Lecture 3

The Best-fit Paradigm of FEA:

Shadows of the Exact

Somenath Mukherjee Gangan Prathap

Scientist, Director,
Structural Technologies Division, National Institute of Science and
National Aerospace Laboratories (NAL), Information Resources (NISCAIR),
Bangalore. New Delhi.

1
 Eminent Scholars of Mathematical and Computational Physics

Carl Friedrich Gauss Aldrien Marie Legendre Joseph Fourier

Pierre Simon Laplace Lord Rayleigh James Clerk Maxwell David Hilbert
2
 Lecture 3
The Best-fit Paradigm of FEA:
Shadows of the Exact
Chapters
1. The Algebra of Vector Spaces and Projections.

2. How the “Principle of Virtual Work” works to make

FEA the best-fit.

3. Real Examples: Simple one-dimensional elements.

4. Consequence of the best-fit nature in FEA.

Optimal points (Gauss’s and Prathap’s points) for exact
strain recovery.
3
“The external world of physics has thus become
a world of shadows. In removing our illusions we
have removed the substance, for indeed we have
seen that substance is one of the greatest of our
illusions…… Reality is a child which cannot
survive without its nurse illusion.”

- Sir Arthur Eddington

The Nature of the Physical World.

4
“The whole scientific inquiry starts from the familiar
world and in the end it must return to the familiar
world; but the part of the journey over which the
physicist has charge is in foreign territory.”
- Sir Arthur Eddington
The Nature of the Physical World.

“ The most practical tool is a good theory.”

- Albert Einstein

5
 Lecture 3

The Mathematics of Shadows…

Chapter 1

The Algebra of Vector Spaces and

Projections

6
 1.1 What are Vectors ?
• A vector is a physical quantity that has both
magnitude and direction.
A1
A = iA1 + jA2 + kA3 = [i j k ] A2 = [i j k ]{A}
A3

• A vector is an ordered array of numbers.

A1
A2
{A} = = [ A1 A2 .. .. An ]T
..
An
7
1.2 Linear Independence of Vectors
A set of N vectors are linearly independent if, for not all zero
values of α
i
N
α i Ai ≠ 0 (1.1)
i =1

A perfectly flat plane is two-dimensional.

Space, as we perceive, is three dimensional.

What is the logic ?

8
 Why do we say that a flat plane is two-dimensional ?
Only two linearly independent vectors can exist on the 2D plane.
B
Let A=5i+4j. Note that kA=k(5i+4j) and A are not
linearly independent because A- kA=0

A Let B=-4i+8j. A and B are two linearly independent

vectors on the plane, since αA B
C
kA Another vector C =13i-12j can be shown
as an resultant of a non-trivial linear
α1A+ α2B=C combination of A and B.
5 −4 13
α1 +α2 = i.e. (1).A + (-2).B=C
4 8 − 12
or or (1).A + (-2).B+(-1)C=0
5 − 4 α1 13
=
4 8 α2 − 12 Together, A, B and C are
linearly dependent.
Solving : α1 = 1, α 2 = −2 9
Any vector on this plane P can be expressed as a linear
combination of any two chosen linearly independent
vectors (for example A and B).
Plane P
k
(two dimensional vector space spanned
R3 by two linearly independent vectors A
A and B. C=α α1A+ α2B, A,B,C ∈ P )
B Important: P should pass through the
j origin O of the Euclidean 3-dimensional
0 D mother space R3, so that P carries the
null vector.
i
C Note: Vector D does not lie in this two-
P ⊂ R3 dimensional plane (vector space) P, but
it lies in the higher dimensional
( A, B, C) ∈ P ⊂ R 3
Euclidean mother space R3. 3
D ∉ P, D ∈ R3
P is a subspace of R3
P⊂R
The dimension of a vector space/subspace is the number of 10
linearly independent vectors needed to span it.
 1.3 Linear Vector Space
A linear vector space V is a set of vectors that satisfy the following
rules

(1) Any linear combination of vectors in the set V of vectors should

yield another vector that belongs to the same vector space.
i.e. A linear vector space is closed under linear combination.

If {u} and {v} are any two vectors in the vector space V, then any
linear combination of these should also lie in the V
If {u},{v}∈V then {w} = c {u}+ c {v} {w}∈V
1 2
where c1, c2 ∈ R. Here R is the set of real numbers.

(2) There exists a null vector {0} as a member of the vector space V,
satisfying the following identity for all {u} ∈ V.
{u}+{0}={u}
11
(3) The dimensions of a Vector Space V is the number of linearly
independent vectors in it. The linearly independent vectors in the space
V are called the basis vectors, spanning the space V.

(4) A Hilbert space is a one in which the norm (or magnitude) of any
vector in it is always positive and is defined by the inner product.
An inner product in the Hilbert space is given by
< a,b >= {a}T [D]{b} (1.2)
Here [D] is a symmetric positive definite matrix.
In Hilbert space the norm of any vector {a} is given by
a = < a, a > (1.3)

In the n-dimensional Euclidean space Rn, the norm of any vector

a (or{ a}) is given by the so-called dot (scalar) product
< a,b >= {a}T {b} = a ⋅b [D] = [I ]
a = < a, a > = {a}T {a} = a ⋅a
12
 Examples of vector spaces
Any “vector” in the real number line R can be a scalar multiple
of 1 as basis, and it belongs to R.
Thus the space R is a one dimensional vector space.
The cartesian product set is a vector space R2 which can be
expressed as
α1
R = R× R = {v}:{v} =
2 ∀ α ,α ∈R
α2 1 2

Basis vectors [1, 0]T and [0, 1]T span R2 . Dimension of R2 is therefore 2.

