Workshop Exercises n°6:

Isoparemetric elements

Theoretical Reminders

Laruelle Cédric Volvert Martin

Assistant Assistant
[email protected] [email protected]
Bât. B52/3 +2/541 Bât. B52/3 +2/418

Author: Gaëtan Wauthelet

May 11, 2022

rev. 9 2

1 Goal
Represent complex geometries with a limited number of elements.

2 Change of variables: geometric aspect

BIJECTION
−→

Real element Reference element

Structural coordinates x Reduced coordinates ξ

i.e.
(1)


x=T ξ
This transformation must be:
ˆ one to one,

ˆ polynomial,

ˆ boundaries must match.

2.1 Transformation modes

One has
(2)



x ξ =M ξ a
where a is the vector of the unknowns of the transformation.
rev. 9 3

Example: 2D Triangle (3 connectors)

Figure 1: Example of a 2D triangular element

Three parameters are required for each dimension:

 
a
 1
 a2 
 
    
 x = a +a ξ+a η  
1 2 3  a3  1 ξ η 0 0 0
  and M =
i.e. a =    
 y = b +b ξ+b η  b1  0 0 0 1 ξ η
1 2 3
 
 
 b2 
 
b3

2.2 Connectors
The line vector of the connectors xN of length ndim. × nconn. is known: they are the
coordinates of the nodes.

xN = Q a (3)

Example of the 2D triangle:

   
x 1 0 0 0 0 0
x 1 = a1  1  
x2  1 1 0 0 0 0
   
x 2 = a1 + a2    
x 3 = a1 + a3  
x3 

1 0 1 0 0

0
y 1 = b1 i.e N
x =
 
 and Q = 



y 2 = b1 + b2  y1  .0 0 0 1 0 0
   
y 3 = b1 + b3  
 y2 

0 0 0 1 1

0
   
y3 0 0 0 1 0 1
rev. 9 4

2.3 Change of variables

Eq. (3) yields a = Q−1 xN and (2) becomes:

x ξ = M ξ Q−1 xN (4)



| {z }
N (ξ )
c

i.e.



x ξ = N c ξ xN
N c ξ = M ξ Q−1

Example of the 2D triangle:

ˆ The relations are inverted:

 
1 0 0 0 0 0
a1 = x1  
−1 1 0 0 0 0
 
a2 = x2 − x1  
= x3 − x1 −1
 
a3 0 1 0 0 0
i.e Q−1 =  
b1 = y1 
 .0 0 0 1 0

0
b2 = y2 − y1
 
0 0 −1 1
 
0 0
b3 = y3 − y1  
0 0 0 −1 0 1

ˆ The change of variables is rewritten:

  
x = (1 − ξ − η) x + ξx + ηx 1−ξ−η ξ η 0 0 0
1 2 3
i.e. Nc =  
y = (1 − ξ − η) y + ξy + ηy 0 0 0 1−ξ−η ξ η
1 2 3
rev. 9 5

3 Shape function of the reference element : Physical as-

pect
The displacements eld writes (cf. theoretical reminder 4):

(5)


u=N ξ q
where N is the matrix of the physical shape functions.

Geometry | Physical eld |

→ If degree N c = degree N ⇒ ISOPARAMETRIC
| |
→ If degree N c < degree N ⇒ UNDER-PARAMETRIC
ex: straight edges (bilin- | biquadratic displacement |
ear)
→ If degree N c > degree N ⇒ SUPER-PARAMETRIC
ex: parabolic edges | bilinear displacements | Conformity problem (i.e.
of displacements continu-
ity along the chord)

Table 1: Summary of the types of transformation.

rev. 9 6

4 Stiness matrix of the reference element

In general, the stiness matrix of an element writes (cf. theoretical reminder 4):
T
∂N ∂N
Z 
K= D dV (6)
V ∂x ∂x

4.1 Change of variables

ˆ Derivatives with respect to the new variables:
   
∂ ∂
 ∂x    ∂ξ

 ∂y  = J −1  ∂η
∂  ∂ 
 (7)
   
∂ ∂
∂z ∂θ
,

where J is the Jacobian matrix of the change of variables, such that:

