11

Variational
Formulation of
Bar Element

11–1
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–2

TABLE OF CONTENTS

Page
§11.1. A New Beginning 11–3
§11.2. Definition of Bar Member 11–3
§11.3. Variational Formulation 11–4
§11.3.1. The Total Potential Energy Functional . . . . . . . . . 11–4
§11.3.2. Admissible Variations . . . . . . . . . . . . . . 11–6
§11.3.3. The Minimum Total Potential Energy Principle . . . . . . 11–6
§11.3.4. TPE Discretization . . . . . . . . . . . . . . . 11–7
§11.3.5. Bar Element Discretization . . . . . . . . . . . . . 11–8
§11.3.6. Interpolation by Shape Functions . . . . . . . . . . 11–9
§11.3.7. The Strain-Displacement Matrix . . . . . . . . . . . 11–9
§11.3.8. *Trial Basis Functions . . . . . . . . . . . . . . 11–9
§11.4. The Finite Element Equations 11–10
§11.4.1. The Stiffness Matrix . . . . . . . . . . . . . . . 11–10
§11.4.2. The Consistent Node Force Vector . . . . . . . . . . 11–11
§11.5. Weak Forms 11–13
§11.5.1. From Strong to Weak . . . . . . . . . . . . . . . 11–13
§11.5.2. Weak Form Based Approximation Example . . . . . . . 11–14
§11.5.3. Balanced Weak Forms . . . . . . . . . . . . . . 11–15
§11.5.4. Principle of Virtual Work as Balanced Weak Form . . . . . 11–15
§11.5.5. *Weighted Residual Methods . . . . . . . . . . . . 11–16
§11.6. *Accuracy Analysis 11–16
§11.6.1. *Nodal Exactness and Superconvergence . . . . . . . . 11–16
§11.6.2. *Fourier Patch Analysis . . . . . . . . . . . . . . 11–17
§11.6.3. *Robin Boundary Conditions . . . . . . . . . . . 11–18
§11. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 11–18
§11. References. . . . . . . . . . . . . . . . . . . . . . 11–19
§11. Exercises . . . . . . . . . . . . . . . . . . . . . . 11–20

11–2
11–3 §11.2 DEFINITION OF BAR MEMBER

§11.1. A New Beginning

This Chapter begins Part II of the course. This Part focuses on the construction of structural and
continuum finite elements using a variational formulation based on the Total Potential Energy. Why
only elements? Because the other synthesis steps of the DSM: globalization, merge, BC application
and solution, remain the same as in Part I. Those operations are not element dependent.
Individual elements are constructed in this Part beginning with the simplest ones and progressing
to more complicated ones. The formulation of 2D finite elements from a variational standpoint is
discussed in Chapters 14 and following. Although the scope of that formulation is broad, exceeding
structural mechanics, it is better understood by going through specific elements first.
From a geometrical standpoint the simplest finite elements are one-dimensional or line elements.
This means that the intrinsic dimensionality is one, although these elements may be used in one,
two or three space dimensions upon transformation to global coordinates as appropriate. The
simplest one-dimensional structural element is the two-node bar element, which we have already
encountered in Chapters 2, 3 and 5 as the truss member.
In this Chapter the bar stiffness equations are rederived using the variational formulation. For
uniform properties the resulting equations are the same as those found previously using the physical
or Mechanics of Materials approach. The variational method has the advantage of being readily
extendible to more complicated situations, such as variable cross section or more than two nodes.

(a) (b)
y

;;
x
cross section axial rigidity EA

;;
u(x)
;;
;;

z q(x) P
P
x cross
section
Longitudinal axis L

Figure 11.1. A fixed-free bar member: (a) 3D view showing reference frame; (b) 2D view on {x, y} plane
highlighting some quantities that are important in bar analysis.

§11.2. Definition of Bar Member

In structural mechanics a bar is a structural component characterized by two properties:

(1) One preferred dimension: the longitudinal dimension or axial dimension is much larger that the
other two dimensions, which are collectively known as transverse dimensions. The intersection
of a plane normal to the longitudinal dimension and the bar defines the cross sections. The
longitudinal dimension defines the longitudinal axis. See Figure 11.1(a).
(2) The bar resists an internal axial force along its longitudinal dimension.

11–3
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–4

Table 11.1 Nomenclature for Mathematical Model of Axially Loaded Bar

Quantity Meaning
x Longitudinal bar axis∗
(.) d(.)/d x
u(x) Axial displacement
q(x) Distributed axial force, given per unit of bar length
L Total bar length
E Elastic modulus
A Cross section area; may vary with x
EA Axial rigidity
e = du/d x = u  Infinitesimal axial strain
σ = Ee = Eu  Axial stress
F = Aσ = E A e = E Au  Internal axial force
P Prescribed end load
∗
x is used in this Chapter instead of x̄ (as in Chapters 2–3) to simplify the notation.

In addition to trusses, bar elements are used to model cables, chains and ropes. They are also used
as fictitious elements in penalty function methods, as discussed in Chapter 9.
We will consider here only straight bars, although their cross section may vary. Our one-dimensional
mathematical model assumes that the bar material is linearly elastic obeying Hooke’s law, and that
displacements and strains are infinitesimal. Figure 11.1(b) pictures some relevant quantities for a
fixed-free bar. Table 11.1 collects the necessary terminology for the governing equations.
Figure 11.2 displays the governing equations of the bar in a graphic format called a Tonti diagram.
The formal similarity with the diagrams used in Chapter 5 to explain MoM elements should be
noted, although the diagram of Figure 11.2 pertains to the continuum bar model rather than to the
discrete one. (The qualifier “strong form” is explained in the next Chapter.)

§11.3. Variational Formulation

To illustrate the variational formulation, the finite element equations of the bar will be derived from
the Minimum Potential Energy principle.

