3. Therefore,
we introduce a so-called volumetric incompatible strain

εVe = εV − εVc . (3.53)

We do not use the word enhanced strain to avoid confusion. But it would not be wrong to interpret εVe
just in this way. The question is, how (3.52) can be fulfilled. One possibility is to choose δp and then
also p to be constant within the element. The same assumption is made for εV and then also δεV . If
δp is constant within the element, the integral in (3.52) will have to vanish over the element domain.
Inserting the assumption that also εV is constant within Be together with the interpolation εVc = B V U e
into (3.52) yields
Z Z
1 1 V
V
ε = εVc dVe = B V det J dξ dη t U e = B̄ U e (3.54)
Ve Be Ve 

as well as
V
εVe = (B̄ − B V ) U e . (3.55)
In contrast to the formulation Q1/E4, we do not introduce a new interpolation function for the incompat-
V
ible strain. The latter is rather given by computing the difference between B̄ and B V , multiplied by
U e . For an element of parallelogram shape, this boils down to εVe = −B Vhg U e , and the relation (3.54)
reduces to εV = B V0 U e . Please note that the split B = B 0 + B hg can be certainly taken over to the 1x8
matrix B V = B V0 + B Vhg , as just performed.

The choice of a suitable interpolation to fulfil (3.46) lets us reduce (3.47) to the statement
Z
((δε̂Vc )T σ̂ V − (δε̂V )T σ̂ Vε ) dV = 0. (3.56)
B

74
This is analogous to the step taken in Section 3.2.3. The latter equation (3.56) is now subtracted from
(3.48) to yield
Z Z Z
V T
(δε̂ ) σ̂ Vε + (δε̂D T D
c ) σ̂ ε ) dV − T
δu f dV − δuT t̄ dA = 0. (3.57)
B B ∂Bt

As in the enhanced strain formulation presented in Section 3.2.3, the independent stress quantity, here
the pressure p, drops out. Instead, we work with σ̂ Vε which we suitably denote by σ̂ Vε = pε i with
pε = 3 K εV . The deviatoric part in (3.57) is completely unchanged with respect to Q1. It is, however,
necessary to find a compact notation for σ̂ D
ε . This can be derived by first writing
 
r+2 r r 0
 r r+2 r 0 
σ̂ 4x1
ε =µ  r
 ε̂4x1 , (3.58)
r r+2 0 
0 0 0 1
| {z }
4x4
Ê
4x4
where we have defined the 4x4 extended elasticity matrix Ê . Using (3.50), the relation between the
D 4x1
deviatoric part σ̂ of the stress vector and the stress vector σ̂ is given by
 
2 −1 −1 0
1  −1 2 −1 0   σ̂ 4x1 = P̂ D σ̂ 4x1 .
σ̂ D =  (3.59)
3  −1 −1 2 0 
0 0 0 3

This holds analogously for the dependent stress vector σ̂ 4x1

ε . It is now possible to write

σ̂ D
D
4x1 D 4x4
b D ε̂4x1 = Ê D ε̂4x1 = Ê D ε̂D = Ê D ε̂D .
ε̂4x1 = 2 µ P
ε = P̂ σ̂ ε = P̂ Ê c (3.60)

In the latter statement, many considerations have been exploited. First of all, we use (3.59), then (3.58).
The matrix product P̂ Ê
D 4x4
b D to be defined as the deviatoric elasticity matrix Ê D .
is equal to 2 µ P
The matrix P b D differs from P̂ D only in the 44 component which is in P
b D the number 1/2 instead of 1
D
visible in P̂ . The matrix Pb D has the property

b D ε̂4x1 = P
P b D (εV i + ε̂D ) = P
b D ε̂D . (3.61)
D D
Thus, we can write Ê ε̂4x1 = Ê ε̂D . Finally, it has been exploited that the incompatible part of the
deviatoric strain is zero. So, ε̂D = ε̂D
c holds. These results are inserted into (3.57) to finally arrive at the
weak form
Z Z Z
D
g= V
(δε 9 K ε + V
(δε̂D T
c ) Ê ε̂D
c ) dV − T
δu f dV − δuT t̄ dA = 0. (3.62)
B B ∂Bt

75
This result is quite an interesting outcome. The first term in the first integral includes with εV a quantity
which is interpolated to be constant within the element. This is a change with respect to Q1, but one
can imagine that the computation of this term is quite simple. The second term in the first integral
is unchanged with respect to Q1 and the loading terms are always the same. So, the present element
formulation turns out to be uncomplicated, although the derivation is challenging. The question is,
whether the small change in the volumetric term will eliminate the locking.