· Any vector {v} in R2 as a unique linear combination of these basis vectors as

α1 1 0
{v} = = α1 +α2
α2 0 1

Likewise, any vector {v} in the 3 dimensional vector space R = R × R × R

3

can be expressed as a unique linear combination of 3 basis vectors

spanning this vector space.
13
1.4 Projection of a vector along another vector
Scalar value of projection of a vector A
A along vector B
p AB = A • bˆ
A-pAB B
ϕ bˆ = is the unit vector along B
B
B
pAB
Vector projection of a vector A along vector B

p AB ( ˆ ˆ
= A•b b = A• )
B B
B B
=
A•B
B•B
B
(1.4)

Orthogonality conditions satisfied by the projection

p AB • (A − p AB ) = 0 (1.5)

2
A − p AB 2 = (A − p AB )
2 (1.6)
14
1.5 Gram-Schmidt algorithm to find orthogonal basis set
Given two linearly independent vectors B1 and B2. Find a set of
orthogonal vectors spanning the space in which B1 and B2 lies.
B2 B2 Orthogonality:
V2 • V1 = 0
V2 B2-p21
ϕ ϕ
B1 V1=B1 Any vector W in
p21 space spanned by B1
Let and B2
V1 = B1 W= α1B1+ α2B2
then or
B 2 • B1 W= 1V1+ 2V2
V2 = B 2 − p 2,1 = B 2 − B1
B1 • B1
V2 • V1 = 0
15
Basis vectors are never unique, and can be arbitrarily scaled.
 Gram-Schmidt algorithm to find orthogonal basis set
spanning an n-dimensional vector subspace.
Given n linearly independent basis vectors {b1}, {b2},….{bn}
spanning the n-dimensional subspace V.
Find a set of orthogonal vectors spanning this space.
Any vector w in the n-dimensional subspace V can be
expressed as
w= αi{bi}= i{vi} <vi,vj>=0 for i j

Gram-Schmidt Algorithm {v1} = {b1}

to determine n numbers of
< b ,v >
orthogonal basis vectors spanning {v2 } = {b2 } − 2 1 {v1}
the n-dimensional vector space V < v1, v1 >
in which the vectors k −1 < b , v >
k j
{b1}, {b2},….{bn} lie. {vk } = {bk } − {v j }
< v ,v >
j =1 j j
(1.7)
16
Basis vectors are never unique, and can be arbitrarily scaled.
Example 1. Find an orthogonal basis set spanning the vector subspace V
of the four-dimensional space R4, given that the following basis set B
spans subspace V.
B = Span(V ) = ({b1}, {b2 }, {b3}) where
T
{b1} = [2 1 2 1] ; T
{b2 } = [2 2 1 0] T
; {b3} = [1 2 1 0]

Using the Gram-Schmidt algorithm {v1} = {b1} = [2 1 2 1]T

< b2 , v1 > 4+2+2+0 1
{v2 } = {b2 } − {v1} = [2 2 1 0]T − {v1} = [2 6 − 3 − 4]T
< v1, v1 > 4 +1+ 4 +1 5
< b ,v >
{v3} = {b3} − 3 1 {v1} −
< b3 , v1 >
{v2 } Normalizing {v2 } = [2 6 − 3 − 4]T
< v1 , v1 > < v2 , v2 >
2+2+2+0
= [1 2 1 0]T − [2 1 2 1]T − 2 + 12 − 3 + 0 [2 6 − 3 − 4]T
4 +1+ 4 +1 4 + 36 + 9 + 16

{v3} = [− 7 5 4 1]T
1
= [− 7 5 4 1]T Normalizing
13

[{v1},{v2},{v3}] is the orthogonal basis set spanning the

three dimensional subspace V of the mother space R4, V ⊂ R4
<vi,vj>=0 for i j 17
 1.6 Orthogonal Projection of a vector in 3D space
onto a 2D subspace (a plane)

Consider a plane P spanned by two orthogonal vectors u and v

Projection Formula for a vector A as A onto the P subspace

A= A •u
u •u u + A• v v
v•v ; u• v =0 ; u, v∈ P (1.8)

A is the orthogonal projection of A onto space P

Properties of orthogonal projection
A
A − A •p = 0 ; p∈ P A−A
900
v
A−A •A =0 ; A∈ P
2 2 u
A
2 P
A−A = A − A Subspace
(1.9) 18
 1.7 Vector components as orthogonal projections
A = iA1 + jA2 + kA3 A ∈ R3
A
A is the Orthogonal Projection of
R3 vector A onto the plane (subspace) P
spanned by orthogonal basis vectors
k i and j.
0 A = iA1 + jA2 A∈P
A
j i Vector A can also be represented as a
P linear combination of its orthogonal
projections on the reference axes.