§11.3.1. The Total Potential Energy Functional

In Mechanics of Materials it is shown that the internal energy density at a point of a linear-elastic
material subjected to a one-dimensional state of stress σ and strain e is U = 12 σ (x)e(x), where
σ is to be regarded as linked to the displacement u through Hooke’s law σ = Ee and the strain-
displacement relation e = u  = du/d x. This U is also called the strain energy density. Integration
over the volume of the bar gives the total internal energy
  L  L  L
 
U= 1
2
σ e dV = 1
2
Fe d x = 1
2
(E Au )u d x = 1
2
u  E A u  d x. (11.1)
V 0 0 0

11–4
11–5 §11.3 VARIATIONAL FORMULATION

Displacement
Prescribed Axial Distributed
end BCs displacement axial load
displacement u(x) q(x)

Kinematic e=u' F'+q=0 Equilibrium

Axial F = EA e Axial Force BCs Prescribed

strain force end load
e(x) Constitutive F(x)

Figure 11.2. Strong-form Tonti diagram for the continuum model of a bar member. Field equations and BCs
are represented as lines connecting the boxes. Yellow (brown) boxes contain unknown (given) quantities.

All integrand quantities in (11.1) may depend on x.

The external work potential is the work performed by applied mechanical loads working on the
bar displacements. This potential is denoted by W . (The external energy V is the negative of the
work potential: V = −W . In the ensuing derivations W will be used instead of V .) It collects
contributions from two sources:
1. The distributed load q(x). This contributes a cross-section density of q(x)u(x) because q is
assumed to be already integrated over the section.
2. Any specified axial point load(s). For the fixed-free example of Figure 11.1 the end load P
would contribute P u(L).
The second source may be folded into the first by conventionally writing any point load P acting at
a cross section x = a as a contribution P δ(a) to q(x), in which δ(a) denotes the one-dimensional
Dirac delta function at x = a. If this is done the external energy can be concisely expressed as
 L
W = q u d x. (11.2)
0

The total potential energy of the bar is given by

=U −W (11.3)

Mathematically is a functional, called the Total Potential Energy functional or TPE. It depends
only on the axial displacement u(x). In Variational Calculus u(x) is called the primary variable
of the functional. When the dependence of on u needs to be emphasized we shall write [u] =
U [u]−W [u], with brackets enclosing the primary variable. To display both primary and independent
variables we write, for example, [u(x)] = U [u(x)] − W [u(x)].

Remark 11.1. According to the rules of Variational Calculus, the Euler-Lagrange equation for is

∂ d ∂
E= − = −q − (E A u  ) (11.4)
∂u d x ∂u 

11–5
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–6

δu(x)
u(x)+δu(x) δu(L)
u
u(L)
u(x)
x
u(0) = 0
x
Figure 11.3. Concept of admissible variation of the axial displacement function u(x). For convenience u(x) is
plotted normal to the longitudinal axis. Both u(x) and u(x) + δu(x) shown above are kinematically admissible,
and so is the variation δu(x). Note that the variation δu(L) is not zero because the BC at x = L is natural.

The stationary condition for is E = 0, or

(E A u  ) + q = 0 (11.5)
This is the strong (pointwise) equation of equilibrium in terms of the axial displacement, which reduces to
E A u  + q = 0 if E A is constant. This equation is not explicitly used in the FEM development. It is instead
replaced by δ = 0, with the variation restricted over the class of finite element interpolation functions.

§11.3.2. Admissible Variations

The concept of admissible variation is fundamental in both variational calculus and the variationally
formulated FEM. Only the primary variable(s) of a functional may be varied. For the TPE functional
(11.5) this is the axial displacement u(x). Suppose that u(x) is changed to u(x) + δ u(x).1 This is
illustrated in Figure 11.3, where for convenience u(x) is plotted normal to x. The TPE functional
changes accordingly as
= [u] ⇒ + δ = [u + δu]. (11.6)
The function δ u(x) and the scalar δ are called the variations of u(x) and , respectively. The
variation δ u(x) should not be confused with the ordinary differential du(x) = u  (x) d x since on
taking the variation the independent variable x is frozen; that is, δx = 0.
A displacement variation δu(x) is said to be admissible when both u(x) and u(x) + δ u(x) are
kinematically admissible in the sense of the Principle of Virtual Work (PVW). This agrees with the
conditions of classical variational calculus, and are restated next.
A kinematically admissible axial displacement u(x) obeys two conditions:
(i) It is continuous over the bar length, that is, u(x) ∈ C 0 in x ∈ [0, L].
(ii) It satisfies exactly any displacement boundary condition, such as the fixed-end specification
u(0) = 0 of Figure 11.1. See of Figure 11.3.
The variation δ u(x) pictured in Figure 11.3 is kinematically admissible because both u(x) and
u(x) + δ u(x) satisfy the foregoing conditions. Note that the variation δu(L) at the free end x = L
is not necessarily zero because that boundary condition is natural; that is, not specified directly in
terms of the displacement u(L). On the other hand, δ(0) = 0.
The physical meaning of conditions (i)–(ii) is the subject of Exercise 11.1.
1 The symbol δ not immediately followed by a parenthesis is not a delta function but instead denotes variation with respect
to the variable that follows.

11–6
11–7 §11.3 VARIATIONAL FORMULATION

u 1, f1 u 2, f2 u 3, f3 u 4, f4 u 5, f5

(1) (2) (3) (4)

1 2 3 4 5
u u(x)
u5
u2 u3 u4
x
u1 = 0
Figure 11.4. FEM discretization of bar member. A piecewise- linear admissible displacement trial
function u(x) is drawn underneath the mesh. It is assumed that the left end is fixed; thus u 1 = 0.

§11.3.3. The Minimum Total Potential Energy Principle

The Minimum Total Potential Energy (MTPE) principle states that the actual displacement solution
u ∗ (x) that satisfies the governing equations is that which renders stationary:2

δ = δU − δW = 0 iff u = u∗ (11.7)

with respect to admissible variations u = u ∗ + δu of the exact displacement field u ∗ (x).

Remark 11.2. Using standard techniques of variational calculus3 it can be shown that if E A > 0 and kinematic
boundary conditions weed out any rigid motions, the solution u ∗ (x) of (11.7) exists, is unique, and renders
[u] a minimum over the class of kinematically admissible displacements. The last attribute explains the
“minimum” in the name of the principle.

§11.3.4. TPE Discretization

To apply the TPE functional (11.3) to the derivation of FEM equations we replace the contin-
uum mathematical model by a discrete one consisting of a union of bar elements. For example,
Figure 11.4 illustrates the subdivision of a fixed-free bar member into four two-node elements.
Functionals are scalars. Therefore, for a discretization such as that shown in Figure 11.4, the TPE
functional (11.6) may be decomposed into a sum of contributions of individual elements:

(1) (2) (Ne )

= + + ... + (11.8)
in which Ne denotes the number of elements. The same decomposition applies to both its internal
energy and external work potential components:

δU = δU (1) + . . . + δU (Ne ) = 0, δW = δW (1) + . . . + δW (Ne ) = 0, (11.9)

as well as to the stationarity condition (11.7):

(1) (2) (Ne )
δ =δ +δ + ... + δ = 0. (11.10)

2 The symbol “iff” in (11.7) is an abbreviation for “if and only if”.
3 See references in Notes and Bibliography at the end of Chapter.