The method is based on a constant interpolation of the pressure and a bilinear interpolation of the dis-
placement. Therefore, it is often termed Q1/P0. Another name for it is “B bar” method. Fundamental
papers are e.g. [NPR74], [MH78] and [STP85]. The version, where J is replaced by J 0 will be termed
Q1RV.

3.7.3 Element stiffness matrix K e in the context of Q1/P0 and Q1RV

It is helpful to recapitulate the interpolations for the volumetric and deviatoric strain. For the former, Eq.
V
(3.54) yields εV = B̄ U e . The vector ε̂V = εV i is then given by
V V
ε̂V = i B̄ U e = B̄ 4x8 U e . (3.63)

Remark. Please note that due to the fact that the vectors ε̂V and ε̂D have the dimension 4x1, we need to
introduce the corresponding 4x8 B matrices. Since the transpose of these matrices is frequently needed,
4x4
the dimension 4x8 is put here as lower index. This is in contrast to the quantities ε̂4x1 4x1
(c) , σ̂ (ε) and Ê ,
where we put the dimension as upper index. 

The interpolation of the deviatoric strain has not yet been discussed. Let us employ index notation such
that using ε̂c = B U e (1.56), each component ε̂c i of ε̂c with i = 1, ..., 4 in (ε̂4x1 T
c ) = {ε̂c 1 , ε̂c 2 , ε̂c 3 , ε̂c 4 }
is represented by ε̂c i = B i U e . As already mentioned, the matrices B i have the dimension 1x8. Please
note that ε̂c 1 = εc xx , ε̂c 2 = εc yy , ε̂c 3 = εc zz and ε̂c 4 = 2 εc xy hold. The deviatoric strain ε̂D
c can then be
written in the way
   

 2 εc xx − εc yy 
 
 2 B1 − B2 

1 
2 εc yy − εc xx
 1 
2 B2 − B1

ε̂D
c = = U e = BD
4x8 U e . (3.64)
3 
 −ε c xx − εc yy 
 3 
 −B 1 − B2 

6 εc xy 3 B3
   

With this information, it is rather simple to identify the element stiffness matrix. Inserting (3.54) and

76
(3.64) into (3.62) leads to the equation
ne Z 
X V T V D
g= δU Te ((B̄ ) 9 K B̄ + (B D
4x8 )
T
Ê BD
4x8 ) det J dξ dη t U e − P e = 0. (3.65)
e=1 
| {z }
Ke

Representation of K e by means of reduced B matrix. Following Section 3.4.2, we seek also here for
a more compact representation by means of a reduced B matrix. This can be done by reformulating the
V V
product (B̄ )T 9 K B̄ according to
V V V V V V
(B̄ )T 9 K B̄ = (i B̄ )T i K iT (i B̄ ) = (B̄ 4x8 )T (K i iT ) B̄ 4x8
V V V
= (B̄ 4x8 )T Ê B̄ 4x8 , (3.66)
D V
where we have introduced – corresponding to Ê – the volumetric elasticity matrix Ê = K i iT .

4x4 4x4
With the knowledge that the total elasticity matrix Ê is alternatively to (3.58) given by Ê =
V D
Ê + Ê , one arrives at the statement
Z
4x4 V
K e = (B̃ red )T4x8 Ê (B̃ red )4x8 det J dξ dη t with (B̃ red )4x8 = (B̄ )4x8 + (B D )4x8 . (3.67)


At this point, we should explain, in which way (B̃ red )4x8 can be interpreted as a reduced B matrix. This
V
is because, in contrast to the full B, the volumetric part B̄ 4x8 is evaluated as an average over the element
volume (see Section 3.7.2). In this way, one eliminates the linear variations of the volumetric part of B
over the element. This will become more evident, when we come to the replacement of J by J 0 .

As before, it is elucidating to compute the expression

V V D 4x1
(B̃ red )4x8 U e = (B̄ 4x8 + B D
4x8 ) U e = ε̂ + ε̂(c) = ε̂ (3.68)

which yields – as expected – the total strain vector ε̂4x1 . Its deviatoric part does not contain an incom-
patible contribution. Therefore, as already mentioned several times, the equality ε̂D = ε̂D
c holds. Very
interesting is the fact that the third component of ε̂4x1 , i.e. εzz , does not vanish. This is because the third
V D
line of the sum B̄ 4x8 + B̄ 4x8 does not yield the zero line vector oT8 .

