ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-1

CHAPTER 5
ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS

The advent of two-dimensional finite elements rather than one-dimensional finite

elements marked the formal beginning of the finite element method. Numerical solution
methods for one-dimensional problems were no doubt of little interest to researchers.
Nonetheless, certain mathematical aspects of FEM essential to two- and three-
dimensional finite element formulations, especially shape functions, are best studied first
in one-dimension. They are:
(1) Low and high order shape functions,
(2) Natural coordinate systems,
(3) Geometry mapping and Jacobians,
(4) Conditions for element convergence,
(5) Gauss quadrature integration.
Physical applications of one-dimensional finite elements are studied in the next chapter in
conjunction with both solid mechanics and heat transfer. For now the focus is mainly on
shape functions and the other four aspects of FEM listed.
The classical Rayleigh-Ritz and Galerkin methods and the finite element method
are alternative methods of numerical approximation in solid and structural mechanics.
Whereas the Rayleigh-Ritz and Galerkin methods (Sections 2-5 and 2-6, respectively) are
global methods of approximation, the finite element method is a local method of
approximation (see also Section 2-7). As such, the finite element method assumes certain
approximations that simplify the behavior of a typical subregion of a structure. These
approximations are based on shape functions defined over the subregion. Thus, shape
functions are a primary source of the approximation nature of FEM. The approximation
introduced may be categorized as discretization error (Section 1-8). Moreover, they are
essential to the kernel in the integral definition of element matrices k( e) and fq( e) for all

finite elements.
The subject structure in this case is a uniform solid bar. It has length L and is
subjected to a distributed axial load q ( x ) as shown below. It is linear elastic with axial

rigidity EA, and is restrained at the left end and unrestrained at the right end. For
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-2

purposes of discussion it may be assumed that the goal is to solve for the axial
displacement of the bar u ( x ) . However, the focus is on how u ( x ) is to be approximated

within a subregion of the bar.

Bar structure subjected to distributed axial load

The first shape functions to be addressed are low order, linear shape functions in
the context of one-dimensional finite elements for the bar, and its noted they may be
somewhat familiar.1 Higher order, quadratic and cubic shape functions follow in the
same context.
Section 5-1. Linear Shape Functions N i ( x ) , N j ( x ) . A finite element model of

the bar is shown below. Four equally spaced two-node elements of length l are arbitrarily
chosen to discretize the bar and its distributed load. Concentrated nodal loads Qn shown
acting on the model stem from an energy-consistent redistribution to the nodes of that
portion of the distributed load acting over an element. No load is applied to the left end
node for that node is fixed (e.g., see Section 1-6 or Example 4-4).

Finite element model of bar structure using four linear elements

The issue is the simplifying approximation to be made on the axial displacement

for a typical element subregion of the bar model. Consider approximating the axial
⌢
displacement u ≈ u ( x ) using a linear polynomial on 0 ≤ x ≤ l . Since two interpolation

1
Linear shape functions were invoked, without fanfare, for the two-node spring element (Section 1-2.2.4)
and the two-node linear Timoshenko beam element (Section 4-9).
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-3

points are required to define a linear polynomial, the elements have precisely two node
points labeled i and j as illustrated for element (4). These are linear elements and they
are the most basic finite elements. They are also referred to as low order elements.
Nodes common to adjacent elements and nodes at its two ends are structure nodes
and their displacements, labeled Dn, are the model DOF. As a result, this finite element
model has four active DOF, D1 through D4 , but only after having acknowledged the

()
essential boundary condition u 0 = 0 by denoting the left most DOF as D0 and setting it

equal to zero. It follows that the structure stiffness matrix K is an order-4 (symmetric)
matrix, and the order-4 structure displacement and load matrices are, respectively,
T T
D = ⎢ D1 D2 D3 D4 ⎥ and Q = ⎢ Q1 Q2 Q3 Q4 ⎥
⎣ ⎦ ⎣ ⎦
A sketch of exact and computed solutions is shown below, both are notional but
also representative for illustrative purposes.2 The exact solution u ( x ) is represented by
⌢
the solid curve and the finite element or approximate solution u ( x ) by the dashed curve.

The latter curve connects the plotted markers having computed solution values D1
through D4. Due to the two-node linear elements employed, the approximate solution
varies piecewise linearly along the model. The more elements used the more accurate the
piecewise linear approximation; such is the nature of local approximation methods.3

Notional solutions - exact solution and approximate finite element solution for the bar structure

2
Actual solutions are given in graphs for 1-D bar problems in Section 6-4.2.5.
3
Advanced, mesh-free, finite element methods are also examples of local approximation methods since,
even though they characteristically employ no elements, they do employ nodes throughout the solution
domain. For the same reason, the classical finite difference method may be classified as a local
approximation method. Local approximation methods are dependent on the digital computer in a way that
the classical global approximation methods of Rayleigh-Ritz and Galerkin are not.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-4

⌢
5-1.1 Linear behavior function u ( x ) . A typical two-node linear element in the

finite element model of the bar is shown below. The displacement in any such element is
assumed given by the linear polynomial function,
⌢
u ≈ u ( x ) = α1 + α 2 x (0 ≤ x ≤ l )
where x is the axial coordinate system local to the element with origin at node i, and α 1
⌢
and α 2 are unknown constants. Defined in this way, u ( x ) is called an assumed element

behavior function since this polynomial is assumed to govern displacement behavior for
the element, irrespective of the exact or actual displacement u ( x ) for the element

subregion of the bar.

Two-node or linear one-dimensional element

Behavior functions are also viewed as linear combinations of modes of

⌢
deformation. In this case, the behavior function u ( x ) restricts the displacement to a

linear combination of two modes of deformation:

(1) constant or uniform displacement,
(2) linear variation in displacement.
The constants α 1 and α 2 are amplitudes of modes (1) and (2), respectively. Observe that
units of the amplitudes differ and are not in general those of displacement.4 The concept
of deformation modes is employed in Section 5-4.
Since the load acting on the bar q ( x ) is a distributed load, the actual

displacement may well be expected to vary nonlinearly as indicated in the previous

 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-5

⌢
notional graph of u ( x ) , so the linear element behavior function u ( x ) represents an

approximation. If the actual load were instead a set of arbitrary concentrated axial forces
acting along the bar the exact displacement u ( x ) would vary either as a constant or
⌢
linearly between nodes. The assumed behavior function u ( x ) = α 1 + α 2 x would then be
⌢
expected to produce an exact solution, i.e., u ( x ) = u ( x ) , since it can replicate both these

modes of deformation.
A similar situation exists for two-node truss members. They are two-force
members subjected only to concentrated axial forces at their ends (Section 3-3). The
member’s axial strain ε is constant and its axial displacement u ( x ) is correspondingly
linear. For this reason, two-node truss members aligned in the x-direction are the same as
uniform, two-node linear elements presented here.5
⌢ ⌢
5-1.2 Linear interpolation function u ( x ) . The linear behavior function u ( x ) may
also be expressed as a linear combination of two, non-dimensional, linear shape functions
N i (x) and N j (x) with the element dof ui and u j serving as amplitudes, i.e., with nodal

displacements serving as amplitudes. That is,

⌢
u ( x ) = N i ( x ) ui + N j ( x ) u j
⌢
where u ( x ) is called a linear interpolation function. It linearly interpolates element
⌢
displacement u ( x ) between the two nodal displacements ui and u j . Nodal

displacements are clearly very much more meaningful to engineers than amplitudes such
as α1 and α2 appearing in behavior functions.
To derive the linear shape functions, the amplitudes α 1 and α 2 in the linear
⌢
behavior function u ( x ) = α 1 + α 2 x are eliminated in favor of amplitudes ui and u j as

follows. Two end conditions are obtained from the definition of the linear interpolation
function,
⌢ ⌢
u ( 0 ) = ui and u ( l ) = u j

4
This was also the case for the amplitudes in the classical Rayleigh-Ritz and Galerkin approximations
(Sections 2-5 and 2-6).
5
The Abaqus element library does not include a two-node continuum finite element. In its place, a two-
node truss element may be used. See Tutorial T6, Example T6-1.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-6

From these two conditions, two simultaneous equations for constants α 1 and α 2 are
obtained from the behavior function,

α 1 + α 2 × 0 = ui ⎫⎪
⎬ or Cα = d
α1 + α 2 × l = u j ⎪
⎭
where,

⎛ 1 0 ⎞ ⎧⎪ α 1 ⎫⎪ ⎧⎪ ui ⎫⎪
C=⎜ , α = ⎨ ⎬ and d = ⎨ ⎬.
⎝ 1 l ⎟⎠ ⎪⎩ α 2 ⎪⎭ ⎪⎩ u2 ⎪⎭

The solution α = C −1d is,

α 1 = ui
(
α 2 = u j − ui l )
⌢
Upon eliminating α 1 and α 2 in the behavior function, the axial displacement u ( x ) is

expressed in terms of nodal displacements,

⌢ u j − ui
u ( x ) = ui + x
l
Upon factoring the nodal displacements,
⌢ ⎛ x⎞ x
u ( x ) = ⎜ 1 − ⎟ ui + u j (0 ≤ x ≤ l )
⎝ l⎠ l
This is the linear interpolation function for the two-node bar element.
In the linear interpolation function the polynomial coefficients of dof ui and u j

are by definition the shape functions N i and N j , respectively,

x x
Ni ( x ) = 1 − and N j ( x) = (0 ≤ x ≤ l )
l l
⌢
Both shape functions are linear polynomials as required since u ( x ) is linear.
⌢
The interpolation function u ( x ) is often written as the product of a row matrix of

the shape functions N and a column matrix of the nodal displacements d,

⎪⎧ ui ⎪⎫
⌢
() () ()
u x = N i x ui + N j x u j = ⎢⎣ N i x () ()
N j x ⎥⎦ ⎨ ⎬ = Nd
⎪⎩u j ⎪⎭
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-7

Shape functions for the linear two-node, finite element are compiled in Table 5-1.
Graphs of N i ( x ) and N j ( x ) sketched over the two-node element are shown below.

They can be used to illustrate two key properties of all finite element shape functions as
described next.

Graphs of shape functions N i ( x ) , N j ( x )

Table 5-1. Linear shape functions in cartesian coordinate system x.