A = iA1 + jA2 + kA3 = (A • i)i + (A • j) j + (A • k )k

(1.10)19
 1.8 Orthogonal Projection of a vector in n-D space
onto a lower dimensional, m-D subspace (m<n)
Consider a m-dimensional subspace P spanned by m numbers of
orthogonal vectors {v1}, {v2}…{vm)
Projection Formula for a vector A existing in a higher n-dimensional
subspace as A onto the P subspace
m A•v
i
A= v v •v =0 for i≠ j
i =1 v i • v i
i i j A
A−A
{A}T {vi }
{A}=
m m < A, v >
{v } =
i
{v }
(1.11) 900
T i i v 1
i =1 {v } {v } i =1 < vi , vi >
i i
< v , v >= 0 for i≠ j v 2 P
A
i j
Subspace
Properties of orthogonal projection
< A − A, p >= 0 ; { p}∈ P A simple geometric representation for
< A − A , A >= 0 visualizing projection in abstract high
2 2 dimensional space.
A− A
2
A − A =
(1.12) 20
Example 2. Let V be a 2-dimensional subspace of 3-dmensional space
R3. The subspace V is spanned by the following given orthogonal basis
vectors T
{v1} = [1 2 1] , T
{v2 } = [1 − 1 1]
Find the orthogonal projection {A} of the vector { A} = [− 2 2 2]
T

onto the subspace V.

< A, v1 > < A, v 2 >
{ A} = {v1} + {v 2 }
< v1 , v1 > < v2 , v2 >
1 1
(−2)(1) + (2)(2) + (2)(1) (−2)(1) + (2)(−1) + (2)(1)
= 2 2 2
2 + 2 2 2
−1 A
(1) + (2) + (1) (1) + (−1) + (1)
1 1
A−A
4
1
(−2)
1 0 900
= 2 + − 1 = 2 = [0 2 0]T { v 1 }
6 3 { v 2 }
1 1 0
V
A
{A − A} = {A} − {A} = [− 2 2 2]T − [0 2 0]T = [− 2 0 2]T
Subspace

2
A− A = (−2) 2 + (0) 2 + (2) 2 = 8,

A
2
− A
2
[ ][ ]
= (−2) 2 + (2) 2 + (2) 2 − (0) 2 + (2) 2 + (0) 2 = 12 − 4 = 8 21
 1.9 Best-fits as Orthogonal Projections
We are given a scatter diagram of points (x1,y1), (x2,y2).. (xn,yn)
etc., and asked to fit a best-fit straight line
y = a + bx y
Using the points (x1,y1), (x2,y2).. (xn,yn)
1 x1 y1 .. y = a + bx
1 x2
a
y2
[ B]{α } = { y} ..
1 xn
b
=

yn
(1.13) ..
x
This is an inconsistent system of equations, since we have more
Equations than the number of unknowns (a,b)
Pre − multiply [ B]{α } = { y} by [ B]T [ B]T [ B]{α } = [ B]T { y}
This is the Normal Equation for best-fit (1.14)
Solving the Normal Equation (1.9) we can get the best-fit
straight line y = a + bx 22
 a
[ B]T ({ y} − [ B]{α }) = 0
y = a + bx = [1 x ]

{y} {{ y} − { y}}= 0
b
T
y = [ B ( x)][α ] (1.15)

Error { y} − { y} is orthogonal to the best fit {y}

{y} { y} − { y}

{ y} − { y*}
900

{y*}
{y}
P Subspace

23
Example 3. Find the best-fit straight line to the following data

x -1 1 2 4 12
y
.
y 2 4 6 12
.
Let y = a + bx = [1 x ]
a
6
. .
b
x
y = [ B ( x)][α ] 0 2 4
1 −1 2
1 1 a 4
=
1 2 b 6
1 4 12

[ B ]{α } = { y}

We have to find the linear best-fit solution to this inconsistent

system of equations 24
 y
x -1 1 2 4
12 .
y 2 4 6 12
.
Method 1: Using Normal equation
.
6
.
T T x
[ B] [ B]{α } = [ B] { y} 0 2 4

4 6 a 24
= {y}
6 22 b 62 { y} − { y}
Solving : a = 3, b = 2 900

Thus the best-fit straight line is {y}

B Subspace

y = 3 + 2x
25
 y
x -1 1 2 4
12 .
y 2 4 6 12
.
Method 2: Using projection formula
.
6
. x
1 −1
1 1 Two-linearly independent 0 2 4
[B] = vectors that lie in (and span)
1 2
the space B are
1 4
{b1} = [1 1 1 1]T , {b2 } = [− 1 1 2 4]T
Using the Gram-Schmidt algorithm, the orthogonal basis vectors
spanning B subspace are to be established

{v1} = {b1} = [1 1 1 1]T

< b2 , v1 > 6 1
{v2 } = {b2 } − {v1} = [− 1 1 2 4]T − {v1} = [− 5 − 1 1 5]T
< v1, v1 > 4 2

Normalizing {v2 } = [− 5 − 1 1 5]T < v , v >= 0

1 2 26
 y
x -1 1 2 4
12 .
y 2 4 6 12
.
Using projection formula, we find the best fit
of the vector {y}=[2,4,6,8]T .
6
. x
4
{y} T
{y}=
m {v } m < y, v >
i i
{v } = {v }
T i < v , v > i {y}
i =1 {v } {v } i =1 i i { y} − { y}
i i
1 −5 1 900

{}
y =
24 1 52 − 1
+
4 1 52 1
=
7
5
B Subspace
{y}

1 5 11
Orthogonal projection of the vector {y} onto the
y = 3 + 2x subspace B 27
1.10 The Polynomial Vector (Function) Space
Consider a Vector Function Space of r-rowed vectors which
are polynomial functions of degree (n-1) of some independent
variable ξ, bounded by –1 and +1.
Definition:
n
Pnr (ξ ) = { p} : { p} = {a}i ξ i −1 , − 1 ≤ ξ ≤ 1, {a}i ∈ R r
i =1
(1.16)