11–7
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–8

_
(a) u e (x) (b1) (b2)
u2 u2 u2

u1 u1 u1
= +
_
1 (e) 2 x
_
x = xγ − x1 (b3)
= Le
= u1 1 + 1 u2
(c) _
1−x/
1 N1e
0
_
x/
N2e 1
0

Figure 11.5. A two-node, TPE-based bar element: (a) element configuration and axial displacement
variation (plotted normal to element axis for better visibility); (b1,b2,b3) displacement interpolation
expressed in terms of linear shape functions; (c) element shape functions.

Using the fundamental lemma of variational calculus,4 it can be shown that (11.10) implies that for
a generic element e we may write

δ e
= δU e − δW e = 0. (11.11)
This variational equation is the basis for the derivation of element stiffness equations once the
displacement field has been discretized over the element, as described next.
Remark 11.3. In mathematics (11.11) is called a first variation form. It is a special case of a more general
expression called a weak form, which is covered in more detail later. In mechanics it states the Principle of
Virtual Work or PVW for each element: δU e = δW e , which says that the virtual work of internal and external
forces on admissible displacement variations is equal if the element is in equilibrium [316].

§11.3.5. Bar Element Discretization

Figure 11.5(a) depicts a generic bar element e. It has two nodes, which are labeled 1 and 2. These
are called the local node numbers.5 The element is referred to its local axis x̄ = x − x1 , which
measures the distance from its left end. The two degrees of freedom are u e1 and u e2 . (Bars are not
necessary since the directions of x̄ and x are the same.) The element length is = L e .
The mathematical concept of bar finite elements is based on approximating axial displacement u(x)
over the element. The exact displacement u ∗ is replaced by an approximate displacement
u ∗ (x) ≈ u e (x) (11.12)
4 See, e.g., Chapter II of Gelfand and Fomin [161].
5 Note the notational change from the labels i and j of Part I. This will facilitate transition to multidimensional elements
in Chapters 14ff.

11–8
11–9 §11.3 VARIATIONAL FORMULATION

over the finite element mesh. This approximate displacement, u e (x), taken over all elements e =
1, 2, . . . N e , is called the finite element trial expansion or simply trial expansion. See Figure 11.4.
This FE trial expansion must belong to the class of kinematically admissible displacements defined
in ?. Consequently, it must be C 0 continuous over and between elements. The most common
choices fpr u e are polynomials in x, as in the development that follows.

§11.3.6. Interpolation by Shape Functions

In a two-node bar element the only possible polynomial choice of the displacement u e that satisfies
the interelement continuity requirement is linear. It can be expressed by the following interpolation
formula, which is graphically developed in Figure 11.5(b1,b2,b3):
 
ue
u (x) =
e
N1e u e1 + N2e u e2 =[ N1e N2e ] 1e = Ne ue . (11.13)
u2

The functions N1e and N2e that multiply the node displacements u 1 and u 2 are called shape functions,
while N is called the shape function matrix. In this case Ne reduces to a row vector.
The shape functions interpolate the internal displacement u e directly from the node values. They
are pictured in Figure 11.5(c). For this element, with x̄ = x − x1 measuring the axial distance from
the left node i, the shape functions are

x̄ x̄
N1e = 1 − = 1 − ζ, N2e = = ζ. (11.14)

Here
x − x1 x̄
ζ = = , (11.15)

is a dimensionless coordinate, also known as a natural coordinate, that varies from 0 through 1
over the element. Note that d x = dζ and dζ = d x/ . The shape function N1e has the value 1 at
node 1 and 0 at node 2. Conversely, shape function N2e has the value 0 at node 1 and 1 at node 2.
This is a general property of shape functions. It follows from the fact that element displacement
interpolations such as (11.13) are based on physical node values.

§11.3.7. The Strain-Displacement Matrix

The axial strain associated with the trial function u e is

du e  dNe dNe   e  1  
e  u1 ue
e= = (u ) = 1 2
e = [ −1 1 ] 1e = B ue , (11.16)
dx dx dx u 2 u2

in which
1
B= [ −1 1], (11.17)

is called the strain-displacement matrix. Note that B is constant over the element.

11–9
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–10

(a) (b)
u N3
(thick line)
(1) (2) (3) (4) N3(2) N3(3) x
1 2 3 4 5
Figure 11.6. Trial basis function (a.k.a. hat function) for node 3 of a four-element bar discretization.

§11.3.8. *Trial Basis Functions

Shape functions are associated with elements. A trial basis function, or simply basis function, is associated
with a node. Suppose node i of a bar discretization connects elements (e1) and (e2). The trial basis function
Ni is defined as  (e1)
Ni if x ∈ element (e1)
Ni (x) = N (e2) if x ∈ element (e2) (11.18)
i
0 otherwise
For a piecewise linear discretization, such as that used in the two-node bar, this function has the shape of a
hat. Thus it is also called a hat function or chapeau function. See Figure 11.6, in which i = 3, e1 = 2, and
e2 = 3. The concept is important in the variational interpretation of FEM as a Rayleigh-Ritz method.

Remark 11.4. In addition to continuity, shape and trial functions must satisfy a completeness requirement
with respect to the governing variational principle. This condition is stated and discussed in later Chapters.
Suffices for now to say that the shape functions (11.14), as well as the associated trial functions, do satisfy this
requirement.