============================================
Ni ( x ) = 1 − x l
N j ( x) = x l
============================================
Observe that the shape function N i equals 1 at node i corresponding to dof ui and

equals 0 at node j corresponding to dof u j . Likewise, the shape function N j equals 1 at

node j and equals 0 at node i corresponding to dof ui . Summarizing,

⎧⎪ 1 at node i ⎧⎪ 0 at node i
Ni ( x ) = ⎨ and N j ( x) = ⎨
0 at node j 1 at node j
⎩⎪ ⎩⎪
These are examples in FEM of the Kronecker delta property of finite element shape
functions. The name derives from indicial (or index) notation as follows.
Assume the two element nodes are numbered 1 and 2 rather than labeled i and j,
respectively. Instead, let i and j be indices that run from 1 to 2. The Kronecker delta
property for the two shape functions may now be summarized (more succinctly) using the
following indicial notation,
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-8

⎧⎪ 1 i = j
( )
N i x j = δ ij where δ ij = ⎨
0 i≠ j
(i, j = 1,2)
⎩⎪
In mathematics the symbol δ ij is known as the Kronecker delta. 6,7

In addition, finite element shape functions also have the partition-of-unity

property, and for the one-dimensional, two-node linear element it is,
Ni ( x ) + N j ( x ) = 1 (0 ≤ x ≤ l )
This property holds that the sum of an element's shape functions is unity.8 It and the
Kronecker delta property apply to the shape functions of all one-, two- and three-
dimensional continuum finite elements.
Interpolation of displacements within bar elements using linear shape functions
N i ( x ) and N j ( x ) is illustrated in the following example.

===============================================================
Example 5-1. For the bar finite element model shown, let l = 0.1 m and L = 0.4 m and
assume that values for the structural nodal loads (i.e., Q1, Q2, Q3 and Q4) are such that the
finite element (notional) solution for nodal displacements D = K −1Q is,
T T
D = ⎢⎣ D1 D2 D3 D4 ⎥⎦ = ⎢⎣1.0 4.0 9.0 16.0 ⎥⎦ × 10 −3 m

6
Named after Leopold Kronecker (1823-1897) who famously declared “God made the integers,” “all the
rest is the work of man.” The declaration was particularly directed at his contemporary K. W. T.
Weierstrass (1815-1897), the father of mathematical analysis. (Men of Mathematics by K. T. Bell, p. 479,
Simon and Schuster, New York, 1937.)
7
For one-dimensional elements herein, nodes are labeled with Latin characters i, j, k, and l exclusively so
use of indicial notation has neither been an option nor contemplated.
8
The partition-of-unity property does not apply to the four cubic shape functions of the two-node beam
member (Section 4-2), for which it can be shown N 1 + N 3 = 1 and N 2 + N 4 = x .
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-9

⌢
()
(a) Interpolate the displacement u x at position X = 0.0667 m :

Since l = 0.1m , position X is in element (1) where x = X = 0.0667 m . For

element (1), ui = D0 (at the left end restraint) and u j = D1 . Displacement in element (1)

is given by its interpolation function,

⎛ x⎞ x
⌢
()
u x = N iui + N j u j = ⎜ 1 −
⎝ ⎟
l⎠
D0 + D1
l

Therefore,
⎛ 0.0667 ⎞ ⎛ 0.0667 ⎞
⌢
u X =x=0.0667
= ⎜1−
⎝ 0.1 ⎠⎟ ( 0) + ⎜
⎝ 0.1 ⎠ ⎟ ( )
1.0 × 10 −3 = 0.667 × 10 −3 m

⌢
()
(b) Interpolate the displacement u x at position X = 0.3333 m :

Position X is in element (4) where x = X − X3 = 0.3333 − 0.3 = 0.0333 m . For

element (4), ui = D3 and u j = D4 . Displacement in element (4) is given by its

interpolation function,
⎛ x⎞ x
⌢
()
u x = N iui + N j u j = ⎜ 1 −
⎝ ⎟
l⎠
D3 + D4
l
Therefore,
⎛ 0.0333 ⎞
⌢
u X =0.3333
⌢
=u x=0.0333
= ⎜1−
⎝ 0.1 ⎠⎟ (
9.0 × 10 −3 +)0.0333
0.1
( )
16.0 × 10 −3 = 11.33 × 10 −3 m

===============================================================

Section 5-2. Quadratic Shape Functions N i ( x ) , N j ( x ) , N k ( x ) . Consider

again a four-element model of the bar structure. This time, however, quadratic elements
are specified as shown in the sketch below of the new model. Quadratic finite elements
are referred to as high order elements. Increased accuracy can be anticipated from their
use at the expense of increased computation. Since three data points are required to
define a quadratic polynomial, the elements have three nodes labeled i, j and k. Node j is
an internal node as shown and it is positioned at the center of an element.
Though there are still only four elements in the bar model, it now has eight active
DOF compared to four in the previous model employing linear elements. Consequently,
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-10

the structure system of equations KD = Q represents eight simultaneous, nodal

equilibrium equations to be solved rather than four.

Finite element model of bar structure using four quadratic elements

⌢
5-2.1 Quadratic behavior function u ( x ) . The assumed approximation for axial

displacement for a typical element of the bar model is a quadratic function, as indicated
in the sketch below. The behavior function is a linear combination of three displacement
modes as follows,
⌢
u ≈ u ( x ) = α1 + α 2 x + α 3 x2 (0 ≤ x ≤ l )
Again, α1, α2 and α3 are unknown constants or displacement mode amplitudes. Clearly
as before, units of these constants differ and are not in general those of displacement.

Three-node or quadratic one-dimensional element

⌢
5-2.2 Quadratic interpolation function u ( x ) . The assumed quadratic behavior
⌢
function u ( x ) is now to be expressed in terms of nodal displacement amplitudes ui, uj

and uk and corresponding shape functions N i ( x ) , N j ( x ) and N k ( x ) . An approach to

deriving finite element shape functions using the behavior function was introduced for
the linear element. It may be formally called the matrix inversion method. Using this
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-11

method, the derivation of N i ( x ) , N j ( x ) and N k ( x ) involves solving for the amplitudes

α 1 , α 2 and α 3 in terms of nodal displacements ui , u j and uk using the inverse

coefficient matrix C−1 .

In this case, the conditions invoked at the nodes are,
⌢ ⌢ ⌢
u ( 0 ) = ui u (l 2) = u j u (l ) = uk

Using the behavior function, these conditions provide three simultaneous equations in the
three unknown amplitudes as follows.
α 1 + α 2 ( 0 ) + α 3 ( 0 ) = ui
2

α 1 + α 2 (l 2) + α 3 (l 2) = u j
2

α 1 + α 2 (l ) + α 3 (l ) = uk
2

or, in matrix form,

Cα = d
where,
⎛ 1 0 0 ⎞ ⎧ α1 ⎫ ⎧ ui ⎫
⎜ ⎟ ⎪ ⎪ ⎪⎪ ⎪⎪
C=⎜ 1 l 2 ( l 2 )2 ⎟ , α = ⎨ α2 ⎬ and d = ⎨ u j ⎬
⎜ 1 l ⎟ ⎪ α ⎪ ⎪ u ⎪
⎝ l2 ⎠ ⎩ 3 ⎭ ⎪⎩ k ⎪⎭

This order-3 system may be inverted, i.e., α = C −1d , or in this case, it can be solved more
expediently as follows. From the first equation it’s clear that α 1 equals u1 . Upon
eliminating α 1 the remaining two equations in matrix form become,

⎛ l 2 ⎞⎧ α ⎫⎪ ⎧⎪ u j − ui ⎫⎪
( l 2 )2 ⎪ 2
(α 1 = ui )
⎜ ⎟⎨ ⎬= ⎨ ⎬
⎜⎝ l l2 ⎟⎠ ⎪ α 3 ⎪⎭ ⎪⎩ uk − ui ⎪⎭
⎩
For any invertible order-2 matrix the inverse is given by,
−1
⎛ a b⎞ 1 ⎛ d −b ⎞
⎜⎝ c d ⎟⎠ =
ad − bc ⎜⎝ −c a ⎟⎠

Thus, the solution for α2 and α3 is

⎧α 2 ⎫ 1 ⎛ l 2 − ( l 2 )2 ⎞ ⎧ u j − ui ⎫
⎨ ⎬= 3 ⎜ ⎟⎨ ⎬
⎩α 3 ⎭ l 2 − l 4 ⎝ −l l 2 ⎠ ⎩u k − u i ⎭
3

or,
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-12

α2 =
4⎡ 2
3 ⎢
l ⎣
( ) ( )(
l u j − ui − l 2 uk − ui ⎤
2

⎥⎦ )
4⎡ 1 ⎤
(
= ⎢u j − ui − uk − ui ⎥
l⎣ 4
)
⎦
and,

α3 =
⎡ 4
( l
) ( ⎤
⎢ −l u j − ui + 2 uk − ui ⎥
⎣ l3 ⎦
)
4⎡ 1 ⎤
= 2 ⎢ −u j + ui + uk − ui ⎥
l ⎣ 2
( )
⎦
⌢
Substituting for α1, α2 and α3 in the behavior function u ( x ) = α 1 + α 2 x + α 3 x 2 yields,

⌢ 4⎡ 1 ⎤ 4 ⎡ 1 ⎤ 2
u ( x ) = ui + ⎢ u j − ui − ( u k − ui ) ⎥ x + 2 ⎢⎣ −u j + ui + 2 ( uk − ui ) ⎥⎦ x
l⎣ 4 ⎦ l
The quadratic shape functions N i ( x ) , N j ( x ) and N k ( x ) are defined as the

coefficients of ui , u j and uk , respectively, in the following interpolation function,

⌢
u ( x ) = N i ui + N j u j + N k u k

For example, determining the coefficient of ui will yield shape function N i as follows,

⎛ 4 1 4 2 ⎞
N i ui = ⎜ 1 − x + x + 2 x 2 − 2 x 2 ⎟ ui
⎝ l l l l ⎠

Upon simplifying the coefficient, the quadratic shape function N i is written,

3 2
Ni ( x ) = 1 − x + 2 x2 (0 ≤ x ≤ l )
l l
The Kronecker delta property for shape function N i is,

⎧⎪ 1 at node i
Ni ( x ) = ⎨
0 at nodes j , k
⎩⎪
That it is satisfied may be easily verified by substituting for x its value at the three nodes
as follows:
Node i : N i ( 0 ) = 1
3 l 2 l2
Node j : N i ( l 2 ) = 1 − + =0
l 2 l2 4
3 2
Node k : N i ( l ) = 1 − l + 2 l 2 = 0
l l
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-13

Thus, the derived expression for N i ( x ) is first a quadratic function and second satisfies

the Kronecker delta property, so it is the shape function corresponding to nodal

displacement ui .
Higher order polynomials (cubic,
quartic, etc.) will also pass through the same

( )( )( )
three data points 0,1 , l 2,0 , l,0 for

nodes i, j, and k. As indicated here, the

quadratic polynomial that does so is unique
among all quadratic polynomials. In general,
one and only one nth-order polynomial Uniqueness of quadratic shape function N i ( x )
Pn ( x ) goes right through a given set of n + 1

data points. In this case n = 2 and therefore the derived quadratic shape function N i ( x )

is unique. This, of course, is also true when n = 1 in the case of linear shape functions
N i ( x ) and N j ( x ) associated with the two-node linear element.