Dimension of this space:

dim[ Pnr (ξ )] = r × n (1.17)

A Polynomial Function {p( )} of degree (n-1) is a one-row vector

(r=1) which is an element of the space
{ n
Pn (ξ ) = Pnr =1 (ξ ) = { p} : { p} = {a}i ξ i −1 ,
i =1
− 1 ≤ ξ ≤ 1, {a}i ∈ R }
dim[ Pnr =1 (ξ )] = n
28
 1.11 The Legendre Polynomials
r =1
• The n-dimensional function space Pn (ξ ) must be spanned by n-
linearly independent vectors.

The n- numbers orthogonal basis vectors spanning Pnr =1 (ξ )

are known as the Legendre Polynomials.

Example 4. A polynomial of degree 3 p4 = a1 + a2ξ + a3ξ 2 + a4ξ 3

r =1
belongs to the function space 4 (ξ )
P
The basis vectors spanning P4r =1 (ξ )
b1 = 1, b2 = ξ , b3 = ξ 2 , b4 = ξ 3

Inner product definition:

1
< a, b >= {a}T {b}.dξ a, b ∈ P41 (ξ ) (1.18)
−1 29
 Using the Gram-Schmidt algorithm, the orthogonal basis
vectors spanning subspace P4r =1 (ξ ) are to be established

P1 = b1 = 1
1
ξ .1.dξ
< b2 , P1 > < ξ ,1 > P2 = ξ
P2 = b2 − {P1} = ξ − 1 = ξ − −1 1 1=ξ
< P1 , P1 > < 1,1 >
dξ
−1
<b ,P > <b ,P >
P3 = b3 − 3 1 {P1} − 3 2 {P2 } P3 = 3ξ 2 − 1
< P1, P1 > < P2 , P2 >

< b4 , P1 > < b ,P > <b ,P > P4 = 5ξ 3 − 3ξ

P4 = b4 − {P1} − 4 2 {P2 } − 4 3 {P3}
< P1 , P1 > < P2 , P2 > < P3 , P3 >

span[ P41 (ξ )] = [ P1, P2 , P3 , P4 ]

1
(1.19)
< Pi , Pj >= Pi Pj dξ = 0, i≠ j
30
−1
 1.12 Orthogonal Projection (Best-fits) of
Polynomial Functions

If we seek a least square error solution of an

original polynomial function {y} then we actually
{}
look for a solution y such that
{y}
{ y} − { y}

< y, y − y >= 0 (1.20)

900
{ y} − { y*}

(NORMAL EQUATION) {y*}

{y}
P Subspace

The minimum squared error is obtained when the

{ }
error { y} − { y} is orthogonal to the approximate
{}
(best-fit) solution y lying on the approximation
31
subspace.
Example 5. Best fit the 2 degree polynomial
y = p3 = 1 + 2ξ + ξ 2
by a straight line.

Solution:
The straight line best fit y to the quadratic y = p3 is the
orthogonal projection of the quadratic function
onto the polynomial space P2r =1 (ξ )

Using Projection 1 1
y.1.dξ y.ξ .dξ
Formula 2
< y, Pi > 4
−1
y= Pi = 1
+ −11 ξ= + 2ξ
i =1 < Pi , Pi > 3
dξ ξ 2 dξ
−1 −1

32
Given Quadratic Curve is
2 4 1
y = p = 1 + 2ξ + ξ = P + 2 P + P −1 ≤ ξ ≤ 1
3 3 1 2 3 3
4 4
Linear Best-fit to y is y = + 2ξ = + 2 P2
3 3

33
 An indication of the best-fit rule in FEA …

CANTILEVER BEAM ANALYSIS USING A SINGLE EULER BEAM

ELEMENT OF LENGTH L
Uniformly distributed loading is q per unit length.
q is such that fixed end curvature (bending strain) is qL2/2EI=4 (m-1)
q

34
1.13 Fourier Series as orthogonal projection of any
periodic function
A periodic function f(t) of period 2 : f (t + 2π ) = f (t )
Fourier series

35
 ∞ ∞ ∞ ∞
< f ( x),1 > < f ( x), cos nx > < f ( x), sin nx >
f ( x) = a0 + an cos nx + bn sin nx = 1+ cos nx + sin nx
n =1 n =1 < 1 ,1 > n =1 < cos nx , cos nx > n =1 < sin nx , sin nx >

π
< a, b >= a.b.dx
−π
π
2
a =< a, a >= a.a.dx
−π
π π π
f ( x)dx f ( x) cos nxdx f ( x) sin nxdx
< f ( x),1 > −π < f ( x), cos nx > −π < f ( x), sin nx > −π
a0 = = , an = = , bn = =
< 1,1 > 2π < cos nx, cos nx > π < sin nx, sin nx > π

In practice, we use a finite number (N) of terms;

N N N N
< f ( x),1 > < f ( x), cos nx > < f ( x), sin nx >
f (x) = a0 + an cos nx + bn sin nx = 1+ cos nx + sin nx
n =1 n =1 < 1 ,1 > n =1 < cos nx , cos nx > n =1 < sin nx , sin nx >

Thus f (x) is the orthogonal projection of f(x) onto a subspace

of 2N+1 dimensions, spanned by orthogonal basis vectors
{1, sin(nx), cos(nx): n=1,2,…N, - x }
2 2 2
f − f = f−f
36
Using Fourier series we can write an ideal square wave as an infinite series of
the form

37
 Lecture 3
FEA creates shadows…
Chapter 2

How the “Principle of Virtual Work”

works to make FEA the best-fit

38
 Archimedes was the first person to use the concept of Virtual Work
in his calculations for the Lever.

"Give me a place to
stand on, and I will
move the Earth."