∂xj ∂ξ
Jij = and Jij−1 = j .
∂ξi ∂xi

ˆ Integral with respect to the new variables:

Z Z +1 Z +1 Z +1
dV = det J dξ dη dθ. (8)
V
| −1 {z
−1 −1
}
Boundaries are now fairly simple!
Remarks:

1. Computing J is straightforward: Jij = ∂N jk N

∂ξi
xk
BUT the calculation of J may only be carried out NUMERICALLY!
−1

2. J must be invertible (bijective transformation) i.e. det J 6= 0.

3. det J > 0 everywhere for the element since dV = det J dξ dη dθ > 0.

Example of the 2D triangle:

   
a2 b2 x − x1 y2 − y1
J = = 2 
a3 b3 x3 − x1 y3 − y1

⇒ det J = a2 b3 − a3 b2 = (x2 − x1 ) (y3 − y1 ) − (x3 − x1 ) (y2 − y1 )

 
1 b 3 b2
J −1 =  
a2 b 3 − a3 b 2 a3 a2
rev. 9 7

4.2 In reduced coordinates

The operator is rewritten ∂
∂x
as a function of the reduced coordinates:

∂ˆ
∂ξ
.  
∂
0
 ∂x 
Reminder: = ∂u
∂x
i.e. in 2D ∂
∂x
=0
 ∂ .
∂y 

∂ ∂
∂y ∂x

!T
Z +1 Z +1 Z +1 ˆ
∂N ˆ
∂N
⇒ K= D detJ dξ dη dθ (9)
−1 −1 −1 ∂ξ ∂ξ

This integral contains non-explicit terms in (ξ, η, θ)

⇒ Dicult to integrate analytically

⇒ COMPUTED NUMERICALLY!

5 Numerical integration method

Since the integrand F (ξ, η, ζ) above the integral sign of the elementary stiness matrix Ke
is no longer an explicit function of (ξ, η, ζ) in general, a numerical integration method has to
be performed in order to get the elementary stiness matrix.

Denition: If f (ξ) is a continuous function dened on the interval [−1, 1], the quadra-
ture formula n
X
I(f ) = wi f (ξi )
i=1

is dened by the given set of n points −1 ≤ ξ1 < ξ2 < . . . < ξn ≤ 1 named integration points
and the given set of n real numbers w1 , w2 , . . . , wn named weight of the quadrature formula.
These n points Rand these n weights will be determined such that I(f ) is an numerical ap-
proximation of −11
f (ξ) dξ .

Numerical integration in two- or three-dimensions can be accomplished where direction

splitting is applied. Thus, the elementary stiness matrix dened by the three-dimensions
integral over the reference cube region is computed by using the quadrature formula. To this
end, we apply successively the one-dimension quadrature formula direction-by-direction:
rev. 9 8

Z 1 Z 1 Z 1
Ke = F (ξ, η, ζ) dζdηdξ
−1 −1 −1
Z 1 Z 1 n
X
= wk F (ξ, η, ζk ) dηdξ
−1 −1 k=1
Z n X
1 X n
= wj wk F (ξ, ηj , ζk ) dξ
−1 j=1 k=1
n X
X n X n
= wi wj wk F (ξi , ηj , ζk ) ,
i=1 j=1 k=1

where wi , wj and wk are the weights of each integration point (ξi , ηj , ζk ) where F is computed.

5.1 Gauss-Legendre quadrature rule

Contrary to classical quadrature formula such as the trapezoid rule, the midpoint rule or
Simpson's rule, the Gauss's idea is to choose the integration points ξ1 , ξ2 , . . . , ξn such that
n Z 1
the quadrature formula I(f ) = wi f (ξi ) is exactly equal to f (ξ) dξ for polynomials f
X

i=1 −1
of degree r as higher as possible. In other words, the objective is to maximize the order of
accuracy of the quadrature rule for a xed number of integration points n whose locations
are no longer dened arbitrary by the user.