The Lagrange interpolation method presented in the following is an alternative to

the matrix inversion method for deriving shape functions. It avoids having to formally
solve simultaneous equations and eliminates much of the associated algebra. Further, as
presented for shape functions, it is an informal, expedient version of Lagrange
interpolation. 9
Table 5-2. Lagrange polynomial interpolation data for deriving N i ( x ) .

Node x y = Ni x()
i 0 1
j l/2 0
k l 0

9
Modern hand-held calculators and computer programs such as Mathematica have impressive capabilities
for symbolically solving linear systems of equations like Cα = d . Nonetheless, an intuitive approach to
deriving shape functions that avoids matrix inversion and algebraic tedium is preferable. With few
exceptions, the presented Lagrange interpolation method is employed hereinafter for deriving shape
functions.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-14

The formal Lagrange interpolation method usually begins with a tabular

collection of points ( xm , f ( xm ) ) through which an interpolating polynomial y = f ( x ) is

to be constructed.10 The data needed to construct the quadratic polynomial shape

function y = N i ( x ) are given in Table 5-2.

The Lagrange formula for the 2nd degree polynomial N i ( x ) is,

( )
N i ( x ) = N ( xi ) li ( x ) + N x j l j ( x ) + N ( xk ) lk ( x )

where the 2nd degree Lagrange polynomials indicated in the formula are,

li ( x ) =
( x − x j ) ( x − xk ) l j ( x) =
( x − xi )( x − xk ) lk ( x ) =
( x − xi )( x − x j )
( xi − x j )( xi − xk ) ( x j − xi )( x j − xk ) ( xk − xi )( xk − x j )
However, the Lagrange formula simplifies to one term due to the Kronecker delta
property of shape functions represented in Table 5-2 requiring N ( xi ) = 1 and

( )
N x j = N ( xk ) = 0 . Therefore, the shape function sought is,

Ni ( x ) = 1 ×
( x − x j ) ( x − xk ) = ( x − l 2 ) ( x − l ) = 2 ⎛ x − l ⎞ ( x − l )
( xi − x j )( xi − xk ) (0 − l 2)(0 − l ) l 2 ⎜⎝ 2 ⎟⎠
Interpolation of the three required
points by this quadratic function is
shown here by its graph. It will always
be the case for derivation of shape
functions by the Lagrange method that
only one term in the Lagrange formula Graph of N i x ()
survives.
The following intuitive version of Lagrange interpolation is even more expedient.
The shape function can be written at least partially by inspection to pass through the
points where the shape function values are zero. In the present case, the points
correspond to nodes j and k. Therefore, it follows that, irrespective of a constant c,
⎛ l⎞
Ni ( x ) = c ⎜ x − ⎟ ( x − l )
⎝ 2⎠

10
Gerald, C. F. and P. O. Wheatley (1999). Applied Numerical Analysis, 6th Ed., Art. 3.2, Addison Wesley
Longman, Inc., New York, 698 Pages, 1999.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-15

That is, irrespective of the constant c, the first linear polynomial is clearly zero at
node j, and the second linear polynomial is clearly zero at node k, so it follows that their
product is zero at these two node points as required by the Kronecker delta property. The
formula is also quadratic, as required, since it is the product of two linear polynomials.
To complete the derivation, the constant c must be such that N i ( x ) equals 1 at
node i, also required by the Kronecker delta property. Thus,
⎛ l⎞ 2
Ni (0 ) = 1 ⇒ c ⎜ 0 − ⎟ (0 − l ) = 1 ⇒ c = 2
⎝ 2⎠ l
The result for N i ( x ) using this more expedient version of Lagrange interpolation is, as

before with the formal version,

2⎛ l⎞
Ni ( x ) = 2 ⎜
x − ⎟ (x − l) (0 ≤ x ≤ l )
l ⎝ 2⎠
That this quadratic shape function is the same as that derived previously by the
matrix inversion method is verified by its expansion,
2⎛ l⎞ 2 ⎡ 2 3l l2 ⎤ 3 2
Ni ( x ) = 2 ⎜
x − ⎟⎠ ( x − l ) 2⎢
= x − x + ⎥ = 1 − x + 2 x2
l ⎝ 2 l ⎣ 2 2⎦ l l

This of course must be the case given the uniqueness of shape functions.
In the following, the Lagrange interpolation method is used to derive the
remaining two quadratic shape functions N j ( x ) and N k ( x ) for the three-node quadratic

element. Given the definition of N j ( x ) , its graph must pass through the three points

(0,0), (l/2,1) and (l,0), and therefore may

be sketched as shown here. It is clear from
the graph that the two zero values at nodes
i and k are interpolated by the quadratic
Lagrange formula,
N j ( x ) = c ( x − 0 )( x − l )
Graph of N j ( x )
Finally, since N j ( x ) = 1 at node j where

x = l 2,

⎛l ⎞⎛ l ⎞ 4
c ⎜ − 0⎟ ⎜ − l⎟ = 1 ⇒ c = − 2
⎝2 ⎠⎝2 ⎠ l
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-16

Therefore, the formula for the shape function using Lagrange interpolation is,
4
N j ( x) = − x(x − l) (0 ≤ x ≤ l )
l2
Likewise, the graph of N k ( x ) may be sketched from its definition as shown here.

Inspection shows that the

two zero values at nodes i
and j shown in the graph
are interpolated by the
quadratic Lagrange
formula,
⎛ l⎞
Nk ( x) = c( x − 0)⎜ x − ⎟
⎝ 2⎠
Also, since N k ( x ) = 1 at node k where x = l ,

⎛ l⎞ 2
c (l − 0 ) ⎜ l − ⎟ = 1 ⇒ c = 2
⎝ 2⎠ l
Therefore, the shape function is,
2 ⎛ l⎞
Nk ( x) = x⎜ x − ⎟ (0 ≤ x ≤ l )
l2 ⎝ 2⎠
Table 5-3 catalogs the shape functions for the three node quadratic finite element
having length l. It may be verified that these shape functions also satisfy the partition-of-
unity property,
Ni ( x ) + N j ( x ) + N k ( x ) = 1 (0 ≤ x ≤ l )
Table 5-3. Quadratic shape functions in cartesian coordinate sytem x.
==============================================

2⎛ l⎞ ⎫
Ni ( x ) = 2 ⎜
x − ⎟ ( x − l )⎪
l ⎝ 2⎠
⎪
4 ⎪
N j ( x) = − 2 x ( x − l) ⎬ (0 ≤ x ≤ l )
l ⎪
2 ⎛ l⎞ ⎪
Nk ( x) = 2 x ⎜ x − ⎟ ⎪
l ⎝ 2⎠ ⎭

==============================================
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-17

⌢
Finally, the displacement interpolation function u ( x ) for the one-dimensional

quadratic finite element may be expressed symbolically as follows,

⎧ ui ⎫
⎪ ⎪
⌢
() () () ()
u x = N i x ui + N j x u j + N k x uk = ⎢⎣ N i x () Nj x() ()
N k x ⎥⎦ ⎨u j ⎬ = Nd
⎪ ⎪
⎩uk ⎭
Section 5-3. Cubic Shape Functions N i ( x ) , N j ( x ) , N k ( x ) , N l ( x ) . Four data

points are required to define a cubic polynomial. Therefore, a cubic finite element is
required to have four nodes, i,j, k and l , as shown for element (3) in the bar model
below. Using cubic elements, the four-element bar model will have 12 active DOF. The
structure equations KD = Q for this model represent a system of 12 simultaneous, nodal
equilibrium equations in the unknown structure displacements D1 through D12.

Finite element model of bar structure using four cubic elements

The axial displacement for an element is an assumed linear combination of four

displacement modes according to the following assumed cubic behavior function,

u ≈ u ( x ) = α1 + α 2x + α 3x2 + α 4 x3
where α 1 through α 4
are unknown mode
amplitudes. A
notional graph of the
function is shown
here plotted over a
typical element in the
bar model. When this
Four-node or cubic one-dimensional element
function is re-
expressed as a linear combination of products of shape functions and nodal displacements
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-18

it becomes the following finite element interpolation function for the cubic element,
⌢
u ( x ) = N i ( x ) ui + N j ( x ) u j + N k ( x ) u k + N l ( x ) ul

In symbolic form the interpolation function is,

⎧ ui ⎫
⎪ ⎪
⎪u j ⎪
⌢
() () ()
u x = ⎢⎣ N i x N j x N k x N l x ⎥⎦ ⎨ ⎬ = Nd () ()
⎪uk ⎪
⎪u ⎪
⎩ l⎭
The Lagrange interpolation method may be used to derive the four shape
functions. The results are listed in Table 5-4 and their corresponding graphs are sketched
in the adjacent figure. Verification of these formulas is left as an exercise (Problem 5-1).
The Kronecker delta properties are:
⎧⎪ 1 at node i
Ni ( x ) = ⎨
⎪⎩ 0 at nodes j, k,l
⎧⎪ 1 at node j
N j ( x) = ⎨
⎪⎩ 0 at nodes i, k,l
⎧⎪ 1 at node k
Nk ( x) = ⎨
0 at nodes i, j,l
⎩⎪
⎧⎪ 1 at node l
Nl ( x ) = ⎨
0 at nodes i, j, k
⎩⎪
The partition-of-unity property is,
Ni ( x ) + N j ( x ) + N k ( x ) + Nl ( x ) = 1 (0 ≤ x ≤ l ) .

