MANE 4240 & CIVL 4240Introduction to Finite Elements

Numerical Integration in 1D

Prof. Suvranu De

Reading assignment:

Lecture notes, Logan 10.4

Summary:

• Newton-Cotes Integration Schemes•Gaussian quadrature

Axially loaded elastic bar

A(x) = cross section at xb(x) = body force distribution (force per unit length)E(x) = Young’s modulusx

y

x=0 x=Lx

F

2

1

TB EB dxx

xk A

Element stiffness matrix2

1i jB EB dx

x

ij xk A

1 2

d1xd2x

For each elementx1

x2

where i

( )B idN x

dx

dxbN2

1x

x

Tbf

Element nodal load vector

dxbN2

1i

x

xibf

2 2 2

1 1 1

T

2 2

2 1 2 1

1 1 1 11 1B EB dx AEdx AEdx

1 1 1 1x x x x

x x x

x x xA

Only for a linear finite element

Question: How do we compute these integrals using a computer?

Any integral from x1 to x2 can be transformed to the following integral on (-1, 1)

1

1)( dfI

Use the following change of variables

21 2

1

2

1xxx

Goal: Obtain a good approximate value of this integral1. Newton-Cotes Schemes (trapezoidal rule, Simpson’s rule, etc)2. Gauss Integration Schemes

NOTE: Integration schemes in 1D are referred to as “quadrature rules”

Trapezoidal rule: Approximate the function f by a straight line gthat passes through the end points and integrate the straight line

f

-1 1

f

fg

)1(2

1)1(

2

1)( ffg

)1()1()()(1

1

1

1

ffdgdfI

•Requires the function f(x) to be evaluated at 2 points (-1, 1) • Constants and linear functions are exactly integrated• Not good for quadratic and higher order polynomials

How can I make this better?

Simpson’s rule: Approximate the function f by a parabola gthat passes through the end points and through f(0) and integrate the parabola

( 1) (1 )( ) ( 1) (1 )(1 ) (0) (1)

2 2g f f f

)1(3

1)0(

3

4)1(

3

1)()(

1

1

1

1

fffdgdfI

f

-1 1

f

fg

f(

•Requires the function f(x) to be evaluated at 3 points (-1,0, 1) • Constants, linear functions and parabolas are exactly integrated• Not good for cubic and higher order polynomials

How to generalize this formula?

Notice that both the integration formulas had the general form

Trapezoidal rule:M=2

11

11

22

11

W

W

Simpson’s rule:M=3

Accurate for polynomial of degree at most 1 (=M-1)

13/1

03/413/1

22

22

11

W

WW Accurate for polynomial of

degree at most 2 (=M-1)

M

iii fWdfI

1

1

1)()(

Weight Integration point

Generalization of these two integration rules: Newton-Cotes

• Divide the interval (-1,1) into M-1 equal intervals using M points• Pass a polynomial of degree M-1 through these M points (the value of this polynomial will be equal to the value of the function at these M-1 points)• Integrate this polynomial to obtain an approximate value of the integral

f

-1 1

f

f

g

With ‘M’ points we may integrate a polynomial of degree ‘M-1’exactly.

Is this the best we can do ?

With ‘M’ integration points and ‘M’ weights, I should be able to integrate a polynomial of degree 2M-1 exactly!!Gauss integration rule

See table 10-1 (p 405) of Logan

Gauss quadrature

M

iii fWdfI

1

1

1)()(

Weight Integration point

How can we choose the integration points and weights to exactly integrate a polynomial of degree 2M-1?

Remember that now we do not know, a priori, the location of the integration points.

Example: M=1 (Midpoint qudrature)

How can we choose W1 and 1 so that we may integrate a (2M-1=1) linear polynomial exactly?

10)( aaf

)()( 11

1

1 fWdfI

0

1

12)( adf

But we want

101011

1

1)()( WaWafWdf

Hence, we obtain the identity

1111002 WaWaa

For this to hold for arbitrary a0 and a1 we need to satisfy 2 conditions

0:2

2:1

11

1

WCondition

WCondition

i.e., 0;2 11 W

f

-1 1

f

g

For M=1 )0(2)(1

1fdfI

Midpoint quadrature rule: • Only one evaluation of fis required at the midpoint of the interval.• Scheme is accurate for constants and linear polynomials (compare with Trapezoidal rule)

Example: M=2

How can we choose W1 ,W2 1 and 2 so that we may integrate a polynomial of degree (2M-1=4-1=3) exactly?

33

2210)( aaaaf

)()()( 2211

1

1 fWfWdfI

20

1

1 3

22)( aadf

But we want

322

3113

222

211222111210

2211

1

1)()()(

WWaWWaWWaWWa

fWfWdf

Hence, we obtain the 4 conditions to determine the 4 unknowns (W1 ,W2 1 and 2 )

0:4

3

2:3

0:2

2:1

322

311

222

211

2211

21

WWCondition

WWCondition

WWCondition

WWCondition

Check that the following is the solution

3

1;

3

1

1

21

21

WW

f

-1 1

For M=2 )3

1()

3

1()(

1

1ffdfI

• Only two evaluations of fis required.• Scheme is accurate for polynomials of degree at most 3 (compare with Simpson’s rule)

)3

1(f )

3

1(f

**

Exercise: Derive the 6 conditions required to find the integration points and weights for a 3-point Gauss quadrature rule

Newton-Cotes Gauss quadrature

1. ‘M’ integration points are necessary to exactly integrate a polynomial of degree ‘M-1’ 2. More expensive

1. ‘M’ integration points are necessary to exactly integrate a polynomial of degree ‘2M-1’ 2. Less expensive3. Exponential convergence, error proportional to M

M

2

2

1

Example

1 3 2

1f( )d f( )I where

Exact integration

2

3I Integrate and check!

Newton-Cotes

To exactly integrate this I need a 4-point Newton-Cotes formula. Why?

Gauss To exactly integrate this I need a 2-point Gauss formula. Why?

1

1f( )d

1 1f(- ) f( )

3 32

3

I

Gauss quadrature:

Exact answer!

Comparison of Gauss quadrature and Newton-Cotes for the integral

1

1)2cos( dxxI

Newton-Cotes

Gauss quadrature

In FEM we ALWAYS use Gauss quadrature

Linear Element1 2

1 1

1 1 1T

1 1 1

1 1 1 11 1B EB d AEd AEd

1 1 1 14 4k A

Stiffness matrix

dbN1

1 Tbf dbN

1

1 iibf

Nodal load vector

Usually a 2-point Gauss integration is used. Note that if A, E and b are complex functions of x, they will not be accurately integrated

Quadratic Element

2

1 1T T

1 1B EB d E B B dk A A

Stiffness matrix

Nodal shape functions 1

1 10

3

1

22

3

( ) 12

( ) 1

( ) 12

N

N

N

You should be able to derive these!

Assuming E and A are constants

31 2 1 1B 2 1 2 2 1

2 2

dNdN dNdN

d d d d

1 1T T

1 1

2 2

12

1 22

B EB d E B B d

1/ 2 2 1/ 2 1/ 4

2 1/ 2 4 2 1/ 2

1/ 4 2 1/ 2 1/ 2

k A A

AE d

Need to exactly integrate quadratic terms.Hence we need a 2-point Gauss quadrature scheme..Why?