5.1.1 Legendre polynomials

Denition The Legendre polynomial of degree n is dened by:
1 dn  2
n 
Ln (ξ) = ξ − 1
2n n! dξ n
Thus, ones has, if ξ ∈ R:
3ξ 2 − 1
L0 (ξ) = 1, L1 (ξ) = ξ, L2 (ξ) = , ...
2
The Legendre polynomials L0 , L1 , L2 , . . . , verify a lot of properties. Among these, one demon-
strates the following ones useful to demonstrate the Gauss-Legendre quadrature formula.

Theorem The Legendre polynomials L0 , L1 , L2 , . . . , verify the following properties:

1. L0 , L1 , L2 , . . . , Ln generate a basis of Pn * ;
Z 1
2. If i 6= j , then Li (ξ)Lj (ξ) dξ = 0; (orthogonality properties)
−1

*P is the vector space generated by all the polynomials of degree up to n. It is well known that Pn is a
n
vector space of dimension (n + 1) and its canonical basis is given by 1, t, t2 , t3 , . . . , tn .
rev. 9 9

3. Ln has exactly n distinct real roots between −1 and 1. These roots are named Gauss
points.

Proof
1. It can be easily veried that Lj (ξ) is actually a polynomial of degree j and thus
L0 , L1 , L2 , . . . , Ln are linearly independant. So they generate a basis of Pn .

2. Suppose that i > j . One gets then by integrating by parts :

1 1
di h 2
i i dj h 2
Z Z
1
j i
Li (ξ)Lj (ξ) dξ = (i+j) ξ −1 ξ −1 dξ
−1 2 i! j! −1 dξ i dξ j
(
1 di−1 h 2
i i dj h 2
j i ξ=1
= (i+j) ξ −1 ξ −1
2 i! j! dξ i−1 dξ j ξ=−1
Z 1 i−1 h j+1

d 2

i d
i h
2

j i
− i−1
ξ −1 ξ −1 dξ
−1 dξ dξ j+1

Since (ξ 2 − 1)i has a root of order i at 1 and at −1, the (i − 1)th derivative of (ξ 2 − 1)i
is null at ξ = 1 and at ξ = −1. Thus one gets
1 1
di−1 h 2
i i dj+1 h 2
Z Z
(−1)
j i
Li (ξ)Lj (ξ) dξ = (i+j) ξ − 1 ξ − 1 dξ
−1 2 i! j! −1 dξ i−1 dξ j+1

By integrating by parts j times as above, one gets

1 1
(−1)j di−j h 2
i i d2j h 2
Z Z

j i
Li (ξ)Lj (ξ) dξ = (i+j) ξ −1 ξ −1 dξ
−1 2 i! j! −1 dξ i−j dξ 2j
| {z }
(2j)!
j Z 1 i−j
(−1) (2j)! d h
2

i i
= ξ −1 dξ
2(i+j) i! j!
−1 dξ
i−j

ξ=1
(−1)j (2j)! di−j−1 h 2
i i
= ξ − 1 =0
2(i+j) i! j! dξ i−j−1 ξ=−1

3. Consider the points ξ1 , ξ2 , . . . , ξs taken strictly between −1 and 1 at which Ln changes

its sign. Clearly these points will be roots of Ln and so one has s ≤ n . If it is assumed
that
p(ξ) = (ξ − ξ1 )(ξ − ξ2 )(ξ − ξ3 ) . . . (ξ − ξs )
then p ∈ Ps and, since p changes also its sign at the points ξj , 1 ≤ j ≤ s, one gets
p(ξ)Ln (ξ) ≥ 0, ∀ξ ∈ [−1, 1] or p(ξ)Ln (ξ) ≤ 0, ∀ξ ∈ [−1, 1]. In all the cases, since
p(ξ)Ln (ξ) is not identically null, one has
Z 1
p(ξ)Ln (ξ) dξ 6= 0
−1
rev. 9 10

By using the rst property, one notice that there are α0 , α1 , α2 , . . . , αs such that
s
X
p(ξ) = αj Lj (ξ)
j=0

and by using the second property, one gets

Z 1 s
X Z 1 Z 1
p(ξ)Ln (ξ) dξ = αj Lj (ξ)Ln (ξ) dξ = αs Ls (ξ)Ln (ξ) dξ
−1 j=0 −1 −1

Z 1
Since p(ξ)Ln (ξ) dξ is not null, one has necessarily s = n and therefore the n roots
−1
of Ln are ξ1 , ξ2 , . . . , ξn .