§11.4. The Finite Element Equations

In linear FEM the discretization process based on the TPE functional leads to the following algebraic
form in the node displacements
def 1 e T def
e
= U e − W e, in which U e = 2
(u ) Ke ue and W e = (ue )T fe . (11.19)
Here Ke and fe are called the element stiffness matrix and the element consistent nodal force vector,
respectively. The three scalars e , U e and W e are only function of the node displacements ue . (This
is a consequence of displacements being the only primary variable of the TPE functional.) Note
that U e and W e depend quadratically and linearly, respectively, on those displacements. Taking
the variation of e with respect to the node displacements gives6
∂ e  e T e e
e T
δ = δu e
= δu K u − fe = 0. (11.20)
∂u e

Because the variations δue can be arbitrary, the bracketed expression must vanish, which yields

Ke ue = fe . (11.21)
These are the familiar element stiffness equations. Hence the foregoing names given to Ke and fe
are justified a posteriori.
6 The 12 factor disappears on taking the variation because U e is quadratic in the node displacements. For a review on the
calculus of discrete quadratic forms, see Appendix D.

11–10
11–11 §11.4 THE FINITE ELEMENT EQUATIONS

§11.4.1. The Stiffness Matrix

We now apply the foregoing expressions to the two-node bar element. Its internal energy U e is
 x2  1
U =2
e 1
e E A e dx = 2
1
e E A e dζ. (11.22)
x1 0

Note that the integration variable x has been changed to the natural coordinate ζ defined in (11.15)
that varies from 0 through 1, whence d x = dζ . This form is symmetrically expanded using the
strain-displacement matrix relation (11.16), by inserting e = e T = (ue )T BT and e = B ue into the
first and second e of (11.22), respectively, to get
 1  1    e
1 −1 1 u
U =2
e 1
(u ) B E A B u dζ = 2
e T T e 1
[ u e1 u e2 ] E A [ −1 1 ] 1e dζ.
0 0 1 u 2
(11.23)
The nodal displacements do not depend on position and can be moved out of the integral. Also
BT E A B = E A BT B since E A is a scalar:
 1  1    e
EA 1 −1 u1
U = 2 (u )
e 1 e T
E A B B dζ u = 2 [ u 1 u 2 ]
T e 1 e e
dζ e . (11.24)
0 0
2 −1 1 u 2
 T
By (11.19) this is expressible as 1
2
ue Ke ue . Since ue is arbitrary, Ke is extracted as
 1  1     1
EA 1 −1 1 1 −1
K =
e
E A B B dζ =
T
dζ = E A dζ.
0 0
2 −1 1 −1 1 0
(11.25)
This is the bar element stiffness matrix. For a homogeneous and prismatic bar of constant rigidity,
1
E A can be moved outside the integral, 0 dζ = 1 and (11.25) collapses to
 
EA 1 −1
K = e
. (11.26)
−1 1

This is the same element stiffness matrix of the prismatic truss member derived in Chapters 2 and
5 by a Mechanics of Materials approach, but now obtained through a variational argument.

§11.4.2. The Consistent Node Force Vector

The consistent node force vector fe defined in (11.19) comes from the element contribution to the
external work potential W :
 x2  1   
 e T 1 1−ζ def  T
W =
e
q u dx = q N u dζ = u
T e
q dζ = ue fe , (11.27)
x1 0 0 ζ

Since ue is arbitrary,
 x2    1  
1−ζ 1−ζ
f = e
q dx = q dζ. (11.28)
x1 ζ 0 ζ

11–11
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–12

;;;;;;;
(b) Load case I: point load P at x=L
2 P
1 1
(a)
;;; (c) Load case II: q(x) varies linearly
constant EA
from q1 at x=0 through q2 at x=L
x 1 1 2
L
(d) Load case III: q(x)=q0 (constant)
from x=0 through x=L/2, else 0
1 1 2

Figure 11.7. Fixed-free, prismatic bar example: (a) configuration; (b,c,d) FEM discretization and load cases.

in which ζ is defined by (11.15). If q is constant over the element, it may be taken out of the
integral:
 1 
1−ζ
f =q
e
dζ. (11.29)
0 ζ

This gives the same results as with the EbE lumping method of Chapter 7. See Exercise 11.3.

Example 11.1. The two-node bar element is tested on the benchmark problem defined in Figure 11.7. A fixed-
free, homogeneous, prismatic bar of length L, elastic modulus E and cross section area A has the configuration
illustrated in Figure 11.7(a). It is discretized with a single element as shown in Figure 11.7(b,c,d), and subjected
to the three load cases pictured there. Case I involves a point load P at the free end, which may be formally
represented as
q I I (x) = P δ(L) (11.30)
where δ() denotes the delta function with argument x.
Case II involves a distributed axial load that varies linearly from q1 = q(0) at the fixed end through q2 = q(L)
at the free end:
q I I (x) = q1 (1 − ζ ) + q2 ζ. (11.31)
Case III involves a distributed axial load q(x) that is constant and equal to q0 from the fixed end x = 0 through
midspan x = L/2, and zero from x = L/2 through the free end x = L:

q I I I (x) = q0 H (x) − H (x − 12 L) , (11.32)

in which H () denotes the Heaviside unit step function with argument x. The master stiffness equations
constructed using the prismatic stiffness matrix (11.26) with = L and x̄ → x are

EA  1 −1

u m1
 
f 1m

= = fm , (11.33)
L −1 1 u m2 f 2m

in which supercript m identifies the load case. The consistent node forces computed from (11.28) with = L
and x̄ → x are  
0 L  2q1 + q2  q0 L  3 
fI = , fI I = , fI I I = . (11.34)
P 6 q1 + 2q2 8 1

11–12
11–13 §11.5 WEAK FORMS

On applying the fixed end support condition u m1 = 0 and solving for u m2 , the free end deflections are

PL (q1 + 2q2 )L 2 q0 L 2
u 2I = , u 2I I = , u 2I I I = . (11.35)
EA 6E A 8E A
The analytical solutions for u(x), obtained on integrating the ODE E Au  + q = 0 with boundary conditions
u(0) = 0, F(L) = E A u  (L) = P for case I and F(L) = E A u  (L) = 0 for cases II and III, are
Px x [3(q1 +q2 )L 2 −3q1 L x+(q1 −q2 )x 2 ] q0 
u I (x) = , u I I (x) = , u I I I (x) = L x−x 2 +x− 12 L2 .
EA 6E A 2E A
(11.36)
In the expression of u I I I (x), x − 12 L2 means (x − 12 L)2 if x ≥ 12 L, else zero (Macauley’s bracket notation
for discontinuity functions). Evaluating (11.36) at x = L and comparing to (11.35), one finds that the three
computed end deflections are exact.
For case I this agreement is obvious: the exact u I (x) is linear in x, which is contained in the span of the linear
shape functions. But for II and III this is far from obvious since the exact solutions are cubic and piecewise
quadratic, respectively, in x. The fact that the exact solution is verified at the free end node is an instance of
the nodal exactness property discussed in §11.6.1.
Note that in cases II and III the FEM displacement solutions within the element, which vary linearly, will
not agree pointwise with the exact solutions, which are not. For example the exact u I I I (x) at midspan is
u I I I ( 12 L) = q0 L 2 /(8E A) = u I I I (L), whereas the FEM interpolation would give q0 L 2 /(16E A), in error by
100%. To reduce such internal discrepancies the member should be divided into more elements.