Table 5-4. Cubic shape functions in cartesian

coordinate system x
=============================
9 ⎛ l⎞⎛ 2l ⎞
Ni ( x ) = − 3 ⎜ x − ⎟ ⎜ x − ⎟ ( x − l )
2l ⎝ 3⎠ ⎝ 3⎠
27 ⎛ 2l ⎞
N j ( x) = x x − ⎟ (x − l)
3 ⎜
2l ⎝ 3⎠
27 ⎛ l⎞
Nk ( x) = − 3 x ⎜ x − ⎟ ( x − l )
2l ⎝ 3⎠ Cubic shape function sketches
9 ⎛ l⎞⎛ 2l ⎞
Nl ( x ) = 3 x ⎜ x − ⎟ ⎜ x − ⎟
2l ⎝ 3⎠ ⎝ 3⎠
=============================
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-19

Section 5-4. Rigid Body and Deformational Modes of Displacement. The

partition-of-unity property of shape functions serves an important purpose. It allows
finite element interpolation functions to replicate rigid body displacement modes.
It is possible that pure rigid body displacement of an element may occur as a
result of certain loading conditions, and it is important that a finite element formulation
be capable of replicating this type of displacement as a matter of element convergence
(Section 5-6). An element undergoes a pure rigid body displacement when all its nodal
displacements have the same value, denoted here by the constant u0 . The element
displaces uniformly a distance u0 in the axial direction.
To see how the partition-of-unity property plays into this, consider that in the
previous four-element model of the bar using cubic elements the loads are for example
Qi > 0 ( i = 1,2,…,9 ) and Q10 = Q11 = Q12 = 0 . Under this load condition element (4)

experiences a rigid body displacement u0 . Now, displacement in element (4) given by its
⌢
interpolation function u will equal the constant u0 as follows,
⌢
u ( x ) = N i ( x ) ui + N j ( x ) u j + N k ( x ) u k + N l ( x ) ul
= ⎡⎣ N i ( x ) + N j ( x ) + N k ( x ) + N l ( x ) ⎤⎦ u0
= u0
In this way, the partition-of-unity property for the element,
Ni ( x ) + N j ( x ) + N k ( x ) + Nl ( x ) = 1 (0 ≤ x ≤ l )
provides for replicating a rigid body displacement of element (4).
Under these same pure rigid body displacement conditions, the cubic element's
assumed behavior function for element (4) is of course also equal the constant u0 as
follows,
⌢
u ( x ) = α 1 + α 2 x + α 3 x 2 + α 4 x 3 = u0 (0 ≤ x ≤ l )
Since the rigid body displacement u0 is independent of position x, the displacement
mode amplitudes α i take the following values,
α 1 = u0 and α 2 = α 3 = α 4 = 0
This illustrates the importance of amplitude α1 in the assumed polynomial for a behavior
function. It’s role is to replicate the rigid body component in an element's displacement
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-20

⌢
behavior. So, α1 plays this role in the behavior function form of u ( x ) and the partition-
⌢
of-unity property plays this role in the interpolation function form of u ( x ) .

This capability is required not only for the case of pure rigid body displacement
discussed above, but also for the general deformation case when an element is subjected
to a combination of rigid body and deformational modes of displacement. When is
displacement a combination of pure rigid body and pure deformation modes? Recall the
earlier, basic analysis of axial deformation (Section1-1.2). The following analysis is
similar.
Consider that the restrained bar shown below is subjected to axial stretching under
some concentrated load at the tip. Let u ≡ u A be the total displacement of some arbitrary
point A on the bar. Then, the total displacement of any point B a distance Δ x to the right
of point A may be expressed as follows,

⎛ rigid shift ⎞ ⎛ stretch of ⎞

⎜ ⎟
u B = ⎜ to the right ⎟ + ⎜ segement Δ x ⎟ = u A + Δu
⎜ ⎟
(
⎝ u ≡ u A ⎠ ⎜⎝ Δu ≡ Δ x * − Δ x )
⎟
⎠

Therefore, the displacement u ( x ) of any point x is in general always a combination of

rigid body and deformational modes.

Displacement of a point is comprised of rigid body and deformational modes

Deformational modes of displacement in, for example, the cubic element are
provided for by the terms α 2 x (linear deformation mode), α 3 x 2 (quadratic deformation
⌢
mode) and α 4 x 3 (cubic deformation mode). The displacement u computed for a one-
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-21

dimensional cubic element is a linear combination of the pure rigid body mode α 1 and

these three pure deformational modes.

===============================================================
Example 5-2. Consider the one-dimensional quadratic element shown. Calculate the
T
displacement mode amplitudes α = ⎢⎣α 1 α 2 α 3 ⎥⎦ and write its behavior function
⌢
u ( x ) = α 1 + α 2 x + α 3 x 2 , for each of the three sample nodal displacement matrices
T
d = ⎢⎣ui uj uk ⎥⎦ given below. Note the amplitudes are given by the equation

α = C −1d (see Section 5-2.2). In this case since l = 2, the equation is,
−1
⎧α i ⎫ ⎛ 1 0 0 ⎞ −1
⎧ ui ⎫ ⎛ 1 0 0 ⎞ ⎧ ui ⎫ ⎛ 2 0 0 ⎞ ⎧ ui ⎫
⎪ ⎪ ⎜ ⎪ ⎪ ⎜ ⎪ ⎪ ⎪ ⎪
(l 2)2 ⎟⎟
1⎜
⎨α j ⎬ = ⎜ 1 l 2 ⎨u j ⎬ = ⎜ 1 1 1 ⎟⎟ ⎨u j ⎬ = −3 4 −1⎟ ⎨u j ⎬
⎪α ⎪ ⎜ 1 l ⎪u ⎪ ⎝ 1 2 4 ⎠ ⎪u ⎪ 2⎜ ⎟
⎩ k⎭ ⎝ l 2 ⎟⎠ ⎩ k⎭ ⎩ k⎭ ⎝ 1 −2 1 ⎠ ⎪⎩uk ⎪⎭

T
(1) Pure rigid body mode displacement, d = ⎢⎣0.1 0.1 0.1⎥⎦ :
⌢
∴ α = ⎢⎣0.1 0 0 ⎥⎦ ⇒ u ( x ) = 0.1
T

T
(2) Pure deformational modes, d = ⎢⎣0.0 0.1 0.1⎥⎦ :
⌢
∴ α = ⎢⎣0 0.15 −0.05 ⎥⎦ ⇒ u ( x ) = 0.15x − 0.05x 2
T

T
(3) Combination of rigid body and deformational modes, d = ⎢⎣0.1 0.2 0.3 ⎥⎦ :
⌢
⇒ u ( x ) = 0.1 + 0.1x
T
∴ α = ⎢⎣0.1 0.1 0 ⎥⎦
===============================================================

Section 5-5. Natural Coordinate Systems for One-Dimensional Elements.

Natural coordinates are dimensionless coordinates that are employed for local, element
coordinate systems. As will be seen in later chapters, they are crucially important to
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-22

formulation of two- and three-dimensional finite elements. This is because such elements
generally have irregular shapes for which natural coordinate systems are more suited than

cartesian coordinate systems. Furthermore, a stiffness matrix k ( ) for a continuum finite

e

element will be defined in the sequel as a definite integral over the volume of the
element. These integrals are generally evaluated using numerical integration. In turn,
numerical integration is greatly facilitated by use of natural coordinates whose limits
range from -1 to 1 or from 0 to 1, irrespective of element size, orientation, and shape.
Alternative one-dimensional coordinate systems are listed in Table 5-5 for a
linear, two-node element as an example. Their two- and three-dimensional counterparts
are encountered in later chapters. The global and local cartesian coordinates X and x,
respectively, have already been employed for the one-dimensional element. Uppercase,
catesian coordinates are used for global or structure coordinate systems exclusively. All
other coordinate systems listed are local or element coordinate systems. The natural
coordinate ξ, and length coordinates Li , L j and related natural coordinate r, have
limited usage in one-dimension. They are introductory to two- and three-dimensional
elements involving quadrilateral, hexahedral, triangular, and tetrahedral element shapes.

Table 5-5. Alternative one-dimensional coordinate systems for the two-node element.
Coordinate System Node i Node j Range
Coordinate Coordinate
Global cartesian
coordinate X X = Xi X = Xj (0 ≤ X ≤ L )
Local cartesian
coordinate x x=0 x=l (0 ≤ x ≤ l )
Local natural
coordinate ξ ξ = −1 ξ=1 ( −1 ≤ ξ ≤ 1)
*Local length Li = 1 Li = 0 (1 ≤ Li ≤ 0 )
coordinates Li, Lj
------------------------
Lj = 0 Lj = 1
(0 ≤ L j )
≤1
Local natural -------------- ----------------
----------------
r=0 r=1
coordinate r (0 ≤ r ≤ 1)
* Any point or node has two length coordinates, (Li, Lj), only one of which is independent.
If it is to be Lj, then let r ≡ Lj be an alternative natural coordinate for the 1-D coordinate system.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-23

As usual the origin of the cartesian coordinate system X is positioned at the left
end of the structural model for the bar with length L, and the origin of the cartesian
coordinate system x for the bar element is at its left end or node i as shown below. The
origin of natural coordinate system ξ is positioned at the center of the element as
indicated. For the present, it is sufficient to introduce the natural coordinate ξ for one-
dimensional finite elements by expressing shape functions in terms of ξ beginning in the
next section. The following coordinate transformations will prove useful when doing so,

X = Xi + x or x = X − Xi
2⎛ l⎞ l
ξ = ⎜x− ⎟ or x= (ξ + 1)
l⎝ 2⎠ 2

Alternative cartesian and natural coordinate systems for 1-D linear element

As shown below, any point P along the element having cartesian coordinate x,
also has length coordinates ( Li , L j ) and natural coordinate r where,

l−x x
Li = and Lj = and r = L j
l l
The length coordinate system has no origin per se. Observe that only one length
coordinate is independent since it is clear that Li + L j = 1 . It follows that point P is

positioned at the centroid of the element when its two length coordinates are equal, i.e.,
when Li = L j = 1 2 . Further, for future reference observe that Li and L j are equal to

linear shape functions N i and N j , respectively, and that Li + L j = N i + N j = 1 . That is,

Li = ( l − x ) l = 1 − x l = N i ( x )
Lj = x l = N j ( x)
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-24

Alternative cartesian and natural length coordinate systems for 1-D linear element

On occasion in FEM certain element volume integrals are encountered expressed

in terms of length coordinates. This allows them to be evaluated closed form using
simple formulas.11 For a one-dimension element of length l the formula is,
l a b a!b!
∫0 Li L j dl = ( a + b!)! × l
where a and b are arbitrary exponents. For example,
l 2 2! 0! 1
∫0 Li dl = ( 2 + 0!)! × l = 3 l
Extended versions of this formula exist for certain two- and three-dimensional elements.
5-5.1 Linear shape functions in natural coordinate ξ. The linear shape functions
N i ( x ) and N j ( x ) may be transformed using the coordinate transformation
x (ξ ) = ( l 2 ) (ξ + 1) as follows,
x l 1 x ( l 2 ) (ξ + 1) 1
Ni = 1 − = 1 − (ξ + 1) = (1 − ξ ) and N j = = = (1 + ξ )
l 2l 2 l l 2
These results are summarized in Table 5-6. Graphs of the two shape functions retain the
same linear profiles as before (Section 5-1.2), but now the origin is at the center of the
element, as shown below. The key properties of shape functions are retained as may be
easily verified from these graphs.
Table 5-6. Linear shape functions in natural coordinate system ξ.
===========================================
1
N i (ξ ) = (1 − ξ )
2
1
N j (ξ ) = (1 + ξ )
2
===========================================

11
M. A. Eisenberg and L. E. Malvern. “On Finite Element Integration in Natural Coordinates,” I.J.N.M.E.,
7, pp. 574-575, 1973.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-25