Obviously, this is a surprising outcome. However, it has to be taken into account that in the context of
the Q1/P0 method, the total strain has only a theoretical meaning. In the present lecture notes, the Q1/P0
formulation is interpreted as a special case of a mixed method based on the notion of an incompatible
strain. Therefore, it was convenient to introduce the incompatible strain and the total strain also in the

77
context of Q1/P0. In the original papers (see [NPR74], [MH78]), however, the method was derived
differently. The quantity of a total strain in the way defined here did not show up. We will come back to
this point further below.

Evaluation of K e by means of standard numerical integration. There is nothing to add. In general,

the B operators for the volumetric and the deviatoric part include with j a matrix which shows a complex
dependence on ξ and η. Consequently, numerical integration has to be applied.

Evaluation of K e by means of analytical integration. As before, the step of replacing J by J 0 is car-

ried out. It was already mentioned in Section 3.7.2 that this leads to a simplification for the computation
of B̄ V . Using (3.54), we obtain
Z
1 1
B̄ V = B V det J 0 dξ dη t = 4 det J 0 t B V0 = B V0 (3.69)
4 det J 0 t  4 det J 0 t
To arrive at the latter statement, we have exploited the fact that the integral over the part B Vhg will vanish
if B hg is computed on the basis of j 0 instead of j. The inflation of the 1x8 matrix B V0 to the 4x8 matrix
(B V0 )4x8 is done as described before in the context of B V . So, by inserting J 0 , the matrix B̃ red (3.67)
reduces to
e red )4x8 = (B V )4x8 + B D .
(B (3.70)
0 4x8

It is instructive to take a closer look at it. Splitting the lines B i according to B i = (B 0 )i + (B hg )i

(i = 1, 2, 3), we can write in detail
     
(B 0 )1 + (B 0 )2 2 B1 − B2 2 (B hg )1 − (B hg )2
1  (B 0 )1 + (B 0 )2  1  2 B 2 − B 1
 = (B 0 )4x8 + 1  2 (B hg )2 − (B hg )1 
  
(Be red )4x8 =  + 
3  (B 0 )1 + (B 0 )2  3  −B 1 − B 2  3  −(B hg )1 − (B hg )2 
o8 3 B3 3 (B hg )3

= (B 0 )4x8 + (B D
hg )4x8

= B 4x8 − (B Vhg )4x8 (3.71)

e red )4x8 − (B 0 )4x8 which is given by

Finally, we seek to compute (B red )4x8 = (B

(B red )4x8 = (B hg )4x8 − (B Vhg )4x8 = (B D

hg )4x8 . (3.72)

It should have become clear now, how the Q1/P0 or B bar formulation works. In the context of Q1, the
matrix K hg which leads to the locking is computed on the basis of B hg . In the present formulation, this
is changed by ignoring the volumetric part of the hourglass part of the B operator. Only the deviatoric
part is used to compute K stab in
Z
4x4
K e = K 0 + (B red )T4x8 Ê (B red )4x8 det J 0 dξ dη t = K 0 + K stab . (3.73)


78
The method based on the replacement of J by J 0 is alternatively called “selective reduced integration”.
Let us abbreviate the method by Q1RV, R stands for reduced integration, V for the volumetric part. The
element stiffness matrix can be obtained by full integration of the deviatoric part and reduced integration
of the volumetric part.

3.7.4 Functional dependence of (B red )4x8 – is locking eliminated?

As in Section 3.5, we are interested to see the effect of working with (B red )4x8 instead of using (B hg )4x8 .
For this purpose, we assume a rectangular element shape with side lengths 2 a by 2 b. For this situation,
the original 3x8 matrix B hg is given by j 0 Lhg M hg . The extension to the 4x8 matrix B hg is easily
carried out by inserting a zero line into j 0 . For the assumed rectangular element shape, the resulting
matrices (j 0 )4x4 and (j 0 )4x4 Lhg take on the formats
 −1
a−1 η
  
a 0 0 0 0
 0 0 0 b−1   0 b−1 ξ 
(j 0 )4x4 = 
 0
 and (j 0 )4x4 Lhg = , (3.74)
0 0 0   0 0 
0 b−1 a−1 0 b−1 ξ a−1 η
respectively. We obtain (B hg )4x8 = (j 0 )4x4 Lhg M hg . To arrive at (B red )4x8 , the matrix (B Vhg )4x8 has
to be subtracted from (B hg )4x8 . It is helpful to represent (B Vhg )4x8 in the format ((j 0 )4x4 Lhg )V M hg ,
where the first factor is given by
a−1 η b−1 ξ
 