§11.5. Weak Forms

Weak forms are expressions notoriously difficult to explain to newcomers. They occupy an inter-
mediate position between differential equations and functionals. There are so many variants and
kinks, however, that their position in the mathematical food chain is fuzzy. Confusion is further
compounded by the use of disparate terminology, some generic, some application oriented. To shed
some sunlight into the murky swamp, we go through a specific example: the bar member.

§11.5.1. From Strong to Weak

The
 governing differential equation for a bar member in terms of axial displacements is

E A u  (x) + q(x) = 0, or E A u  (x) + q(x) = 0 if the rigidity E A is constant. Replace the
zero by r (x), which stands for residual, and move it to the left-hand side:
 
r (x) = E A u  (x) + q(x), or if E A is constant: r (x) = E A u  (x) + q. (11.37)

The governing ODE may be compactly stated as r (x) = 0. This must hold at each point over the
member span, say x ∈ [0, L]. Hence the term strong form (SF) used for this kind of mathematical
model. No ambiguity so far. But suppose that insisting on r (x) = 0 everywhere is too demanding.
We would like to relax that condition so it is satisfied only in an average sense. To accomplish that,
multiply the residual by a function v(x), integrate over the problem domain, and set the result to
zero:  L
J= r (x) v(x) d x = 0. (11.38)
0
Here v(x) is supposed to be sufficiently well behaved for the integral to exist. Ignoring boundary
conditions for now, (11.38) is called a weak form, which is often abbreviated to WF in the sequel.

11–13
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–14

Function v(x) receives two names in the literature: test function in a general mathematical context,
and weight function (also weighting function) in the context of approximation methods based on
weak forms. In what follows both terms will be used within the appropriate context.

§11.5.2. Weak Form Based Approximation Example

To show how weak forms can be used to generate an approximate solution, consider again a fixed-
free, prismatic homogeneous bar member (that is, E A is constant), under uniform load q(x) = q0
along its length. The WF (11.38) becomes
 L

J= E A u  (x) + q0 v(x) d x = 0. (11.39)
0

subject to the end conditions

u(0) = 0, F(L) = E A u  (L) = 0. (11.40)
We will restrict both u(x) and v(x) to be quadratic polynomials:
u(x) = a0 + a1 x + a2 x 2 , v(x) = b0 + b1 x + b2 x 2 . (11.41)
in which ai and bi are numerical coefficients, real in this case. Once assumptions such as those
in (11.41) are made, new terminology kicks in. The assumed u(x) is now called a trial function,
which is here spanned by the linear-space basis {1, x, x 2 } of dimension 3. The assumed v(x) is
called a weight function, which is here spanned by exactly the same basis. There is a special name
for the scenario when the trial and weight function bases coalesce: the Galerkin method.7 . We will
call the end result a Galerkin solution. Replacing (11.41) into (11.39) we get
L
J= (6b0 + 3b1 L + 2b2 L 2 )(2E A a2 + q0 ). (11.42)
6
Now J = 0 for any arbitrary value of {b0 , b1 , b2 }. On extracting the expressions that multiply
those coefficients we obtain the same equation thrice: 2E A a2 + q0 = 0. Thus a2 = −q0 /(2E A),
whereas a0 and a1 remain arbitrary. Consequently the Galerkin solution before BC is
q0 2
u(x) = a0 + a1 x − x . (11.43)
2E A
ODE aficionados would recognize this as the general solution of E Au  + q0 = 0 so Uncle Boris
has done the job. Applying the end conditions (11.40) gives a0 = 0 and a1 = q0 /(E A) whence the
final solution is q0
u(x) = x(2L − x). (11.44)
2E A
Replacing this into the governing ODE it may be verified that this is the exact analytical solution.
Instead of applying the end coditions a posteriori we may try to incorporate them a priori into
the trial function assumption. On enforcing (11.40) into the assumed u(x) of (11.41) we find that
a0 = 0 and a1 = −2a2 L. The trial function becomes
u(x) = a2 x (x − 2L), (11.45)
7 For a historical account of such terminology, see Notes and Bibliography

11–14
11–15 §11.5 WEAK FORMS

and only one free coefficient remains. Accordingly only one weight basis function is needed: either
1, x or x 2 does the job, and the exact solution (11.44) is obtained again.8
Even for this fairly trivial example, several procedural choices are apparent, although the final
answer is unaffected. If we allow the trial and weight function spaces to differ, volatility zooms up.
Furthermore, we can apply transformations to the residual integral as done in the next subsection.
Compared to the well ordered world of variational-based FEM, confusion reigns supreme.

§11.5.3. Balanced Weak Forms

Some method in the madness can be injected by balancing. A look at (11.39) reveals an unpleasant
asymmetry. Second derivatives of u(x) appear, but none of v(x). This places unequal restrictions
on smoothness of the trial and test function spaces. Integration by parts restores derivative order
L
balance. Replacing 0 E A u  v d x = − 0 E A u  v  d x + (E Au  )v 0 and rearranging terms yields
L L

 L  L
 

q(x) v(x) d x − E A u  (x) v(x) 0 .
L
J= E A u (x) v (x) d x − (11.46)
0 0

This will be called a balanced-derivative weak form, or simply a balanced weak form (BWF). It
displays obvious advantages: (i) same smoothness requirements for assumed u and v, and (ii) end
BC appear explicitly in the non-integral term, neatly factored into essential and natural. A minor
flaw is that the original residual is no longer clearly visible.
For a bar with variable axial rigidity replace E A u  by (E A u  ) in the first integrand.
On repeating the Galerkin procedure of the previous subsection with the assumptions (11.41) one
finds an identical J , as may be expected, and the same final solution. Again one has the choice of
pre- or post-imposing the end conditions (11.40). Generally the latter choice is far more convenient
in a computer implementation.