⎧⎪ 1 at node i ⎧⎪ 0 at node i
N i (ξ ) = ⎨ N j (ξ ) = ⎨
⎪⎩ 0 at node j ⎪⎩ 1 at node j

N i (ξ ) + N j (ξ ) = 1

Graphs of linear shape functions with key properties in natural coordinate system ξ

⌢
Axial displacement u is interpolated in the linear element between the two nodal
values ui and u j using shape functions N i and N j as before. Since these shape

functions are now expressed in the natural coordinate ξ the interpolation function is,

⎧⎪ ui ⎫⎪
⌢
() () ()
u ξ = N i ξ ui + N j ξ u j = ⎢⎣ N i ξ () ()
N j ξ ⎥⎦ ⎨ ⎬ = Nd
⎩⎪u j ⎪⎭
5-5.2 Quadratic shape functions in natural coordinate ξ . Applying the coordinate
transformation x = ( l 2 ) (ξ + 1) to N i ( x ) , N j ( x ) and N k ( x ) yields,
2⎛ l⎞ 2 ⎡l l ⎤⎡l ⎤
Ni = 2 ⎜
x − ⎟ ( x − l ) = 2 ⎢ (ξ + 1) − ⎥ ⎢ (ξ + 1) − l ⎥
l ⎝ 2⎠ l ⎣2 2⎦ ⎣2 ⎦
4 4 ⎡l ⎤⎡l ⎤
N j = − 2 x ( x − l ) = − 2 ⎢ (ξ + 1) ⎥ ⎢ (ξ + 1) − l ⎥
l l ⎣2 ⎦ ⎣2 ⎦
2 ⎛ l⎞ 2 ⎡l ⎤⎡l l⎤
Nk = x ⎜ x − ⎟ = 2 ⎢ (ξ + 1) ⎥ ⎢ (ξ + 1) − ⎥
l 2
⎝ 2 ⎠ l ⎣2 ⎦ ⎣2 2⎦

Table 5-7. Quadratic shape functions in natural coordinate system ξ .

=============================================
1
N i (ξ ) = ξ (ξ − 1)
2
N j (ξ ) = (1 + ξ ) (1 − ξ )
1
N k (ξ ) = ξ (ξ + 1)
2
=============================================
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-26

Upon simplifying these expressions, the three quadratic shape functions listed in Table 5-
7 are obtained. Their graphs are shown below. From these graphs the reader may verify
that the key properties of these shape functions are satisfied.

Graphs of quadratic shape functions in natural coordinate system ξ

Instead of using the coordinate transformation method, the Lagrange interpolation

method may be used to derive the quadratic shape functions in terms of natural
coordinate ξ , perhaps more intuitively. For example, by definition for N i (ξ ) ,

⎧⎪ 1 at node i
N i (ξ ) = ⎨
⎪⎩ 0 at nodes j, k

N i (ξ ) must also be quadratic and its graph must therefore look exactly like that given

above. By inspection, the following quadratic (Lagrange) function passes through zero at
nodes j and k as required, irrespective of the constant c,
N i (ξ ) = c (ξ − 0 ) (ξ − 1)

Also, for the constant c it’s noted that N i ( −1) = 1 so that,

N i ( −1) = c ( −1) ( −2 ) = 1 ⇒ c = 1 2
So, by the Lagrange interpolation method,
1
N i (ξ ) = ξ (ξ − 1)
2
N j (ξ ) and N k (ξ ) as given in Table 5-7 may be derived similarly by this method.

Finally, with the quadratic shape functions in hand, the corresponding quadratic
interpolation function for displacement may be written,
⎧ ui ⎫
⌢ ⎪ ⎪
u (ξ ) = N i (ξ ) ui + N j (ξ ) u j + N k (ξ ) uk = ⎢⎣ N i (ξ ) N j (ξ ) N k (ξ ) ⎥⎦ ⎨u j ⎬ = Nd
⎪u ⎪
⎩ k⎭
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-27

5-5.3 Cubic shape functions in natural coordinate ξ. According to the key

properties of shape functions, the graph of the cubic shape function N i (ξ ) , for example,

is like that shown here.

Cubic shape function graph for node i in natural coordinate system ξ

The cubic Lagrange formula that interpolates zero values at node points j, k, and l is by
inspection,
⎛ 1⎞ ⎛ 1⎞
N i (ξ ) = c ⎜ ξ + ⎟ ⎜ ξ − ⎟ (ξ − 1)
⎝ 3⎠ ⎝ 3⎠
Irrespective of the value for the unknown constant c, the first, second and third linear
polynomials in ξ appearing in this expression guarantee that N i (ξ ) equals zero at nodes

j, k and l, respectively. Further, the product of three linear polynomials is a cubic

⌢
polynomial as required by a cubic interpolation function u(ξ ) . Finally, by applying the
condition that N i = 1 at node i where when ξ = −1 the value of the constant c is obtained,

⎛ 1⎞ ⎛ 1⎞
N i ( −1) = 1 ⇒ c ⎜ −1 + ⎟ ⎜ −1 − ⎟ ( −1 − 1) = 1
⎝ 3⎠ ⎝ 3⎠
⎛ 2⎞ ⎛ 4⎞ ⎛ 6 ⎞
c⎜ − ⎟ ⎜ − ⎟ ⎜ − ⎟ = 1
⎝ 3⎠ ⎝ 3⎠ ⎝ 3 ⎠
27 9
c=− =−
48 16
Therefore, the cubic shape function N i (ξ ) is,

9 ⎛ 1⎞ ⎛ 1⎞
N i (ξ ) = − ⎜⎝ ξ + ⎟⎠ ⎜⎝ ξ − ⎟⎠ (ξ − 1)
16 3 3
Dividing the coefficient by -9 and multiplying the rest by -9 yields an alternative form,
1
N i (ξ ) = (1 − ξ ) ( 3ξ + 1) ( 3ξ − 1)
16
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-28

Similar application of Lagrange interpolation yields the formulas for cubic shape
functions for nodes j, k and l in terms of natural coordinate ξ . All the results are
summarized in Table 5-8. They of course satisfy the key properties of shape functions.
The corresponding cubic interpolation function for displacement is,
⌢
u ( ξ ) = N i ( ξ ) ui + N j ( ξ ) u j + N k ( ξ ) u k + N l ( ξ ) ul
⎧ ui ⎫
⎪u ⎪
⎪ j⎪
= ⎢⎣ N i (ξ ) N j (ξ ) N k (ξ ) N l (ξ ) ⎦ ⎨ ⎬ = Nd
⎥
⎪u k ⎪
⎪⎩ ul ⎪⎭

Table 5-8. Cubic shape functions in natural coordinate system ξ

==========================================
1
N i (ξ ) = (1 − ξ ) ( 3ξ + 1) ( 3ξ − 1)
16
9
N j (ξ ) = (1 + ξ ) (ξ − 1) ( 3ξ − 1)
16
9
N k (ξ ) = (1 + ξ ) (1 − ξ ) ( 3ξ + 1)
16
1
N l (ξ ) = (1 + ξ ) ( 3ξ + 1) ( 3ξ − 1)
16
==========================================

5-5.4 Interpolation of element geometry in natural coordinates. Element

geometry considered in a global cartesian coordinate system can be approximated using
shape functions expressed in natural coordinates. For example, the geometry X for the
two-node element may be interpolated between element nodal coordinates Xi and X j

using the linear shape functions N i (ξ ) and N j (ξ ) in the same manner shown previously
⌢
for interpolation of element displacement u between the element’s nodal displacements
ui and uj (Section 5-5.1).
Proceeding with the coordinate transformation method, global coordinate X is
expressed as a function of local natural coordinate ξ , and an element geometry mapping

function X = X (ξ ) is obtained as follows,

 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-29

X (ξ ) = Xi + x (ξ ) = Xi +
l (
X −X )
(ξ + 1) = Xi + j i (ξ + 1)
2 2
= Xi +
1
2
( X j ξ + X j − X iξ − X i )
1 1
= (ξ + 1) X j − (ξ − 1) Xi
2 2
1 1
= (1 − ξ ) Xi + (ξ + 1) X j
2 2
= N i (ξ ) Xi + N j (ξ ) X j

Or in matrix form, this geometry mapping function is,

⎧ X ⎫
⎪ i ⎪
X (ξ ) = NX where N = ⎢ N i (ξ ) N j (ξ ) ⎥ andX=⎨ ⎬
⎢⎣ ⎥⎦
⎪⎩ X j ⎪⎭
Compare the geometry interpolation function X (ξ ) = NX with the displacement
⌢
interpolation function u (ξ ) = Nd for the two-node element. When element geometry is

interpolated in the same manner as element displacement, i.e., by using the same shape
functions (linear, quadratic, etc.), the corresponding continuum finite element is called an
isoparametric element.
The isoparametric concept will be developed further in the sequel when two- and
three-dimensional curved finite elements are addressed (Chapter 8 forward). Using
higher order shape functions together with the isoparametric concept enables greater
geometrical fidelity when meshing arbitrarily curved edges and surfaces of components
and structures. This is a very important attribute of FEM for it relates directly to
accuracy of simulation. One-dimensional shape functions in natural coordinates, as
discussed here, are the basic building blocks for formulating curved isoparametric
elements.12
5-5.5 Jacobian of the Geometry Mapping X = X ( ξ ). It will become necessary to
produce isoparametric formulations of element strain energy. To do this, element strain
⌢
ε = du dX needs to be expressed in terms of shape functions N( ξ ) and their derivatives.
⌢
Upon substituting for u element strain becomes,

12
The importance of isoparametric elements is not easily appreciated in 1-D. However, it may be glimpsed
later in Problem 6-7 where the geometry of a two-node, tapered bar element is approximated using the
element’s linear shape functions.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-30

d
N (ξ ) d
ε=
dX
Since engineering strain is defined in the global cartesian system X, the chain rule is
needed to carry out this derivative of N( ξ ). That is,

d d dξ
N (ξ ) = N (ξ ) ⋅
dX dξ dX
Therefore, element strain becomes,
d dξ
ε= N (ξ ) ⋅ d
dξ dX
The derivative dξ dX is not available in this expression. However, its inverse dX dξ
is available, i.e., computable, and is denoted J( ξ ),

dX d dN
J (ξ ) ≡ = ( NX ) = X
dξ dξ dξ

()
J( ξ ) is known as the Jacobian of geometry mapping X = X ξ .13 Therefore, element
strain becomes,
d −1 d
ε = J (ξ ) ⋅ N (ξ ) d = B (ξ ) d where B (ξ ) = J (ξ ) N (ξ )
−1
dξ dξ
B is the strain-displacement matrix, relating element strain to its nodal displacements.
Since J( ξ ) is in the denominator of element strain, concern naturally exists for its
robustness. The derivative of the mapping should be smooth and positive over the
domain of an element. Generally, it is sufficiently robust. However, there are exceptions
some of which are important (Section 8-7.1).
Also, since
dX
dX =
dξ
dξ = J ξ dξ ()
the Jacobian has a geometric interpretation. It may be viewed as a length scale factor
between line segments dX and dξ in the cartesian and natural coordinate systems,
respectively.
Finally, in passing, an isoparametric representation of element strain energy
would be expressed in terms of the natural coordinate ξ and the Jacobian J( ξ ) as
follows. In this case, it is for a quadratic element with the three nodes Xi , X j , X k .