1  a−1 η b−1 ξ 
−((j 0 )4x4 Lhg )V = −  . (3.75)
3  a−1 η b−1 ξ 
0 0
Analogously to Section 3.5, we seek to compute the components of the total strain vector ε̂4x1 =
e red )4x8 U e = ((B 0 )4x8 + (B hg )4x8 + ((B red )4x8 − (B hg )4x8 )) U e . This computation leads to
(B
η T η T ξ T
εxx = bTx (U e )x + h (U e )x − h (U e )x − h (U e )x ,
a 3a 3b
ξ η T ξ T
εyy = bTy (U e )y + hT (U e )y − h (U e )x − h (U e )x ,
b 3a 3b
η T ξ T
εzz = − h (U e )x − h (U e )x ,
3a 3b
ξ T η
2 εxy = bTy (U e )x + bTx (U e )y + h (U e )x + hT (U e )y . (3.76)
b a

Limit of incompressibility – volumetric locking? Again, it must be possible to satisfy the incompress-
ibility constraint tr ε = 0 without enforcing the deformation to vanish. It has already been mentioned

79
and is highly interesting that εzz is not zero. It is now clearly visible that this is due to the green, i.e. the
incompatible part of the strain. The blue and the red part which belong to the compatible strain do not
provide any contribution to εzz . Adding up the first three lines of (3.76) leads to the result

εxx + εyy + εzz = bTx (U e )x + bTy (U e )y = 0 (3.77)

which is absolutely identical to (3.31). So, the volumetric locking is eliminated which is on the one
hand a very positive outcome. On the other hand, the way how this result is achieved, is not completely
convincing, since the total strain vector ε̂4x1 violates the assumption of a plane strain state. As already
pointed out, the latter is only fulfilled for the compatible strain vector ε̂4x1
c .

Pure bending – shear locking? In order to arrive at a behaviour which is free of shear locking, the shear
strain εxy must be able to vanish without enforcing U e to approach the zero vector o8 . It is immediately
visible that the expression (3.764 ) for the shear strain is identical to the one of Q1 which exhibits shear
locking. So, unfortunately, the Q1/P0 (or B bar) and the Q1RV formulations do not cure the shear
locking.

3.7.5 Eigenvalue analysis at element level and comparison with analytical result

As in Section 3.6, we intend to compute the eigenvalues of K e to confirm the previous findings about
the locking behaviour. The investigation is carried out just in the same way as already documented in
Section 3.6. We obtain the eigenvalues
4 1 4 1
ω4 = ( s−1 + s) µ t and ω8 = ( s + s−1 ) µ t. (3.78)
9 3 9 3
The finite element-based strain energy ψeFEM for Q1/P0 and Q1RV reads
4 1 8 2
ψeFEM = ( s−1 + s) t 2 p2x = ( s−1 + s) µ t p2x (3.79)
|9 {z 3 } T|{z} 2 9 3
ω4 (U e ϕ4 ) /2

which is clearly different from the continuum mechanical result (2.61). The eigenvalues and the strain
energy ψeFEM are independent of r. Thus, the quantities do not approach infinity for r → ∞ which
indicates that volumetric locking is eliminated. However, both eigenvalues ω4 and ω8 go to infinity for
either s → ∞ or s−1 → ∞, respectively. This means that the stiffness to withstand bending, when
the structure is very thin in the respective direction (for instance in y direction for the bending mode
(U e )x = h̄), is far overestimated. So, we see shear locking. The fact that the eigenvalue ω4 approaches
∞ for s−1 → ∞ is physically reasonable. These statements hold analogously for ω8 .

80
Summary. The element formulations Q1/P0 and Q1RV are free of volumetric locking. As such, they
show in many examples a better behaviour than Q1. If the replacement of J by J 0 is not carried out,
Q1/P0 might be slightly more efficient than Q1/E4, because the inversion of a matrix such as K ww in
the case of Q1/E4 is not necessary. If it comes to non-linear applications, Q1/P0 will have the advantage
that no additional degrees-of-freedom have to be determined. This is then necessary for Q1/E4.