§11.5.4. Principle of Virtual Work as Balanced Weak Form

There is a close connection between the BWF (11.46) and one of the fundamental tools of Analytical
Mechanics: the Principle of Virtual Work (PVW). Set the test function to be an admissible variation
of u(x): v(x) = δ u(x), in which δu(x) strongly satisfies all essential boundary conditions. Then
assume that J is the first variation of a functional :
 L  L
 

q(x) δu(x) d x − E A u  (x) δu(x) 0 = δ .
L def
J= E A u (x) δu (x) d x − (11.47)
0 0

But this is in fact the first variation of the TPE functional:

 L  L
 
=U −W = 2 1
u (x) E A u (x) d x − q(x) u(x) d x (11.48)
0 0

Hence J = 0 is the same as δ = 0 or δU = δW , which is the PVW for an elastic bar member.
This relationship can be used to prove an important property: Galerkin method is equivalent to a
variational formulation if the residual is the Euler-Lagrange equation of a functional.

8 Some early works covering weighted residual methods, for example Crandall [81], proclaim that the trial function must
satisfy all BC ab initio. Later ones, e.g., Finlayson [136], relax that to BC of essential type. Both rules are nonsense. As
the DSM implementation of FEM makes clear, BC application can wait after the discrete equations are assembled.

11–15
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–16
 L
Remark 11.5. Where does the boundary term E A u  (x) δu(x) 0 in (11.47) go? Actually, into δW . This
immersion is a bit tricky, and depends on redefining q(x) to include prescribed end point forces such as
N (L) = E A u  (L) = P through delta functions. This is the subject of Exercise 11.8.

§11.5.5. *Weighted Residual Methods

Galerkin method is widely used in computational mechanics, but does not exhaust all possibilities of using a
weak form as source for obtaining numerical solutions. The main generalization consist of allowing trial and
test (weight) functions to be different. This leads to a rich class of approximation methods unified under the
name Method of Weighted Residuals or MWR.
The key idea is as follows. Both u(x) and v(x) are restricted to belong to linear function spaces of finite
dimension Nu and Nv . These are the trial function space and the test function space, respectively. which are
spanned by basis functions φi (x) and ψ(x), respectively:
   
u(x) = span φi (x), 1 ≤ i ≤ Nu , v(x) = span ψi (x), 1 ≤ i ≤ Nv (11.49)

in which usually Nu = Nv . Since the spaces are linear, any u(x) and v(x) can be represented as linear
combination of the basis functions:

Nu 
Nv
u(x) = ai φi (x), v(x) = bi ψi (x). (11.50)
i=1 i=1

Here ai and bi are scalar coefficients, which may be real or complex depending on the nature of the problem.
Insert these into the weak form, perform the necessary integrations, and extract the Nv expressions that are
coefficients of the bi . Solve these equations for the coefficient ai , and replace in (11.50) to get the approximate
solution u(x).
The MWR methodology is of course not restricted to one dimension, and also extends to time-dependent
problems. It can be merged smoothly with the FEM concept of piecewise approximation using shape functions.
Some references are provided under Notes and Bibliography.

§11.6. *Accuracy Analysis

Low order 1D elements may give surprisingly high accuracy. In particular the lowly two-node bar element
can display infinite accuracy under some conditions. This phenomenon is studied in this advanced section as
it provides an introduction to modified equation methods and Fourier analysis along the way.

§11.6.1. *Nodal Exactness and Superconvergence

Suppose that the following two conditions are satisfied:
1. The bar properties are constant along the length (prismatic member).
2. The distributed load q(x) is zero between nodes. The only applied loads are point forces at the nodes.
If so, a linear axial displacement u(x) as defined by (11.13) and (11.14) is the exact solution over each element
since constant strain and stress satisfy, element by element, all of the governing equations listed in Figure 11.2.9
It follows that if the foregoing conditions are verified the FEM solution is exact; that is, it agrees with the
analytical solution of the mathematical model.10 Adding extra elements and nodes would not change the

9 The internal equilibrium equation p  + q = E A u  + q = 0 is trivially verified because q = 0 from the second
assumption, and u  = 0 because of shape function linearity.
10 In variational language: the Green function of the u  = 0 problem is included in the FEM trial space.

11–16
11–17 §11.6 *ACCURACY ANALYSIS

(a)
(b)
q(x)
Trial basis function Nj
for node j
xi = xj −
i j k 1 − (xj −x)/ = 1−ψ 1 + (x j−x)/ = 1+ψ
EA = const 1
xj Two-element
xk = xj + patch ijk
L i j k

Figure 11.8. Superconvergence patch analysis: (a) lattice of bar elements; (b) two element patch.

solution. That is the reason behind the truss discretizations used in Chapters 2–3: one element per member is
enough if they are prismatic and loads are applied to joints. Such models are called nodally exact.
What happens if the foregoing assumptions are not met? Exactness is then generally lost, and several elements
per member may be beneficial if spurious mechanisms are avoided.11 For a 1D lattice of equal-length, prismatic
two-node bar elements, an interesting and more difficult result is: the solution is nodally exact for any loading
if consistent node forces are used. This is proven in the subsection below. This result underlies the importance
of computing node forces correctly.
If conditions such as equal-length are relaxed, the solution is no longer nodally exact but convergence at the
nodes is extremely rapid (faster than could be expected by standard error analysis) as long as consistent node
forces are used. This phenomenon is called superconvergence in the FEM literature.

§11.6.2. *Fourier Patch Analysis

The following analysis is based on the modified differential equation (MoDE) method of Warming and Hyett
[406] combined with the Fourier patch analysis approach of Park and Flaggs [292,293]. Consider a lattice of
two-node prismatic bar elements of constant rigidity E A and equal length , as illustrated in Figure 11.8. The
total length of the lattice is L. The system is subject to an arbitrary axial load q(x). The only requirement on
q(x) is that it has a convergent Fourier series in the space direction.
From the lattice extract a patch12 of two elements connecting nodes xi , x j and xk as shown in Figure 11.8. The
FEM patch equations at node j are
 
EA ui
[ −1 2 −1 ] uj = fj, (11.51)
uk
in which the node force f j is obtained by consistent lumping:
 xk  0  1
fj = q(x)N j (x) d x = q(x j + ψ )(1 + ψ) dψ + q(x j + ψ )(1 − ψ) dψ. (11.52)
xi −1 0

Here N j (x) is the “hat” trial basis function for node j, depicted in Figure 11.8, and ψ = (x − x j )/ is a
dimensionless coordinate that takes the values −1, 0 and 1 at nodes i, j and k, respectively. If q(x) is expanded
in Fourier series

M
q(x) = qm eiβm x , βm = mπ/L , (11.53)
m=1

11 These can happen when transforming such elements for 2D and 3D trusses. See Exercise E11.7.
12 A patch is the set of all elements connected to a node; in this case j.