5.1.2 Gauss-Legendre quadrature theorem

Denition The quadrature formula
n
X
I(f ) = wi f (ξi )
i=1

is the Gauss-Legendre quadrature formula with n integration points if

1. the integration points ξ1 < ξ2 < . . . < ξn are the n roots of Legendre polynomial Ln of
degree n, i.e. the Gauss points ;
2. the weights w1 , w2 , . . . , wn are dened by
Z 1
wi = lj (ξ) dξ, i = 1, 2, . . . , n
−1

, where l1 , l2 , . . . , ln are the Lagrange basis of Pn−1 associated with n Gauss points.
With the previous properties of Legendre polynomials, one can demonstrate the following
theorem :

Theorem : The Gauss-Legendre quadrature formula with n integration points (n is

an integer ≥ 1) is exact for polynomials of degree up to r = 2n − 1.
n
Proof Consider the Gauss-Legendre quadrature formula I(f ) = wi f (ξi ) with n in-
X

i=1
tegration points and f a polynomial of degree r = 2n − 1. Clearly, it can be dened for ξ ∈ R
: n X
f˜(ξ) = li (ξ)f (ξi )
i=1

, where
(ξ − ξ1 ) · · · (ξ − ξi−1 )(ξ − ξi+1 ) · · · (ξ − ξn )
li (ξ) =
(ξi − ξ1 ) · · · (ξi − ξi−1 )(ξi − ξi+1 ) · · · (ξi − ξn )
rev. 9 11

is the Lagrangian polynomial which generates the Lagrange basis of Pn−1 associated with
the Gauss points. In other words, the polynomial f˜ is the interpolation polynomial of f of
degree n − 1 generated from the n Gauss points ξ1 , ξ2 , . . . , ξn :

li (ξj ) = δij ⇒ f˜(ξi ) = f (ξi )

Consider the error polynomial q dened by :
q(ξ) = f (ξ) − f˜(ξ) ∀ξ ∈ R.

The error polynomial q is a polynomial of degree 2n−1 which cancels at the points ξ1 , ξ2 , . . . , ξn ,
i.e. q(ξi ) = 0 if i = 1, 2, . . . , n. Therefore q can be divided by a polynomial v of degree n
dened by :
v(ξ) = (ξ − ξ1 )(ξ − ξ2 )(ξ − ξ3 ) . . . (ξ − ξn ) ∀ξ ∈ R,
i.e. there is a polynomial w of degree n − 1 such that
q(ξ) = v(ξ)w(ξ) ∀ξ ∈ R.

Since v is a polynomial of degree n which cancels at the n roots of Ln (Legendre polynomial

of degree n), there is a real number α such that
v(ξ) = αLn (ξ) ∀ξ ∈ R.

Since w is a polynomial of degree n − 1, it can be expressed in the Legendre polynomials

basis. So, there are β0 , β1 , β2 , . . . , βn−1 ∈ R such that
n−1
X
w(ξ) = βk Lk (ξ).
k=0

So, by using the property of orthogonality of Legendre polynomials, one gets,

Z 1 Z 1 n−1 Z
X 1
q(ξ) dξ = v(ξ)w(ξ) dξ = α Ln (ξ)Lk (ξ) dξ = 0.
−1 −1 k=0 −1

By denition of q , it has be proven that

Z 1 Z 1
f (ξ) dξ = f˜(ξ) dξ,
−1 −1

and nally, by denition of f˜, one gets

Z 1 n
X Z 1 n
X
f (ξ) dξ = f (ξi ) li (ξ) dξ = wi f (ξi ) = I(f ).
−1 i=1 −1 i=1

This last relation is exactly what we want to demonstrate.

rev. 9 12

5.1.3 Gauss Theorem

A single variable polynomial function of degree r may be integrated exactly by:
Z 1 n
X
f (x) dx = wi f (xi ) + Rn
−1 i=1

i.e. by evaluating the function f (x) in n optimized points (Gauss Point)

where:
ˆ n is the number of points where f is evaluated, such that r ≤ 2n − 1 ;

ˆ wi are the weight associated to each Gauss point;

ˆ Rn is the residual for n Gauss points;

ˆ wi and xi are free ⇒ 2n parameters.