Equilibrium: Virtual Work Principle

F δ −F δ =0
11 2 2
δ δ
Geometric Compatibility 1= 2
L L
1 2
δ2
L1 Together
δ1 L2
F1 L1= F2 L2
F2
F1 39
 2.1 Application of the Principle of Virtual Work
2.1.1 The weak form at element level with exact solution u in the
differential equation
• A linear differential equation of a conservative system (with as a self-adjoint
operator),
(u) = f (2.1)
• Weak form at the element level is obtained from applying the virtual work principle

u h ( u − f ) dx = 0 (2.2)
e

This leads to the weak form as [

a (u h , u ) e = (u h , f ) + u h , R e ]
B (2.3)

a(u,uh)e is the symmetric, bilinear functional a(u,uh)e=a(uh,u)e (2.4)

(u h , f ) = u h. f dx (2.5)
e

u h , Re = {u h }B {Re } = Virtual Work done by the ANALYTICAL nodal

T
B
reaction vector {Re} from connectivity at the element boundary

40
2.1.2 The weak form at element level with approximate solution for the
differential operator (FEA form)
u h ( u h − f )dx = a(u h ,u h )e − (u h , f ) − u h ,Q he ≠ 0 (2.6)
e B
Here {Q } = FE computed Nodal Stress Resultant using u h
he

However, we may replace the approximate nodal stress resultant vector

appropriate nodal reaction vector for equilibrium
a(u h ,u h )e = (u h , f ) + u h , Rhe (2.7)
B
R he ≠ Q he
(2.8)
Computed nodal reaction vector NOT EQUAL TO nodal stress resultant vector
With FE approximation, each element is only in external
equilibrium, but not in internal equilibrium.
The FEA computed Nodal Reaction Vector for the element is

{Rhe }= [K e ]{δ e }− {F e } (2.9)

41
 2.2 FEA Error Statement

Equation (2.3): [
a(u h , u, )e = (u h , f ) + u h , R e ]B

Equation (2.7): [
a(u h , u h ) e = (u h , f ) + u h , R he ]B

(Actual FEA)

Subtracting equation (2.3) from (2.7)

[
a(u h , u h ) e − a (u h , u ) e = u h , R he − R e ]B
(2.10)
= {δ } {R
e T he
− Re }
This equation governs all FEM errors in a very general sense

42
2.3 The Bilinear Symmetric Form and the inner product

• Element Strain Vector by FEA: {ε } = [B]{δ }

h e
(2.11)

• Analytical element strain vector : {ε }

• Bilinear form and Inner product definition :

a(u h , u ) e = {ε } [D]{ε }dx =< ε

h T h
,ε > (2.12)
e

a (u h , u h ) e =
e
{ε } [D]{ε }dx =< ε
h T h h
, ε h >= ε h
2
(2.13)

Here [D] is the element rigidity matrix

43
 2.4 The Best-Fit Paradigm in FEA

CASE A: Agreement of Nodal Reactions in Elements

• In all determinate problems, it has been observed that the analytical nodal
reactions are exactly reproduced by finite element computations, i.e.
{R he } = {R e } (2.14)
Then from equation (2.10), we have
a (u h , u h )e = a (u , u h )e (2.15)
We get NORMAL EQUATION (for Orthogonal Projection)
< ε h , ε h >=< ε h , ε > (2.16)
or < ε h , ε − ε h >= 0 (2.17)

[B]T [D][B]{δ e }dx = [B]T [D]{ε }dx (2.18)

e e
Conclusion: Finite Element Strain {ε h } is actually an Orthogonal Projection
(best-fit) of the analytical strain {ε } onto the B Subspace
Using the Gram-Schmidt process, then orthogonal basis vectors spanning the
m-dimensional space B can be found out. Hence by Projection Formula
< ε ,ν i >
{ε } = {ε } =
h m
{ν i }, < ν i ,ν j >= 0 for i≠ j (2.19)
i =1 < ν i ,ν i > 44
FEA strain vector is the orthogonal projection of the analytical strain
vector, onto the B Subspace.

The Error-Energy Rule (Pythagoras Theorem)

2 2
ε −ε h = ε − ε h
2
(2.20)
i.e.
The Energy of the Error= Error of the Energies

45
 2.4 The Best-Fit Paradigm in FEA (..Continued)

CASE B: When FEM Computed Nodal Reactions in Elements do not agree

with the Analytical ones.

{R he } ≠ {R e }
• This leads to a prima facie violation of the best-fit rule:

a (u h , u h )e − a (u , u h ) e = [u h , R he − R e ]B
= {δ e } {R he − R e }
T

≠0
• This means that the discretisation and approximation process of FEA
have induced a stiffening force in the system from the error of the reaction
{R he
− Re }

• However, note that if we consider the modified analytical solution for the
stiffened system from the extraneous force from the reaction error, then
the FEA results are actually the best-fit strains of the stiffened analytical
solution.