11–17
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–18

(the term m = 0 requires special handling) the exact solution of the continuum equation E A u  + q = 0 is

M
qm eiβm x
u ∗ (x) = u ∗m eiβm x , u ∗m = . (11.54)
m=1
E Aβm2

Evaluation of the consistent force using (11.52) gives


M
sin2 ( 12 βm )
fj = f jm , f jm = qm eiβm x2 . (11.55)
m=1
β
1 2 2
4 m

To construct a modified differential equation (MoDE), expand the displacement by Taylor series centered at
node j. Evaluate at i and k: u i = u j − u j + 2 u j /2! − 3 u j /3! + 4 u ivj /4! + . . . and u k = u j + u j +
2 
u j /2 + 3 u j /3! + 4 u ivj /4! + . . .. Replace these series into (11.51) to get
 2 4

1 
−2E A u j + u ivj + u vij + . . . = fj. (11.56)
2! 4! 6!
This is an ODE of infinite order. It can be reduced to an algebraic equation by assuming that the response of
(11.56) to qm eiβm x is harmonic: u jm eiβm x . If so u jm = −βm2 u jm , u ivjm = βm4 u jm , etc, and the MoDE becomes
 
1 β2 2 β4 4 sin2 ( 12 βm ) iβm x j
2E A βm2 − m + m − . . . u jm = 4E A sin2 ( 12 βm ) u jm = f jm = qm e .
2! 4! 6! β
1 2 2
4 m
(11.57)
Solving gives u jm = qm eiβm x j /(E Aβm2 ), which compared with (11.54) shows that u jm = u ∗m for any m > 0.
Consequently u j = u ∗j . In other words, the MoDE (11.56) and the original ODE: E Au  + q = 0 have the
same value at x = x j for any load q(x) developable as (11.53). This proves nodal exactness. In between
nodes the two solutions will not agree.13
The case m = 0 has to be treated separately since the foregoing expressions become 0/0. The response to a
uniform q = q0 is a quadratic in x, and it is not difficult to prove nodal exactness.

§11.6.3. *Robin Boundary Conditions

Suppose that for a bar of length L one has the following end conditions: u  (0) = au(0) + b at x = 0 and
u  (L) = au(L) + b at x = L, in which a and b are given coefficients. Those are called Robin BCs in the
literature. Adjoining them as Courant penalty terms gives the functional
 L
F(u) = [ 12 E A (u  )2 − q u] d x + 12 [u  (0) − au(0) − b]2 + 12 [u  (L) + au(L) + b]2 . (11.58)
0

Divide [0,L] into Ne elements and N = Ne + 1 nodes. Do C 0 linear interpolation over each element, insert
into F(u) to get Fd (u) = 12 uT Ku − fT v, in which u is the vector of node values, K the master stiffness matrix
and f the master force vector. Coefficients a and b will affect both K and f.
Vanishing of the first variation: δ Fd = 0 yields the FEM equations Ku = f to be solved for u. The Robin BCs
at x = 0 and x = L will affect the stiffness and force contributions of the first and last elements, but not those
of interior elements.
This kind of boundary value problem (i.e., with Robin BCs) is common in heat conduction and heat transfer
with convection given over cooling surfaces. In that case the heat flux is proportional to the difference of the
(unknown) surface temperature and that of the cooling fluid. Elements that ”touch” the convention boundary
are affected.

13 The FEM solution varies linearly between nodes whereas the exact one is generally trigonometric.

11–18
11–19 §11. References

Notes and Bibliography

The foregoing account pertains to the simplest structural finite element: the two-node bar element. For bar
members these developments may be generalized in several directions, three of which are mentioned next.
Refined bar models. Adding internal nodes we can pass from linear to quadratic and cubic shape functions.
These elements are rarely useful on their own right, but as accessories to 2D and 3D high order continuum
elements (for example, to model edge reinforcements.) For that reason they are not considered here. The
3-node bar element is developed in exercises assigned in Chapter 16.
Use in 2D and 3D truss structures. The only additional ingredients are the local-to-global transformations
discussed in Chapters 3 and 6.
Curved bar elements. These can be derived using isoparametric mapping, a device introduced later.
Matrices for straight bar elements are available in any finite element book; for example Przemieniecki [321].
Tonti diagrams were introduced in the 1970s in papers now difficult to access, for example [387]. Scanned
images are available, howewer, from http://www.dic.units.it/perspage/discretephysics
The fundamentals of Variational Calculus may be studied in the excellent textbook [161], which is now
available in an inexpensive Dover edition. The proof of the MPE principle can be found in texts on variational
methods in mechanics. For example: Langhaar [234], which is the most readable “old fashioned” treatment
of the energy principles of structural mechanics, with a clear treatment of virtual work. (Out of print but
used copies may be found via the web engines cited in §1.5.2.) The elegant treatment by Lanczos [233] is
recommended as reading material although it is more oriented to physics than structural mechanics.
It was noted that weak forms occupy an intermediate position between two older classical areas: differential
equations (introduced in the XVII Century by the Calculus founders) and variational forms (introduced by
Euler in the XVIII Century). Some weak forms in disguise are also ancient; e.g., the PVW was placed on
firm mathematical grounds by Lagrange in the late XVIII Century [231]. But their rapid development as tools
for producing approximate solutions of ODEs and PDEs took place in the early XIX Century. Five important
variants are: Galerkin (1915), subdomain (1923), least squares (1928), moments (1932), and collocation
(1937). These, as well as a few others of less importance, were unified in 1956 under the label Method of
Weighted Residuals or MWR, by Crandall [81]. Other attempts at unification during this period may be found
in [10,71]. The use of MWR methods, especially Galerkin’s, as enabling devices to generate finite element
equations developed rapidly following the pioneer paper [427]. The chief motivation was to accommodate
application problems where a classical variational formulation does not exist, or is inconvenient to use.
The first accuracy study of FEM discretizations using modified equation methods is by Waltz et. al. [404];
however their procedures were faulty, which led to incorrect conclusions. The first correct derivation of
modified equations appeared in [406]. The topic has recently attracted interest from applied mathematicians
because modified equations provide a systematic tool for backward error analysis of differential equations: the
discrete solution is the exact solution of the modified problem. This is particularly important for the study of
long term behavior of discrete dynamical systems, whether deterministic or chaotic. Recommended references
along these lines are [170,176,365].
Nodal exactness of bar models for point node loads is a particular case of a theorem by Tong [385]. For
arbitrary loads it was proven by Park and Flaggs [292,293], who followed a variant of the scheme of §11.6.2.
A different technique is used in Exercise 11.8. The budding concept of superconvergence, which emerged in
the late 1960s, is outlined in the book of Strang and Fix [360]. There is a monograph [405] devoted to the
subject; it covers only Poisson problems but provides a comprehensive reference list until 1995.