Nevertheless, if the replacement of J by J 0 is performed, Q1SP will have only advantages over Q1RV.
This is because the computational efficiency is more or less the same. Please note that for Q1SP, the
matrix K ww can be inverted analytically. A major argument is that Q1SP is not only free of volumetric
locking but also of shear locking.

81
Summary and conclusions

Chapter 1. The text book is restricted to plane strain linear elasticity. The relevant equations of the
strong form are derived in Section 1.1.2. In Section 1.2, the corresponding weak form is formulated
which can be seen as a one-field variational functional, where only the displacement vector u is unknown.
Afterwards, the isoparametric element concept is introduced. Based on that, bilinear interpolations for
u and the geometry x are defined (Section 1.3). It should be emphasized that the present lecture notes
are based on the viewpoint that one seeks to work with a finite element formulation of low order. The
idea is to keep the bilinear interpolation of u and x for all formulations. A major step of the present
derivation is to split the shape function vector N into linear and hourglass parts (Section 1.4). The split
is transferred to the B operator which is split accordingly, in this case into a constant and an hourglass
part (Section 1.5).

The Jacobi matrix J which represents the mapping from the reference to the real element in the isopara-
metric element concept plays a major role in finite element technology. The properties of J are investi-
gated in detail in Section 1.6. In particular interesting is the notion of the equivalent parallelogram which
is the element shape obtained by neglecting the linear parts in J . This is nothing else than evaluating
the Jacobi matrix in the centre of the element which renders the frequently used J 0 . In Section 1.7, the
main focus is on the computation of the element stiffness matrix K e . It can be computed by numerical
integration or by analytical integration, whereas it has to be taken into account that the analytical integra-
tion requires the replacement of J by J 0 . In this way, tacitly, every element is replaced by its equivalent
parallelogram. It has been shown in mathematical literature that the error of this approximation vanishes
with increasing mesh density. The chapter closes with a short remark on the assembly procedure and
references for further reading on the standard finite element method.

The finite element formulation derived in Chapter 1 is called Q1, where the letter Q refers to the notion
quadrilateral (four-node element) and the number 1 to the polynomial order of the shape functions. Let
us use the abbreviation Q1J0 for the formulation based on J 0 .

82
Chapter 2. The second chapter is devoted to the investigation of locking. The locking effect is charac-
terized by an overestimation of the stresses and an underestimation of the deformation. In Section 2.1,
two examples are given, where locking is highly evident. The two cases refer, first of all, to material be-
haviour in the limit of incompressibility and, secondly, to pure bending. One observes volumetric locking
and shear locking, respectively. It is well-known that these undesirable phenomena can be investigated
by carrying out an eigenvalue analysis of the element stiffness matrix K e (Section 2.2).

To gain a deeper understanding of locking, it is of high value to derive eigenvalues and eigenvectors
analytically. This is achieved by representing K e in the format M T (K
c0 + Kchg ) M , where M is an
orthogonal matrix and the additive split of K e into the two summands K 0 and K hg has been employed.
Further, a rectangular element shape is assumed. After several mathematically challenging steps, it
turns out that K hg possesses two eigenvalues which show non-physical behaviour. These two so-called
hourglass eigenvalues approach infinity in the limit of incompressibility and for extreme aspect ratios of
the underlying element geometry. So, the volumetric and the shear locking are already visible here.

To confirm this finding, the finite element-based strain energy of one element is compared to the cor-
responding strain energy at continuum mechanical level (Section 2.3). This analysis brings out the
interesting result that the finite element-based strain energy computed on the basis of K 0 is physically
correct, whereas the one derived from K hg is not. So, it can be clearly stated that the locking is attributed
to the so-called hourglass part K hg of the element stiffness matrix. One can also state that the hourglass
part of the shape function vector, i.e. N hg = ξ η γ resulting in the hourglass part B hg of the B operator
causes the difficulties.

Chapter 3. In Chapter 3, various ideas to cure the locking are presented. The goal must be to change
K hg in an intelligent way. Looking at the shape functions NI (I = 1, ..., 4), one possible idea is to omit
the disturbing bilinear part. Another good strategy is to add the terms ξ 2 and η 2 in order to render the
interpolation somewhat complete up to second order (Section 3.1). The goal is to arrive at a strain which
is complete up to first order.