13
In 2-D and 3-D, the Jacobian is a square matrix of computable derivatives, denoted J.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-31

⌢ 2
U( ) =
1 ⎛ du ⎞ 1 1 1 ⎛ 1 ⎞
dX = ∫ EA⋅( Bd ) Bd J dξ = dT ⎜ ∫ BT ( EA) B J dξ ⎟ d
e Xk T
2 ∫Xi
EA ⎜ ⎟
⎝ dX ⎠ 2 −1 2 ⎝ −1 ⎠

Section 5-6. Completeness, Compatibility and Convergence. The assumed

⌢ ⌢
element behavior function u ( x ) and its derivative du dx appear in the expression for the

total potential energy functional Π ( e) for a one-dimensional bar element. The former
appears in the potential energy term associated with element loads and the latter in the
strain energy term as follows (e.g., see Section 1-6 or 2-5),
l ⌢ 2 l
( e) ⌢ 1 ⎛ du ⎞ ⌢
Π ( u ) = ∫ EA ⎜ ⎟ dx − ∫ q ( x ) u dx
0
2 ⎝ dx ⎠ 0

It is essential that this function and its derivative be sufficiently smooth so that both
integrals exist on the interval ( 0 ≤ x ≤ l ) , so that in turn the element's total potential
⌢
energy is properly represented and exists. Alternatively, the derivative du dx must be
sufficiently smooth so that the integral appearing in the weak form functional of the solid
mechanics boundary value problem exists (Section 2-6). The existence of the total
potential energy functional or, alternatively, the existence of the weak form functional,
⌢ ⌢
requires this smoothness or continuity in the function u ( x ) and its derivative du dx .14
Completeness is a crucial property of the assumed polynomial, α 1 + α 2 x + ... , for
⌢
the element behavior function u ( x ) . It is necessary for convergence of the finite element
solution in the limit of mesh refinement. The mathematical meaning of completeness is
somewhat restricted in the context of FEM. 15 For our purposes completeness simply
⌢ ⌢
means that the element displacement function u ( x ) and its derivative du dx , i.e., the
element strain ε ( x ) , become constants in the limit of mesh refinement.
⌢
The term α 1 guarantees that u ( x ) is a constant in the limit of mesh refinement.
⌢
Physically, this means that u ( x ) is capable of replicating rigid body displacement (see
⌢
Section 5-4). The term α 2 x guarantees that du dx is constant in the limit of mesh

14
For the two-node beam element, these considerations must be extended to the second derivative or
curvature, since curvature appears in the element's strain energy term (Section 4-3). See, for example,
Chapter 11, L. P. Felton and R. B. Nelson, Matrix Structural Analysis, John Wiley & Sons, Inc., New York,
700 pages, 1997.
15
The term "completeness" derives from the mathematics of functional analysis in connection with metric
spaces. A complete space is one in which every Cauchy (i.e., convergent) sequence converges to a point
within the space. The space of real numbers is a simple example of a complete space.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-32

⌢
refinement. Physically, this means that the function u ( x ) is also capable of replicating a
constant state of strain. For example, for the cubic element, the displacement and strain
behavior functions are,
⌢ ⎫
u(x) = α 1 + α 2 x + α 3 x 2 + α 4 x 3
⎪
2 ⎬
(0 ≤ x ≤ l )
du dx ≡ ε ( x ) = α 2 + 2α 3 x + 3α 4 x ⎪
⎭
⌢
Therefore, u(x) → α 1 and ε ( x ) → α 2 as element length l → 0 . Essentially, the element
is capable of shrinking to a point that displaces as a rigid entity (a distance α 1 ), and at
which position the strain is constant (with value α 2 ).
The first two terms, α 1 and α 2 x , in the assumed polynomial expression for the
⌢
behavior function u ( x ) are sufficient for its completeness. Equivalently, that the
⌢
function u ( x ) is at least a linear polynomial (i.e., α 1 + α 2 x ) provides for completeness.
Completeness is a necessary condition for convergence to the exact solution in the limit
of mesh refinement.
Recall that the displacement based finite element method includes elements that
satisfy displacement compatibility both internally and at their boundaries with contiguous
elements (Section 1-2.2.4). When an element formulation has both the properties of
completeness and compatibility, the element is said to be conforming. This is a sufficient
condition for convergence to the exact solution in the limit of mesh refinement. Further,
convergence is monotonic; meaning that solution accuracy continuously improves from
one side, as the mesh is refined.
Though they are omitted herein, mention should be made of a very useful class of
finite elements known as incompatible elements many of which exist in the libraries of
commercial finite element computer programs. They are displacement-based elements
⌢
but their behavior functions u ( x ) are discontinuous at element boundaries leading to
infinite strain at element interfaces. Convergence of these non-conforming elements has
been demonstrated by a numerical testing protocol known as the patch test.16 In some
cases the convergence rate is actually superior to that of conforming elements (though not
necessarily monotonic). The patch test is a sufficiency test for convergence. It shows
that a small group (i.e., a patch) of non-conforming elements, or in some cases a single
element, when subjected to appropriate load and boundary conditions, can nonetheless
replicate the required rigid body motion and constant strain behavior discussed above for
conforming elements.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-33

⌢
Section 5-7. Strain Energy Lower Bound. Substituting u ( x ) = Nd into the
total potential energy functional for a one-dimensional element yields the function
Π ( e) ( d ) in matrix form,
1
Π ( e) ( d ) = U ( e) ( d ) + V ( e) ( d ) = dT k( e)d − dT fq( e)
2
where as usual,
l T l
⎛ dN ⎞ ⎛ dN ⎞
k ( e)
= ∫⎜ ⎟⎠ EA ⎜⎝ ⎟⎠ dx and fq( e) = ∫ q ( x ) N T dx
0
⎝ dx dx 0

The finite element equation is obtained by invoking the stationary condition of the
principle of minimum total potential energy, i.e.,
∂Π ( e)
=0 ⇒ k( e)d = fq( e)
∂d
Using this finite element equation, the element's total potential energy Π ( e) and its strain
energy U ( e) can be written, respectively,
1 1
Π ( e) = − dT fq( e) and U ( e) = dT fq( e)
2 2
This reveals an opposite relationship between these energies. When the total potential
energy decreases, the total strain energy increases.
Since energy is a scalar, the total potential energy and the total strain energy of a
N N
finite element model with N elements are ∑ Π (e) and ∑U ( ) , respectively.
e
As N → ∞
e=1 e=1

in the limit of mesh refinement, the total potential energy of the system approaches the
exact (minimum) value Π monotonically from
above, and the total strain energy of the system
approaches the exact (maximum) value U
monotonically from below,
N

∑ Π ( ) → Π ( from above )
e

e=1

and
N

∑ U ( ) → U ( from below )
e
Convergence of Energies
e=1

16
O. C. Zienkiewicz and R. L. Taylor, The Finite Element Method, Chapter 10, Vol. 1, Fifth Edition,
Butterworths-Heinemann, Inc., London, 689 pages, 2000.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-34

N N
The term ∑ Π (e) is an upper bound, and the term ∑U ( )e
is a lower bound, as depicted
e=1 e=1

in the convergence sketch above.

The strain energy lower bound
theorem says: a finite element solution
yields a lower bound on the exact strain
energy of the system. The practical value of
this theorem is that the stiffness of a
displacement-based finite element model
(i.e., the matrix K), which is derived from Concept of monotonic convergence
from below for displacements
strain energy alone, is greater than the
stiffness of the structural idealization that it aims to replicate. It follows that computed
displacements (for conforming elements) will converge monotonically from below to
exact displacements, i.e., computed displacements are a lower bound on exact
displacements, as indicated in the sketch.17 Convergence is faster for higher order
elements than for lower order elements (see also Section 5-9).
The properties of completeness and compatibility, the concept of monotonic
convergence, and the strain energy lower bound theorem, discussed here in the context
one-dimension, carry over to two- and three-dimensional FEM.

Section 5-8. Gauss-Legendre Quadrature. Just as the Bernoulli-Euler beam

member stiffness matrix k ( ) was defined as a definite integral over the member's length,
e

finite element stiffness matrices for two- and three-dimensional continuum elements,
derived in later chapters, are defined as definite integrals over the element's area or
volume, depending on element type. Unfortunately or otherwise, element matrices for
only a very few basic finite elements are amenable to closed form integration. Most
elements must be integrated numerically. For example, recall that the order-4 stiffness
matrix for the two-node Bernoulli-Euler beam element was integrated closed form (see
Example 4-2),

17
It's physically reasonable to argue that a system with finite number of DOF is stiffer than a system whose
number of DOF approaches infinity. It follows, that the former system may be viewed as an artificially
constrained system.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-35

⎡ integrated ⎤
L ⎢ ⎥
k ( ) = ∫ BT ( EI ) B dx = ⎢
e
closed form ⎥
⎢ ⎥
0
⎢⎣ (exact ) ⎥⎦4×4
However, the order-n stiffness matrices for most finite elements, which for the present

may be denoted k ( ) = ∫( ) (B EB ) det J dVol , are integrated numerically,

e T

e
Vol

⎡ numerically ⎤
⎢ ⎥
k( ) = ∫( ) ( )
e
BT EB det J dVol ≈ ⎢ integrated ⎥
⎢ ⎥
Vol
e

⎢⎣ (approximate ) ⎥⎦ n×n

( )
This is generally because the integrand BT EB det J will prove to be an n × n array of
rational functions, defying any reasonable attempt at closed form integration.
Numerical integration is "costly" computationally when compared to closed form
integration, usually amounting to a significant fraction of overall compute time in a finite
element analysis. Further, numerical integration of element matrices introduces another
layer of approximation into the overall modeling task known as quadrature error.
However, there is more to the story. As it happens, exact integration even if or when
possible is not necessarily desirable; numerical integration can actually be exploited to
improve accuracy. This is discussed later under the heading of "reduced integration" (see
Chapter 9).
The numerical integration method of choice in FEM is Gauss quadrature. It is
shown to require fewer evaluations of integrand functions for the same level of accuracy
than alternative methods, at least in one dimension. Gauss quadrature rules are no longer
necessarily optimal in two- and three-dimensional problems. 18 In addition, as will be
seen in later chapters, integrand functions are expressed in natural coordinates having
limits of integration that are, it would seem, natural to Gauss quadrature.
Consider numerical evaluation of the integral of a function f (ξ ) over the interval
-1 to 1. The Gauss quadrature formula is,
1 n

∫ f (ξ ) dξ ≈ ∑ wm f (ξ m )
−1 m=1

18
Hughes, T. J. R. (1987). The Finite Element Method, p. 145, Prentice-Hall, Inc. Englewood Cliffs, New
Jersey, 803 Pages, 1987.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-36

where: (a) f (ξ ) is a given function,

(b) wm are weighting factors,
(c) ξ m are natural coordinate values of optimal sampling points,
aka Gauss points in the interval -1 to 1,
(d) n is a prescribed order of numerical integration.