References
Referenced items moved to Appendix R.

11–19
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT 11–20

Homework Exercises for Chapter 11

Variational Formulation of Bar Element

EXERCISE 11.1 [D:10] Explain the kinematic admissibility requirements stated in ? in terms of physics,
namely ruling out the possibility of gaps or interpenetration as the bar material deforms.

EXERCISE 11.2 [A/C:15] Using (11.25), derive the stiffness matrix for a tapered bar element in which the
cross section area varies linearly along the element length:

A = Ai (1 − ζ ) + A j ζ, (E11.1)

where Ai and A j are the areas at the end nodes, and ζ = x e / is the dimensionless coordinate defined in
§11.3.6. Show that this yields the same answer as that of a stiffness of a constant-area bar with cross section
1
2
(Ai + A j ). Note: the following Mathematica script may be used to solve this exercise:14

ClearAll[Le,x,Em,A,Ai,Aj];
Be={{-1,1}}/Le; ζ =x/Le; A=Ai*(1-ζ )+Aj*ζ ;
Ke=Integrate[Em*A*Transpose[Be].Be,{x,0,Le}];
Ke=Simplify[Ke];
Print["Ke for varying cross section bar: ",Ke//MatrixForm];

In this and following scripts Le stands for .

EXERCISE 11.3 [A:10] Find the consistent load vector fe for a bar of constant area A subject to a uniform
axial force q = ρg A per unit length along the element. Show that this vector is the same as that obtained with
the element-by-element (EbE) “lumping” method of §8.4, which simply assigns half of the total load: 12 ρg A ,
1
to each node. Hint: use (11.29) and 0
ζ dζ = /.

EXERCISE 11.4 [A/C:15] Repeat the previous calculation for the tapered bar element subject to a force
q = ρg A per unit length, in which A varies according to (E11.1) whereas ρ and g are constant. Check that if
Ai = A j one recovers f i = f j = 12 ρg A . Note: the following Mathematica script may be used to solve this
exercise:15

ClearAll[q,A,Ai,Aj,ρ,g,Le,x];
ζ =x/Le; Ne={{1-ζ ,ζ }}; A=Ai*(1-ζ )+Aj*ζ ; q=ρ*g*A;
fe=Integrate[q*Ne,{x,0,Le}];
fe=Simplify[fe];
Print["fe for uniform load q: ",fe//MatrixForm];
ClearAll[A];
Print["fe check: ",Simplify[fe/.{Ai->A,Aj->A}]//MatrixForm];

EXERCISE 11.5 [A/C:20] A tapered bar element of length , end areas Ai and A j with A interpolated as
per (E11.1), and constant density ρ, rotates on a plane at uniform angular velocity ω (rad/sec) about node i.
Taking axis x along the rotating bar with origin at node i, the centrifugal axial force is q(x) = ρ Aω2 x along
the length, in which x ≡ x e . Find the consistent node forces as functions of ρ, Ai , A j , ω and , and specialize
the result to the prismatic bar A = Ai = A j . Partial result check: f j = 13 ρω2 A 2 for A = Ai = A j .

14 The ClearAll[...] at the start of the script is recommended programming practice to initialize variables and avoid
“cell crosstalk.” In a Module this is done by listing the local variables after the Module keyword.
15 The ClearAll[A] before the last statement is essential; else A would retain the previous assignation.

11–20
11–21 Exercises

EXERCISE 11.6 [A:15] (Requires knowledge of Dirac’s delta function properties.) Find the consistent load
vector fe if the bar is subjected to a concentrated axial force Q at a distance x = a from its left end. Use
(11.28), with q(x) = Q δ(a), in which δ(a) is the one-dimensional Dirac’s delta function at x = a. Note: the
following script does it by Mathematica, but it is overkill:

ClearAll[Le,q,Q,a,x];
ζ =x/Le; Ne={{1-ζ ,ζ }}; q=Q*DiracDelta[x-a];
fe=Simplify[ Integrate[q*Ne,{x,-Infinity,Infinity}] ];
Print["fe for point load Q at x=a: ",fe//MatrixForm];

EXERCISE 11.7 [C+D:20] In a learned paper, Dr. I. M. Clueless proposes “improving” the result for the
example truss by putting three extra nodes, 4, 5 and 6, at the midpoint of members 1–2, 2–3 and 1–3, respectively.
His “reasoning” is that more is better. Try Dr. C.’s suggestion using the Mathematica implementation of Chapter
4 and verify that the solution “blows up” because the modified master stiffness is singular. Explain physically
what happens.

EXERCISE 11.8 [A:15]. Prove that (11.47) is the first variation of (11.48), thus linking the PVW and the
TPE functional. See Remark 11.5 for a hint on how to treat the boundary term in (11.47).

EXERCISE 11.9 [A:35, close to research paper level]. Prove nodal exactness of the two-node bar element
for arbitrary but Taylor expandable loading without using the Fourier series approach. Hints: expand q(x) =
q(x j ) + ( ψ)q  (x j ) + ( ψ)2 q  (x j )/2! + . . ., where ψ = x − x j is the distance to node j, compute the
consistent force f j (x) from (11.52), and differentiate the MoDE (11.56) repeatedly in x while truncating all
derivatives to a maximum order n ≥ 2. Show that the original ODE: E Au  + q = 0, emerges as an identity
regardless of how many derivatives are kept.

11–21