The question arises as to how this can be achieved in a variationally consistent manner. To approach
this aim, in Section 3.2, a mixed variational formulation is presented which is based on considering three
fields, the displacement vector u, the strain vector ε̂ and the stress vector σ̂ as independent quantities. At
this point, the notion of the enhanced strain vector ε̂e = ε̂− ε̂c should be repeated which is the difference
between the total strain vector ε̂ and its compatible part ε̂c .

By choosing the independent stress quantity σ̂ to be constant within the element, the field σ̂ drops out

83
of the formulation and the so-called orthogonality condition for ε̂e is derived. The original three-field
variational functional reduces to a two-field variational functional, where u and ε̂e are the independent
quantities. Subsequently, Section 3.3 is devoted to find a suitable interpolation for the enhanced strain
ε̂e . The resulting functional dependence is such that the total strain vector ε̂ now depends linearly on
both local coordinates ξ and η, as desired. The interpolation includes four additional unknowns listed in
the vector W e .

Using this information, the element stiffness matrix is derived (Section 3.4). As mentioned before, the
standard procedure to compute the integral is numerical integration which is easily carried out. The finite
element formulation derived up to this point is called Q1/E4, where E4 stands for the four additional
(enhanced) degrees-of-freedom. As already discussed in Chapter 1, it is possible to approximate the
Jacobi matrix J by its value J 0 evaluated in the centre of the element. This leads to the possibility
to determine the integral in K e analytically. This element formulation is called Q1SP, S stands for
stabilization (of K 0 ), P for equivalent parallelogram. Coming up with the so-called reduced B matrix
B red , one arrives at the representation K e = K 0 + K stab with K stab =  B Tred Ê B red det J 0 dξ dη t.
R

So, the goal – formulated at the beginning – to change the part K hg in K e has been reached. The matrix
K hg is replaced by K stab . The reduced matrix B matrix B red is given in the table below.

For the special case of a rectangular element, it is possible to formulate the components of the strain
vector ε̂ analytically (Section 3.5). It is now possible to check, whether the two constraints tr ε = 0
(incompressibility) and εxy = 0 (pure bending) lead to the vanishing of the deformation. In fact, this
is no longer the case! Both types of locking are cured by working with Q1/E4 or Q1SP. This is also
confirmed by the eigenvalue analysis and the comparison with the continuum mechanical result for the
strain energy carried out in Section 3.6.

Both element formulations Q1/E4 and Q1SP are already perfect in the regard that the continuum mechan-
ical strain energy is exactly matched. Locking is cured. Nevertheless, in a text book on finite element
technology, the so-called pressure-displacement approach should not be forgotten (Section 3.7). It is
mainly motivated by the observation that Q1 is not able to display the volumetric part of the strain (or
stress) in a proper way. For this reason, a split of stress and strain into deviatoric and volumetric parts is
carried out. The volumetric part is interpolated differently than the deviatoric part. To be more concrete,
frequently, the volumetric part of the stress and the strain are assumed to be constant within the element.
The deviatoric parts are interpolated in the same way as done in the context of the element formulation
Q1.

In the present text book, we formulate this so-called Q1/P0 formulation – P0 stands for the constant in-

84
terpolation of the pressure which is equal to tr σ/3 – as special case of the mixed variational formulation
started in Section 3.2. The following steps are analogous to the ones taken in Section 3.3 to 3.6. A major
difference is, however, seen in the interpolation of the incompatible strain which is not carried out by
using additional degrees-of-freedom but via a suitably formulated average of the so-called volumetric B
matrix multiplied with the already existing element vector U e of nodal displacements. As before, the
element stiffness matrix K e can be computed, first of all, by numerical integration and, secondly, by
analytical integration. In the second case, the Jacobi matrix J is replaced by J 0 . We call this element
formulation Q1RV, where R stands for reduced integration of the volumetric part (V). Again, we come
up with the representation K e = K 0 + K stab , where K stab is now computed by means of a different
reduced B matrix (see the table below). Unfortunately, the latter choice allows to eliminate only the
volumetric locking, not the shear locking. This is confirmed by an eigenvalue analysis, the analytical
computation of the strain components and finally the comparison with the continuum mechanical result.

element formulation B red present types of locking

Q1J0 B red = B hg = j 0 Lhg M hg volumetric and shear locking

Q1SP B red = j 0 (Lhg − Le K −1
ww K wu ) M hg
f free of locking
Q1RV (B red )4x8 = (j 0 )D
4x8 Lhg M hg shear locking