Optimal sampling points ξ m and weighting factors wm and are provided in Table 5-9 for

∑
n
orders of integration n = 1 through n = 5 . Observe that m=1
wm equals 2 for any n.
While Gauss quadrature of order n is clearly a numerical method that in general
yields approximate results, it will in fact integrate exactly all polynomials of degree k,
denoted Pk (ξ ) , where k ≤ ( 2n − 1) . For example, when the specified order of numerical
integration is n = 2 , the following integrals would be evaluated exactly by Gauss
quadrature,
1 1 1 1

−1
∫ P3 (ξ ) dξ ∫ P2 (ξ ) dξ
−1
∫ P1 (ξ ) dξ
−1
∫ P (ξ ) dξ
−1
0

Table 5-9. Sample points and weight factors for Gauss quadrature rules.*
n ξm wm 2n − 1
1 0.0 2.0 1
2 -0.57735027 1.0 3
0.57735027 1.0
3 -0.77459667 0.55555555 5
0.0 0.88888889
0.77459667 0.55555555
4 -0.86113631 0.34785485 7
-0.33998104 0.65214515
0.33998104 0.65214515
0.86113631 0.34785485
5 -0.90617975 0.23692689 9
-0.53846931 0.47862867
0.0 0.56888889
0.53846931 0.47862867
0.90617975 0.23692689
*These rules apply in 2-D and 3-D as well (see Chapters 9 and 10).

Suppose instead the integrand is f ( x ) and the interval is a ≤ x ≤ b . In this event,

∫ f ( x ) dx to the ξ -
b
Gauss quadrature is employed after transforming the integral
a

coordinate system using the transformation,

 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-37

x (ξ ) =
(b − a )ξ + (b + a )
2
Regarding transformation of the interval of integration, observe that,
⎧ a for ξ = −1
⎪⎪
x (ξ ) = ⎨ ( b + a ) 2 for ξ = 0
⎪ b for ξ = 1
⎪⎩
Also, upon differentiating the coordinate Mapping interval of integration
transformation with respect to ξ ,

⎛ b − a⎞
dx = ⎜ dξ
⎝ 2 ⎟⎠
Thus, the transformed integral is obtained by substituting for x and dx and changing the
integration interval to −1 ≤ ξ ≤ 1 as follows,

⎛ b − a⎞ ⎡ (b − a )ξ + (b + a ) ⎤
b 1

∫ f ( x ) dx = ⎜ ∫ f ⎢⎣ ⎥ dξ
a
⎝ 2 ⎟⎠ −1
2 ⎦
Using this equation, Gauss quadrature can be applied directly to the numerical evaluation

∫ f ( x ) dx .
b
of the integral
a

===============================================================
Example 5-3. Evaluate the two integral forms,
1
π 2
∫ N (ξ ) dξ
−1
i and ∫ 0
sin x dx

using Gauss quadrature:

(1) ∫ N (ξ ) dξ .
−1
i Use integration order n = 2 for the linear, quadratic and cubic cases of

()
node i shape function N i ξ . This integral equals the area beneath the graph of N i ξ , ()
but otherwise has no particular significance.
1
(a) Linear case (polynomial degree k = 1 ): N i = (1 − ξ )
2
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-38

1 1
1
∫ N i (ξ ) dξ =
−1
∫ 2 (1 − ξ ) dξ
−1
2
≈ ∑ wm N i (ξ m )
m=1

⎡1 ⎤ ⎡1 ⎤
≈ w1 ⎢ (1 − ξ1 ) ⎥ + w2 ⎢ (1 − ξ 2 ) ⎥
⎣2 ⎦ ⎣2 ⎦
⎡1 ⎤ ⎡1 ⎤
≈ 1.0 ⎢ (1 + 0.57735027 ) ⎥ + 1.0 ⎢ (1 − 0.5773507 ) ⎥
⎣2 ⎦ ⎣2 ⎦
≈ 1.0000...

This result is exact to machine precision for the area under the graph of the given

()
linear polynomial N i ξ , as is easily verified.

1
(b) Quadratic case (polynomial degree k = 2 ): N i = ξ (ξ − 1)
2
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-39

1 1 2
1
∫ N i (ξ )dξ = −1∫ 2 ξ (ξ − 1)dξ ≈ ∑
−1 m=1
wm N i (ξ m )

⎡1 ⎤ ⎡1 ⎤
≈ w1 ⎢ ξ1 (ξ1 − 1) ⎥ + w2 ⎢ ξ 2 (ξ 2 − 1) ⎥
⎣2 ⎦ ⎣2 ⎦
⎡1 ⎤
≈ 1.0 ⎢ ( −0.57735027 ) ( −0.57735027 − 1) ⎥
⎣2 ⎦
⎡1 ⎤
+ 1.0 ⎢ ( 0.5773507 ) ( 0.57735027 − 1) ⎥
⎣2 ⎦
≈ 0.33333...
This result is also the exact area under the graph of the given quadratic

()
polynomial N i ξ , to machine precision.

1
(c) Cubic case (polynomial degree k = 3 ): N i = (1 − ξ ) ( 3ξ + 1) ( 3ξ − 1)
16

1 1 2
1
∫ Ni (ξ )dξ = −1∫ 16 (1 − ξ )( 3ξ + 1)( 3ξ − 1) dξ ≈ ∑ wm N i ( ξ m )
−1 m=1

⎡1 ⎤ ⎡1 ⎤
≈ w1 ⎢ (1 − ξ1 ) ( 3ξ1 + 1) ( 3ξ1 − 1) ⎥ + w2 ⎢ (1 − ξ 2 ) ( 3ξ 2 + 1) ( 3ξ 2 − 1) ⎥
⎣ 16 ⎦ ⎣ 16 ⎦
Observe that ξ m = ∓0.57735027... = ∓ 1 3 ⇒ ξ1 = − 1 3 and ξ 2 = + 1 3
Therefore,

(1 + )( )( )
1

∫ N (ξ )dξ ≈ 1.0 ⎡⎣1 16

−1
i 1 3 1 - 3 -1 - 3 ⎤
⎦

( )(
+1.0 ⎡1 16 1 - 1 3 1 + 3 -1 + 3 ⎤
⎣ ⎦ )( )
≈ 0.25000…

The result is the exact area under the graph of the given cubic polynomial N i ξ ()
to machine precision.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-40

The previous three integral evaluations were exact (within machine precision)
since the integrands in each case are polynomials no higher than degree k = 3 and Gauss
quadrature order n = 2 was employed. The criterion for exact integration, k ≤ ( 2n − 1) ,
was satisfied in all three cases.
π 2
(2) ∫0
sin x dx . Use integration order n = 2 .

The coordinate transformation for this integral is,

x (ξ ) =
(π2 − 0 ) ξ + (π 2 + 0 ) π
= (ξ + 1)
π
(dx = dξ )
2 4 4
π ⎛ b − a⎞ π
a = x ξ =−1 = 0 and b = x ξ =1 = and ⎜ =
2 ⎝ 2 ⎟⎠ 4
Therefore,
f (ξ )
!##" ## $
π ⎡π π 2
1
π 2 ⎤
∫0 sin x dx = ∫ ⎢⎣ 4
4 −1
sin ( ξ + 1) ⎥⎦ dξ ≈ ∑ wm f ( ξ m )
4 m=1
Noting again ξ m = ∓0.57735027... = ∓ 1 3 ⇒ ξ1 = − 1 3 and ξ 2 = + 1 3
Thus,
π 2 π⎧ ⎡π ⎤ ⎡π ⎤⎫
∫ sin x dx ≈ ⎨w1 sin ⎢ (ξ1 + 1) ⎥ + w2 sin ⎢ (ξ2 + 1) ⎥ ⎬
0 4⎩ ⎣4 ⎦ ⎣4 ⎦⎭
π⎧ ⎡π ⎡π
4⎩
≈
⎣4
( ⎤
⎨(1.0 ) sin ⎢ − 1 3 + 1 ⎥ + (1.0 ) sin ⎢
⎦
)
⎣4
⎤⎫
1 3 + 1 ⎥⎬
⎦⎭
( )
π
≈ ⎡⎣sin ( 0.10566 π ) + sin ( 0.3943π ) ⎤⎦
4
≈ 0.99847
Since the function f ( x ) = sin x is not a polynomial of requisite degree, the result
is approximate. Nonetheless, the result compares very well with the exact value
π 2
∫0
sin x dx = 1 using Gauss quadrature.
===============================================================
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-41

Section 5-9. Computational Efficiency. The question of whether to use linear,

quadratic or cubic elements for the bar model is one of computational efficiency.
Clearly, the cubic element is more accurate than the quadratic element, and the quadratic
element is more accurate than the linear element. However, higher order elements
require more computation than lower order elements.
Computational efficiency is a combination of accuracy and compute time. In
practice, most of the compute time for a finite element problem is taken up by the
solution of the structure system of equations KD = Q . The Gaussian elimination
algorithm or, more precisely, one of its direct solver derivatives is usually implemented
to accomplish this task. Relative solution time is taken approximately proportional to
N3/2 where N is the number of equations or equivalently the number of active DOF for a
finite element model of the bar.

Table 5-10. Six alternative finite element models of the bar.*

Model # Finite element models N N3 2

1 4 8

2 8 22.6

3 12 41.6

4 8 22.6

5 6 14.7

6 6 14.7
* White nodes are internal nodes or DOF, no different than black end nodes for tallying the value of N.