5.2 Application of Gauss method for the calculation of the stiness

matrix
Approximation of the integrand F The integrand F ξ of degree r = 2n − 1 must



ξ
be approximated along each dimension by a polynomial function, i.e. by the sum of:

ˆ an interpolating polynomial of degree n − 1 (Lagrange interpolation)

n
(10)
X
P (ξ) = lj (ξ) F (ξj ),
j=1

where
(ξ − ξ1 ) · · · (ξ − ξj−1 )(ξ − ξj+1 ) · · · (ξ − ξn )
lj (ξ) = . (11)
(ξj − ξ1 ) · · · (ξj − ξj−1 )(ξj − ξj+1 ) · · · (ξj − ξn )
Remark:

lj (ξi ) = δij ⇒ ψ(ξi ) = F (ξi ). (12)

ˆ a correction polynomial

E(ξ) = F (ξ) − P (ξ) = V (ξ) W (ξ) , (13)

| {z } | {z }
order n order n−1

where
V (ξ) = (ξ − ξ1 ) · · · (ξ − ξn ). (14)
Remark:

V (ξi ) = 0. (15)
In other words, the integrand F (ξ) may be written under the following form:
n
(16)
X
F (ξ) = lj (ξ)F (ξj ) + V (ξ)W (ξ).
j=1
rev. 9 13

Integration of the polynomial form

Z 1 n Z 1 Z 1
F (ξ) dξ ∼ (17)
X
= lj (ξ) dξ F (ξj ) + V (ξ)W (ξ) dξ
−1 j=1 −1 −1
| {z } | {z }
wj Rn

Choice of the integration points

The integration points ξj are selected in order obtain an null error (or at least minimum)
in the case of a polynomial of degree r = 2n − 1, i.e. a residual Rn minimum. If the
weight polynomial W (ξ) is decomposed in Legendre polynomial base Li (ξ) where i denotes
the maximum of the Legendre polynomial:
n−1
(18)
X
W (ξ) = αi Li (ξ).
i=0

Therefore, the integration points ξj are as such that:

Z 1
V (ξ)Li (ξ)dξ = 0 for j = 0, 1, · · · , n − 1 ⇒ n conditions, (19)
−1

where ξj are the roots of the Legendre polynomial of order n.

In this way, one may determine n coordinates of Gauss points in each dimension, i.e.,
if ndim denotes the number of dimensions of the problem at hand, F (ξ) will be evaluated at:

nndim Gauss points

Calculation of weights associated to the Gauss points

The weights are selected so as to respect the F values at the integration points:
Z b
wj = lj (ξ) dξ. (20)
a
rev. 9 14

Table of the Gauss points and their weights

Let there be:

Z +1 n
X
f (r) dr = wi f (ξi ).
−1 i=1

n ξi wi
1 0.000000000000000 2.000000000000000

2 ±0.577350269189626 1.000000000000000

3 ±0.774596669241483 0.555555555555556

0.000000000000000 0.888888888888889

4 ±0.861136311594053 0.347854845137454

±0.339981043584856 0.652145154862546

5 ±0.906179845938664 0.236926885056189

±0.538469310105683 0.478628670499366

0.000000000000000 0.568888888888889

6 ±0.932469514203152 0.171324492379170

±0.661209386466265 0.360761573048139

±0.238619186083197 0.467913934572691

Table 2: Gauss-Legendre integration points and associated weights

The Gauss integration points and the corresponding weights are given in the numerical
table (2). Knowing the Gauss quadrature rule for n integration points, it can be proved that
the magnitude of the global error Rn of the integration of an arbitrary given function f (ξ)
over [−1, 1] is bounded by :