85
Chapter 4

Glossary

4.1 Notions in the text

bubble modes Section 3.3

compatible strain Section 3.2.1
displacement vector Section 1.1.2
eigenvalue problem Section 2.2.2
elasticity constants Section 1.1.2
elasticity law Section 1.1.2
element load vector Section 1.7
element nodal coordinates Section 1.3
element nodal displacement vector Section 1.3
element stiffness matrix Section 1.7
enhanced strain Section 3.2.3
equilibrium Section 1.1.2
equivalent parallelogram Section 1.6.4
global load vector Section 1.8
global stiffness matrix Section 1.8
global displacement vector Section 1.8
Hooke’s law Section 1.1.2
incompatible modes/displacements Section 3.1, Section 3.3
isoparametric concept Section 1.3
Jacobi matrix Sections 1.4 and 1.6
kinematical relation Section 1.1.2
(continuum mechanics)
locking Motivation

86
numerical integration Section 1.7
orthogonal matrix Section 2.2.2
potential Section 1.2
pressure Section 3.7.1
Q1 Beginning of Chapter 2
Q1/E4 Section 3.3, Section 3.4, Section 3.5, Section 3.6
Q1/E12 Section 3.3
Q1/P0 Section 2.1.1, Section 3.7
Q1RV Section 3.7.3
Q1SP Section 2.1.1, Section 3.4, Section 3.5, Section 3.6
reduced B matrix Section 3.4.2, Section 3.7.3
reduced integration Section 3.1, Section 3.4.4, Section 3.7.3
shape function (vector) Section 1.3
shear locking Section 2.1
spectral decomposition Section 2.3
stabilization vector Section 1.4
strain Section 1.1.1 (1D), Section 1.1.2 (2D, 3D)
strain-displacement operator Section 1.5
(B operator)
strain energy Section 2.3
stress Section 1.1.1 (1D), Section 1.1.2 (2D, 3D)
strong form Section 1.1
test function Section 1.2
variation Section 1.2
virtual work Section 1.2
volumetric-deviatoric split Section 3.7.1
volumetric locking Section 2.1
weak form Section 1.2

87
4.2 Mathematical symbols (only abbreviations)

This list contains only symbols which are abbreviations and are not associated with the quantities listed
above.

(B .... )4x8 various B matrices extended to the dimension 4x8 Section 3.7.3
bx , by part of N lin Section 1.4
cx , cy part of B hg Section 1.5
ε̂4x1 4x1 Voigt notation of strain Section 3.7.1
εD deviatoric part of strain matrix Section 3.7.1
εV volumetric part of strain matrix Section 3.7.1
εV trace of ε divided by three Section 3.7.1
4x4
Ê 4x4 elasticity matrix Section 3.7.2
V D
Ê , Ê volumetric and deviatoric parts of Ê Section 3.7.2
g left hand side of weak form Section 1.2
(virtual work)
G 3x4 interpolation matrix for enhanced strain Section 3.3
i 4x1 vector, equal to {1, 1, 1, 0}T Section 3.7.2
j part of B hg Section 1.5
k factor in K
c0 Section 2.2.1
k
b factor in K
csub K (K = 1, ..., 4) Section 2.2.2
K
c part of element stiffness matrix Section 2.2.1
K
csub K submatrices of Kc (K = 1, ..., 4) Section 2.2.3
K uu , K uw , K ww parts of K e for Q1/E4 Section 3.4.1
L0 part of B 0 Section 2.2.1
Le part of G Section 3.3
Lhg part of B hg Section 1.5
4x8
Lhg part of B hg Section 2.2.1
M 8x8 matrix Section 2.2.1
M hg part of B hg Section 1.5
P 2x1 vector which includes quadratic interpolation Section 3.3
r material parameter ratio Section 2.2.1
s side length ratio Section 2.2.2
S summarizes interpolation of u Section 1.7
σ̂ 4x1 4x1 Voigt notation of stress Section 3.7.1
σε stress matrix epressed in terms of ε Section 3.2.1
σD deviatoric part of stress matrix Section 3.7.1
σV volumetric part of stress matrix Section 3.7.1
We 4x1 element vector of additional degrees-of-freedom Section 3.3

88
Chapter 5

Literature for further reading

[Arm00] F. Armero, On the locking and stability of finite elements in finite deformation plane strain
problems, Computers and Structures 75 (2000), 261-290.