It has been shown that computed finite element displacements are a lower bound
on exact displacements for a structural idealization of the bar (Section 5-7).
Alternatively, a finite element model is internally constrained and thus is stiffer than the
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS 5-42

structural idealization it presumes to replicate. Six alternative models for the constrained
bar are compared in Table 5-10 with respect to computational efficiency. Models with
higher N should be more accurate simply because they have more displacement DOF, and
therefore represent models that are less over-stiff (less internally constrained).
Model #3 should be the most accurate model because it has the highest value for
N, just as Model #1 should be the least accurate because it has the lowest value of N.
However, for linear problems, experience has shown that for the same N, models using
higher order elements are more efficient computationally. (See the numerical study of
low and high order elements in Section 7-6). In this respect, Model #6 comprised of two
cubic elements should prove more efficient than Model #5 comprised of three quadratic
elements, even though they both have 6 active DOF and the same relative compute time.
Likewise, Model #2 comprised of four quadratic elements should be more efficient than
Model #4 comprised of eight linear elements, even though they both have eight DOF and
the same relative compute time.
In the past, linear finite elements were sometimes preferred when very large,
linear, finite element models stretched available computing power. In general, this is no
longer the case with modern finite element solvers and computer hardware. Linear finite
elements are still in use for some specialized nonlinear finite element models because
they are generally considered more robust than higher order elements. This is
particularly important in research projects that are computationally intensive. One reason
is that higher-order elements can become difficult or even pathological when internal
nodes (or midside nodes) migrate from their natural position under nonlinear geometric
displacements or distortions of the element.
In general, the higher the element order, the greater the accuracy of the element
per model DOF. However, quadratic finite elements are most popular for linear stress
analysis, providing a good combination of accuracy and acceptable compute time. (See
the numerical study of quadratic elements in Section 8-10). Cubic finite elements are
generally considered unnecessary in engineering stress analysis, and are indeed omitted
from the element libraries of most finite element software products.19

19
The two-node cubic Bernoulli-Euler beam/frame element may be considered an exception (Chapter 4
and Tutorial T4).
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-1

PROBLEM SET 5

Problem 5-1
Use the Lagrange interpolation method to derive the four cubic shape functions compiled
in Table 5-4.

Problem 5-2
() ()
Verify that the quadratic shape functions in natural coordinates N i ξ , N j ξ , N k ξ ()
satisfy the following key properties:
(a) Kronecker delta property, i.e., the shape function N γ ( γ = i, j, k ) has a value of unity
at node γ and a value of zero at the other two nodes.
(b) Partition-of-unity property, i.e., summation of the three shape functions equals unity.

Problem 5-3
Assume the finite element solution D = K −1Q for the subject bar finite element model in
Section 5-2 is,
D = [ 0.24875 0.995000 2.23875 3.98000
6.21875 8.95500 12.18875 15.9200 ] × 10 −3 m
T

Let L = 0.4 m , and for the two points X = 0.0667 m and X = 0.333 m :
⌢
(a) Interpolate the axial displacement u ( X ) using the appropriate interpolation function.
(b) Calculate the average axial strain ε avg near each point.
(c) If u ( X ) = 0.1X 2 is the exact axial displacement, what is the axial strain ε ?

Problem 5-4
(a) Sketch the shape functions for the two-node linear finite element shown indicating
their key properties.

i j

(b) Derive the shape function for node i in terms of natural coordinate ξ , −1 ≤ ξ ≤ 1 .
1
1
(c) Given N j = (ξ + 1) , evaluate ∫ N dξj numerically using order n = 2 Gauss
2 −1
quadrature.

(d) Explain why the result in (c) should either be not exact or exact to machine precision.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-2

Problem 5-5
(a) Sketch the shape functions in natural coordinate ξ for the 3-node quadratic finite
element shown indicating their key properties.
ξ
i j k

(b) Derive the shape function for node i in terms of natural coordinate ξ .
1

(c) Given N j = (1 + ξ ) (1 − ξ ) , evaluate ∫ N dξ j numerically using order n = 2 Gauss

−1
quadrature.
(d) Explain why the result in (c) should either be not exact or exact to machine precision.

Problem 5-6
(a) Sketch the shape function for node k, of the four-node cubic finite element shown,
indicating its key properties.

(b) Using the Lagrange interpolation method, derive the shape function for node k in
terms of natural coordinate ξ .
1
1
(c) Given N l = (ξ + 1) ( 3ξ + 1) ( 3ξ − 1) , evaluate ∫ N dξ
l numerically using order
16 −1
n = 2 Gauss quadrature.

(d) Explain why the result in (c) should either be exact (to machine precision) or an
approximation.

Problem 5-7
The shape functions for nodes i and k of the quadratic, one-dimensional bar finite
1 1
element shown are N i = − ξ (1 − ξ ) and N k = ξ (1 + ξ ) , respectively.
2 2

x ξ

i l 2 l 2
j k
ξ = −1
ξ=0 ξ=1
(a) Derive the shape function N j (ξ ) .
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-3

⌢
(b) Calculate the axial displacement u at ξ = 0.5 given the following nodal
displacements in units of micrometers,
ui = 2 uj = 5 uk = 3
(c) For the displacements given in part (b), find the axial strain ε at ξ = 0.5 if the
2⎛ l⎞
element length is l = 1 cm . (Hint: The transformation ξ = ⎜ x − ⎟ may be useful.)
l⎝ 2⎠
(d) For the displacements given in part (b), find the axial stress σ ( e) at ξ = 0.5 if Young's
modulus is E = 200 GPa .

Problem 5-8
Specialty elements are not uncommon in FEM. The special element formulation
introduced in this problem is the basis for so-called quarter-point elements useful in
linear elastic fracture mechanics for numerically analyzing stresses around crack tips.1
()
(a) Obtain the geometry mapping X ξ for the basic 3-node quadratic bar element
having nodes positioned at,
X = Xi X j = Xi + l 2 Xk = Xi + l
(b) That interior nodes such at Xj do not have to be evenly spaced is an advantage of
isoparametric elements. Let the nodes of a special 3-node bar element be positioned at,
X = Xi X j = Xi + l 4 Xk = Xi + l
()
Compute the geometry mapping X ξ , and its Jacobian J ξ = dX dξ . ()
⌢
(c) Show that the strain ε = du dX is unbounded at node i of the specialty element.
() () ( )
Hint: first compute ε ξ , and then invert the mapping X ξ for ξ X to find ε X . ( )
(d) What is the order a of the singularity as X → X , i.e., what is a such that ε ( r ) ∼ r −α
i

as r ≡ X − X i → 0 ?

Problem 5-9
Consider an isoparametric 3-node bar element with natural geometry as shown.

(a) Explain the meaning of the “Kronecker delta” and “partition-of-unity” properties for
this element.
1
Barsoum, R. S. (1976). “On the Use of Isoparametric Finite Elements in Linear Fracture Mechanics,”
International Journal of Numerical Methods in Engineering, Vol. 10, 1976, pp. 25-37.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-4

(b) Use the Lagrange interpolation method to derive the quadratic shape functions in
terms of natural coordinates. Show that they may be reduced to this form,

⎢ 1 1 ⎥
N = ⎢ N 1 (ξ ) N 2 (ξ ) N 3 (ξ ) ⎥ = ⎢ ξ (ξ − 1) − (ξ + 1) (ξ − 1) ξ (ξ + 1) ⎥
⎣ ⎦
⎣ 2 2 ⎦
(c) Assuming a distorted element, meaning non-uniform spacing of nodes ( Xi , X j , X k ) in
the element’s physical geometry, compute the mapping function X(ξ) and its jacobian
J(ξ) = dX/dξ. Then specialize the result to the case where the element is undistorted, i.e.,
X j = ( Xi + X k ) 2 , so that nodes are uniformly spaced.

(d) In the FEM formulation of 1-D elastodynamics, the quadratic element mass matrix
emerges in the form
Xk 1

m( e) = ρ ∫ N T N dX = ρ ∫ N T N J d ξ
Xi −1

where ρ is mass density. What Gauss quadrature rule n would be needed to evaluate this
matrix exactly for an undistorted element. Explain your answer. How (if at all) does this
change if the element is distorted?

Problem 5-10
()
xk =l
For the quadratic shape function N i , evaluate the integral ∫ xi =0
N i x dx :
(a) In terms of cartesian coordinate x,

(b) In terms of natural coordinate ξ after transforming,

(c) Using order n = 2 Gauss quadrature,

(d) Explain why the result from part (c) should either be an approximation or exact (to
machine precision).

Problem 5-11
(a) Generate the geometry of a solid object of your choosing that is comparable feature-
wise to the Pin's geometry in Example T5-1, using a CAD system. Document your result
by printing the solid object.

(b) Generate the geometry of the same solid object using Abaqus/CAE. Document your
result by printing the solid part.

(c) Compare the experience with solid geometry generation with the CAD system and
Abaqus/CAE.

(d) Import the solid object created in (a), and mesh this part with Abaqus/CAE.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-5

Problem 5-12
Referring to Example T5-2, use Abaqus/CAE to defeature the CAD bracket geometry in
the STEP file, prt_203.stp, to produce FEA bracket geometry as follows:

CAD Geometry FEA Geometry

(a) Use the standalone interface to import the part. Provide a printout of the imported
CAD geometry.

(b) Determine whether there are invalid entities and/or imprecise entities. If necessary,
repair the part to precise geometry. Provide a printout of the repaired CAD geometry.

(c) Delete edges smaller than 1.0 units using the Repair small edges tool. Provide a
zoom-in screen print of highlighted small edges.

(d) Remove the more complex of the two bosses next to one another, using the Replace
faces tool with Extend neighboring faces turned on. Using the same tools, fill in the
remaining boss holes. Provide a printout of the defeatured or FEA geometry.

(e) Free mesh the part with tetrahedral element shapes using a global seed size of 6 units,
and provide a printout of the meshed part.

Problem 5-13
(a) Repeat Example T5-3 using the needle bearing assembly or another assembly of your
choosing. Provide a printout of the CAD system geometry, and the imported
Abaqus/CAE part.

(b) Suppress any feature/s in the CAD system (other than InnerRace, in the case of needle
bearing assembly). Document this with a print out of the CAD system geometry and the
regenerated Abaqus/CAE part.

(c) Configure an analysis of the part including loads and boundary conditions using
Abaqus/CAE. Document with printout/s of the part showing loads and boundary
conditions.

(d) Modify the geometry of the part in the CAD system. Document the modified
geometry in both the CAD system and in Abaqus/CAE showing the regenerated loads
and boundary conditions.

(e) Mesh the modified part, and document by printing the meshed part.
 ONE-DIMENSIONAL FINITE ELEMENT SHAPE FUNCTIONS P5-6

Problem 5-14
(a) Generate the geometry of an object of your choosing (wire, planar or solid topology)
using your CAD system. Document your result by printing the object.

(b) Generate the geometry of the same object using Abaqus/CAE. Document your result
by printing the resulting part geometry.

(c) Compare the experience with geometry generation using CAD and Abaqus/CAE.

(d) Import the object created in (a), and mesh the part using Abaqus/CAE. Document
this result by printing the meshed part.