Z 1 n
X
f (ξ) dξ − wi f (ξi ) ≤ Ch2n
−1 i=1
| {z }
Rn

, where f (ξ) is supposed to be 2n continuously dierentiable, C is a constant that does not

depend on the integration points ξi , i = 1, . . . , n chosen to divide the domain [−1, 1] and
h = max |ξi+1 − ξi |.
0≤i≤n−1

For example, we will determine the Gauss-Legendre quadrature rule with 2 integration
points.
rev. 9 15

First, we have the Legendre polynomial L2 (ξ) = 12 (3ξ 2 − 1) with its two roots :
1 1
ξ1 = − √ and ξ2 = √
3 3
The Lagrange basis l1 , l2 of P2 associated to the points ξ1 and ξ2 is dened by :
√ √
1− 3ξ 1+ 3ξ
l1 (ξ) = and l2 (ξ) =
2 2
Thus, the weights associated with the integration points are :
Z 1 Z 1
w1 = l1 (ξ) dξ = 1 and w2 = l2 (ξ) dξ = 1
−1 −1

Therefore, the Gauss-Legendre quadrature formula is :

   
1 1
I(f ) = f − √ +f √
3 3
Remark :

ˆ The Gauss-Legendre quadradure formula introduced have all the integration points
that lies strictly between −1 and 1. The Gauss-Legendre-Lobatto quadrature formula
takes into account of the integration interval end points −1 and 1 and the roots of
the rst derivative of Legendre polynomial Ln . Finally, it can be also proved that
Gauss-Legendre-Lobatto quadrature formula integrate exactly polynomials of degree
2n − 1.

ˆ There are Gauss quadrature formula based on the Tchebyche polynomials instead of
Legendre polynomials ;
ˆ In the case where the integration domain is semi-innite or innite, there are Gauss
quadrature formula based on the Laguerre or Hermite polynomials.

5.3 Two-dimension example

Consider a quadrangular element with a thickness t

isparametric
−→
transformaion
rev. 9 16

Change of variables 4 connectors ⇒ 4 parameters (degree 1)


 x = a + a ξ + a η + a ξη
1 2 3 4
 y = b + b ξ + b η + b ξη
1 2 3 4
 
a2 + a4 η b 2 + b 4 η
⇒ J =  → linear
a3 + a4 ξ b3 + b4 ξ

Stiness matrix
Z Z +1 Z +1
K= T
∂N D ∂N dV = t ˆ T D∂N
∂N ˆ det J dξ dη
V −1 −1 | {z }
F (ξ,η)

Gauss integration F (ξ) is evaluated at nndim = n2 Gauss points, and one has:
n
X
K≈ wi wj F (ξi , ξj ),
i,j=1

where n is such that r ≤ 2n − 1

with r → degree of: ˆ T

∂N D ∂N det J
(If ˆ of order 1)
N of order 2 ⇒ ∂N ↓ ↓ ↓ ↓
⇒ r= 1 + 0 + 1 + 1 =3
⇒ n= 2

Therefore, there are two Gauss points per dimension, i.e. 2 × 2 = 4 Gauss points„ .
Referring to the table 2, one has:
√
3
ξ1 , ξ2 = ± = ±0.57735.
3
The four Gauss points are
√ √ 
3 3
1 → (ξ1 , ξ1 ) = , ,
 3√ 3√ 
2 → (ξ2 , ξ1 ) = − 3 , 33 ,
 √3 √ 
3 → (ξ2 , ξ2 ) = − 33 , − 33 ,
√ √ 
3 3
4 → (ξ1 , ξ2 ) = 3
, − 3
.

The corresponding weights are (cf Tab. 2):

„ 3D, with n=2 , one has 2×2×2=8 Gauss points
rev. 9 17

w1 = w2 = 1.

⇒ In the end, the stiness matrix can be rewritten

√ √ ! √ √ ! √ √ ! √ √ !
3 3 3 3 3 3 3 3
K=F , +F − , +F − , − +F , −
3 3 3 3 3 3 3 3
| {z } | {z } | {z } | {z }
1 2 3 4

Figure 2: Representation of the Gauss points for a quadrangular element.