46
2.5 How the B Subspace (for strain projection) emerges
The strain displacement matrix [B] emerges from the FE formulation :
{ε h } = [B]{δ e }
The B subspace is the vector function space in which all the column
vectors of the [B] matrix lies. But these vectors in [B] need not be all
linearly independent.
The B space is a subspace of the general Polynomial Space of r-rows.
B ⊂ Pnr ; { n
Pnr = {p} : {p} = {ai }ξi −1 ,−1 ≤ ξ ≤ 1 , { ai } ∈ R r
i =1
}
m = dim( B ) < dim Pnr = r × n
Using Gram-Schmidt process, one can generate m-numbers of the non-zero
orthogonal basis vectors spanning the m-dimensional B subspace.

m=dim(B)=Rank of the element stiffness matrix

If the element does not suffer from rank deficiency, then the dimension m of the
B space is given by
dim(B)=N-R
N=Total number of degrees of freedom of the element
R=Total number of physical rigid body motions of the element 47
 Lecture 3
The proof of the pudding is in the eating…

Chapter 3

Real Examples:
Simple one-dimensional elements

48
 3.1 The Simple Bar Element

Le 2
1
u1,F1 u2,F2
Differential Equation: R1 R2
d du
EA=Elastic rigidity
− EA −q =0
dx dx (3.1)

Weak form (Exact solution for the differential equation):

d du
− EA − q .u h .dx = 0
e dx dx (3.2)
Integration by parts,

du h du du du (3.3)
EA dx − u h qdx − u1 − EA + u2 EA =0
e dx dx e dx dx
x = x1 x = x2

49
• The Weak form (not used in FEA) is
du h du du du
EA dx = u h qdx + u1 − EA + u2 EA
e dx dx e dx dx
x = x1 x = x2
(3.4)
Here we have the general form
u1,F1 Le 2
h e h e
a (u , u ) = (q, u ) + [u1R1e + u2 R2e ] R1 1
(3.5)
u2,F2
The Analytical Reactions at the nodes match the internal stress
resultants
R2
du du
R1e = − EA R2e = EA
dx x = x1 dx x= x2 (3.6)

Equation (3.6) is satisfied only if the exact function u is used in the

differential equation.

This means that with the exact function u in the differential

equation, the element is in both external and internal equilibrium.
50
• Weak form for the FEA: Le 2
1
d du h
u1,F1 u2,F2
− EA − q .u h .dx R1
e dx dx R2
du h du h (3.7)
= a (u , u ) − ( q , u ) − u1 − EA
h h e h e
+ u2 EA
dx x = x1
dx x = x2

≠0 !!
But with proper nodal reactions for only external equilibrium of the
element (using this approximation),
a (u h , u h )e − (q, u h )e − [u1R1h ,e + u2 R2h ,e ] = 0
or (3.8)
a (u , u ) = (q, u ) + [u R
h h e h e
1 1
h ,e
+ u2 R ]h ,e
2

Penalty for using approximate function is that the approximate internal stress
resultants do not agree with the corresponding nodal reactions
du h du h
R 1
h ,e
≠ − EA and Rh ,e
2 ≠ EA (3.9)
dx x = x1
dx x = x2

Element is in external equilibrium; it is NOT in internal

equilibrium (!)
51
 3.2 The best-fit rule in simple bar element

The following relationship holds good in general for any element

a (u h , u h )e − a (u , u h )e = [u h , R he − R e ]B
= {δ } {R
e T he
− Re }
For the bar element:

du h du h du h du
dx = {δ e } {R he − R e }
T
EA .dx − EA . (3.10)
e dx dx e dx dx
Here the Analytical Reaction Vector components are
du du
R1e = − EA R2e = EA
dx x = x1 dx x = x2

The FEA computed Nodal Reaction Vector for an element is

{R } = [K ]{δ }− {F }
he e e e

52
 3.2.1 Case A: When the Analytical Reactions are
conserved by FEA
• In this case {R } = {R }
he e

• Hence we have the following as direct consequences,

a (u h , u h )e = a (u , u h )e < ε h , ε − ε h >= 0

e
[B ]T [D][B ]{δ e }dx = e [B ]T [D]{ε }dx
• For the linear bar element (with linear Lagrangian shape functions)
1 1 u1 1 1
ε = [ B]{δ } = −
h e
[ B] = −
L L u2 L L
N = Degrees of freedom = 2, R = Rigid body motion = 1
dim( B) = 1, B spanned by ν = 1 ( Applying Gram − Schmidt )

• FEA gives Best-fit strain {ε } = {ε } = < ε ,1 > 1

h

< 1,1 >

53
 EXAMPLE 1. Cantilever bar analysis with a single linear bar element.
(a) Bar with constant section (b) Bar with varying
area A subjected to sectional area ( A1=Ao,
uniformly distribute axial A2=3Ao)subjected to axial
load q0. tip load P.

Rigidity matrix [D] EA EA0 (2 + ξ )

Orthogonal basis vector {ν }1 = 1 {ν }1 = 1
spanning 1-D B subspace
Analytical strain vector q0 L(1 + ξ ) P
{ε } 2 EA EA0 (2 + ξ )

{ε } = {ε } = < ε ,ν
h 1>
{ν 1 } q0 L P
< ν 1 ,ν 1 > 2 EA 2 EA0

2 2 2
2
q0 L
3
P 2 L(ln 3 − 1)
ε −ε h = ε − ε h 2 EA
12EA
54
EXAMPLE 1. Cantilever bar analysis with a single linear bar element.