[AR95a] K. Arunakirinathar, B.D. Reddy, Some geometrical results and estimates for quadrilateral
finite elements, Computer Methods in Applied Mechanics and Engineering 122 (1995), 307-314.

[AR95b] K. Arunakirinathar, B.D. Reddy, Further results for enhanced strain methods with isoparam-
eteric elements, Computer Methods in Applied Mechanics and Engineering 127 (1995), 127-143.

[BLM+14] Belytschko, Ted; Liu, Wing Kam; Moran, Brian, Elkhodary, Khalil, Nonlinear finite ele-
ments for continua and structures. Wiley, 2014.

[BOL+84] T. Belytschko T., J. S.-J. Ong, W.K. Liu, J.M. Kennedy, Hourglass control in linear and
nonlinear problems, Computer Methods in Applied Mechanics and Engineering 43 (1984), 251-
276.

[Cha12] Chadwick, Peter, Continuum mechanics: concise theory and problems. Dover, 2012.

[DPH+95] E. de Souza Neto, D. Perić, G. Huang, D.R.J. Owen, Communications in Numerical Meth-
ods in Engineering 11 (1995), 951-961.

[Hug00] Hughes, Thomas J.R., The finite element method: linear static and dynamic finite element
analysis. Dover, 2000.

[KT00] E.P. Kasper, R.L. Taylor, A mixed-enhanced strain method: Part I: Geometrically linear prob-
lems, Computers & Structures 75 (2000), 237-250.

89
[MH78] D.S. Malkus, T.J.R. Hughes, Mixed finite element methods - reduced and selective integration
techniques: a unification of concepts, Computer Methods in Applied Mechanics and Engineering
15 (1978), 63-81.

[NPR74] J.C. Nagtegaal, D.M. Parks, J.R. Rice, On numerically accurate finite element solutions in the
fully plastic range, Computer Methods in Applied Mechanics and Engineering 4 (1974), 153-177.

[OBC71] Oden, J. Tinsley; Becker, Eric B.; Carey, Graham F., Finite elements: an introduction. Volume
I. Prentice Hall, 1971.

[RKR99] S. Reese, M. Küssner, B.D. Reddy, A new stabilization technique for finite elements in non-
linear elasticity, International Journal for Numerical Methods in Engineering 44 (1999), 1617-
1652.

[RS95] B.D. Reddy, J.C. Simo, Stability and convergence of a class of enhanced strain methods, SIAM
Journal on Numerical Analysis 32 (1995), 1705-1728.

[RW00] S. Reese, P. Wriggers, A stabilization technique to avoid hourglassing in finite elasticity, Inter-
national Journal for Numerical Methods in Engineering 48 (2000), 79-109.

[RWR00] S. Reese, P. Wriggers, B.D. Reddy, A new locking-free brick element technique for large
deformation problems in elasticity, Computers & Structures 75 (2000), 291-304.

[SAT93]J.C. Simo, F. Armero, R.L. Taylor, Improved versions of assumed enhanced strain tri-linear
elements for 3D finite deformation problems, Computer Methods in Applied Mechanics and En-
gineering 110 (1993), 359-386.

[SR90] J.C. Simo, S. Rifai, A class of mixed assumed strain methods and the method of incompatible
modes, International Journal for Numerical Methods in Engineering 29 (1990), 1595-1638.

[STP85] J.C. Simo, R.L. Taylor, K. Pister, Variational and projection methods for the volume constraint
in finite deformation elasto-plasticity, Computer Methods in Applied Mechanics and Engineering
51 (1985), 177-208.

[TBW76] R.L. Taylor, P.J. Beresford, E.L. Wilson, A nonconforming element for stress analysis, Inter-
national Journal for Numerical Methods in Engineering 10 (1976), 1211-1219.

[WR96] P. Wriggers, S. Reese, A note on enhanced strain methods for large deformations, Computer
Methods in Applied Mechanics and Engineering 135 (1996), 201-209.

[ZTT71]O.C. Zienkiewicz, R.L. Taylor, J.M. Too, Reduced integration technique in general analysis of
plates and shells, International Journal for Numerical Methods in Engineering 3 (1971), 275-290.

90
[ZTZ05] Zienkiewicz, Olek C.; Taylor, Robert L.; Zhu, Jian Z., The finite element method: its basis
and fundamentals. Elsevier, 2005.

91