(b) P
(a)
Ao 2Ao 3Ao

A = A (2 + ξ )
O
L
qo
L
u2 = 0 A2 2 ξ=1
ξ =-1 A1 1
1 2
u2=0
ξ = −1 ξ = 1 ξ=0

∗
ξ = 0 ∗ P
q L ° ° ∗ ° P
qo L ° ∗ ° ° o
EAo ∗
∗
2EA EA 3 EAo
0 0
P /( 2 EAo )
AXIAL STRAIN AXIAL STRAIN

EAε
h
/P [1 ( )]
+ ξ 2

Analytical strain °
∗ °
FEM Strain 1 °
° 0
ξ=0 ξ=1
, ξ =-1

SPURIOUS OSCILLATIONS OF FORCE

RESULTANT
55
 EXAMPLE 2. Cantilever bar analysis with a two linear bar
elements.
Results of the element 1-2
P
Ao 2Ao 3Ao
Rigidity matrix [D] EA0
(3 + ξ )
2
L L
Orthogonal basis
vector spanning 1-D {ν }1 = 1
B subspace
A2 2 A3 3
A1 1
u3=0
Analytical strain
P ∗ ° 2P
°
∗
vector { ε }
EAo ° ∗ ° 2P EA0 (3 + ξ )
5 EAo
0 P
2P
3 EAo 3EAo {ε } = {ε } = < ε ,ν
h 1 >
{ν 1 }
2P
< ν 1 ,ν 1 > 3EA0
Cantilever bar analysis for the linearly
tapering bar with two linear bar
elements.
h 2 2 h 2 P2L 2
ε −ε = ε −ε ln 2 −
EA0 3

56
3.2.2 Case B: When the Analytical Reactions are NOT
conserved by FEA

{R he } ≠ {R e }

a (u h , u h )e − a (u , u h ) e = [u h , R he − R e ]B
= {δ e } {R he − R e }
T

≠0

•When the FE computed reactions {Rh,e} disagree with

the analytical reactions {Re}, the approximate (FEA)
element strains are best fits to the strains from stiffened
analytical solutions u* with error of the nodal reactions.
(With analytical reactions matching the FE reactions).
a (u h , u h )e − a (u*, u h )e = 0

57
 EXAMPLE 3.
Fixed-fixed tapered bar analysis with two linear bar elements.
Case : Uniform loading q = 1

E = 1, l = 1, A2 = 1 and A1 = 0.01, u(x=0)=u(x=l)=0

The exact (analytical) solutions for this case :

ε

Displacement : u = -1.010101x + 0.219341 ln (1 + 99x)

Strain: = -1.010101 + 21.714759 / (1 + 99x)

Stress resultant: Q = 0.2070 – x

A1

A2

58
EXAMPLE 3. Fixed-fixed tapered bar analysis with two linear bar
elements.
Case: Uniform loading q = 1
Element e 1 2

FEM Strain {ε he }
0.4950 -0.4950

Best-fit e -0.1671 -0.7216

Strain {ε }

e 0.6621 0.2266
{ε he } − {ε }
FEM
Reaction -0.3775 0.1225
{Rhe} -0.1225 -0.6225

Analytical
Reaction -0.2070 0.2930
{Re} -0.2930 -0.7930

Reaction Error
{Rhe}-{Re} -0.1705 -0.1705
0.1705 0.1705

59
EXAMPLE 3. Fixed-fixed tapered bar analysis with two linear bar
elements.
Case: Uniform loading q = 1

Displacements in a tapered bar

0.50

U-Exact
0.45
U-FE-N2=0.5
0.40
Series3
Local Error in
0.35 1st element Series4
E=1
Global Error in A1=0.01
Displacement

0.30
2nd element A2=1
0.25
q= 1
Local Error in 2nd l=1
0.20 element

0.15
Global Error in
1st element
0.10

0.05

0.00
0 0.2 0.4 0.6 0.8 1 1.2

Distance along length

60
 EXAMPLE 3. Fixed-fixed tapered bar analysis with two linear bar
elements.
Case : Uniform loading q = 1
Stress resultants in a tapered bar

0.60

Exact-Q
0.40 Q-FE-N2=0.5
Q'
0.20

E=1
0.00 A1=0.01
0 0.2 0.4 0.6 0.8 1 1.2
A2=1
-0.20 q= 1
Q

l=1

-0.40

-0.60

-0.80

-1.00

Distance along length

The FEA solution suffers an additional spurious stiffening from a tension

of 0.1705 (force units) which originates from the error in the nodal
reaction vectors. FEA strains are best fits to the stiffened analytical strains.
61
Modified problem to EXAMPLE 3. Free-fixed tapered bar analysis with
two linear bar elements.
Case : Uniform loading q = 1 and point load P=0.3775 at the free end
same as the FEA Reaction for the original fixed end of example 3.
0.3775

X
The FEA strain of
this problem exactly
A1
agrees with its own
Element e 1 2 A2 best-fit (Reactions agree).
FEM Strain {ε he } 0.4950 -0.4950
Best-fit e FEA strain of Example 3
{ε } 0.4950 -0.4950
Strain (FIXED-FIXED) responds
e
{ε he } − {ε } 0
as the best-fit strain of
0
this modified problem,
FEM -0.3775 0.1225
Reaction -0.1225 -0.6225
(with additional stiffening of
{Rhe} 0.1705 units of force from
Analytical -0.3775 0.1225
Reaction -0.1225 -0.6225
reaction error that stiffens
{Re} the original system of
Reaction Error 0 0
{Rhe}-{Re} 0 0
Example 3).
62
 0.3775
X
X

A
1 A1

Case 1. Fixed-fixed indeterminateA2 Case 2. Free-fixed bar with u.d.l. q=1, and
A2
bar with u.d.l. q=1 Load P=0.3775 ( ) at the free left end

63
 Example 4: The spherically symmetric Laplace Equation

The Laplace Equation: ∇ 2u = 0

1 d 2 du
The spherically symmetric Laplace Equation: 2
r =0
r dr dr
Analytical Solution for Potential u (disp) and 1 du 1
u (r ) = ε = = − 2
Potential Gradient (strain) : r dr r
1 1
Satisfies Boundary Conditions (DD): u (r1 ) = u1 = ; u ( rn +1 ) = un +1 =
r1 rn +1
Boundary Conditions (DD) are applied to the ends of the continuum
discretised with n elements:
The Weak form in an element ‘i’: a (u , u h ) = ( Riui + Ri +1ui +1 )
du du h
The Bilinear form: a (u , u ) =
h ri +1
ri r2
dr
dr dr
du du
Analytical Reactions (Flux): Ri = − r 2 Ri +1 = r 2
dr r = ri dr r = ri +1
64
 ELECTRIC FIELD FROM A POINT CHARGE Q
An Example of Application of Spherically Symmetric
Laplace Equation
1 d 2 du
r =0
r 2 dr dr
Q
Potential : u=
4πε 0 r
du Q
Intensity: E=− =
dr 4πε 0 r 2
Q
Q
Here = k2 = 1
4πε 0
r3

r1 The Continuum is
discretised into 2 Linear
Finite Elements.
First
Element(1-2) r2 Second Element(2-3)
65
 FE Solution of Symmetric Laplace Equation (using 2 linear elements)
Original Problem with (DD) Modified problem with (DN)
FE Flux=2.1645 FE Flux=-2.1645 Applied End Force F3 = −2.1645

Anal Flux=1 Anal Flux= -1 Flux=2.1645

(Incoming at 1) (Outgoing at 3) Flux= -2.1645
(Incoming at 1) (Outgoing at 3)
r1 = 0.1, r2 = 0.2, r3 = 10
u1 = 1 / r1 = 10, F3 = −2.1645
u1 = 1 / r1 = 10, u3 = 1 / r3 = 0.1
Element 1 2 Element 1 2
e e

FEM -92.764 -0.0636 FEM -92.764 -0.0636

{ε he }
Strain Strain { ε he *}

Best-fit -42.8571 -0.0294 Best-fit -92.764 -0.0636

Strain e Strain e

{ε } { ε *}
e e
{ε he } − {ε } -49.9065 -0.0342 { ε he *} − { ε *} 0.0 0.0

FEM 2.1645 2.1645 FEM 2.1645 2.1645

Reaction -2.1645 -2.1645 Reaction -2.1645 -2.1645
{Rhe} {Rhe}
Analytical Analytical
Reaction 1.0 1.0 Reaction 2.1645 2.1645
{Re} -1.0 -1.0 {Re} -2.1645 -2.1645

Reaction Reaction Error

Error 1.1645 1.1645 {Rhe}-{Re} 0.0 0.0
{Rhe}-{Re} -1.1645 -1.1645 0.0 0.0 66
FE Solution of Symmetric Laplace Equation (using 2 linear
elements)

67
 Bilinear symmetric forms and the corresponding strains for different elements
T
du h du
a(u h , u ) e = EA dx h
Bar element e
dx dx
{ε }h
=
du
dx
a(u h , u ) e = {ε } EA{ε }dx =< ε
h T h
,ε >
e
2 h T
d w d 2w
a ( wh , w) e = EI dx
dx 2 dx 2 h
Euler beam
{ε } d 2w
e
h
= 2
element h
a ( w , w) = e
{ε } EI {ε }dx =< ε
h T h
,ε > dx
e

Timoshenko
beam
h
a (u , u ) =e
{ε }
h T EI
0
0
kGA
{ε }dx {ε }
h
=
dθ h / dx
element e ( θ h − dwh / dx
=< ε h , ε >
Euler beam
element
elastic
on a ( wh , w) e = {ε } h T EI
0
0
k
{ε }dx
{ε } h
d 2w
= dx 2
h

foundation with e
wh
stiffness k per =< ε h , ε >
unit length
68
 3.3 The Simple Euler Beam Element (Cubic)
The differential equation for the Euler beam with section rigidity EI and distributed
loading q(x) is given by
d2 d 2w (3.11)
EI − q = 0
dx 2 dx 2
Virtual work principle leads to the Galerkin Equation in the element ‘e’.
{ }{ }
T
a( wh , w) e = ( f , wh ) e + δ e R e (3.12) w1,F1 w2,F2
h
d 2 wh d 2w dw
2
EI dx = w h
qdx + w1
h e
R1 ( x = x ) + R2
e
( x = x1 ) Le
dx 2
1
e dx e dx x = x1

h
h e dw e
+ w2 R3 ( x = x 2 ) + R4 ( x = x2 )

1,M1 2,M2
dx x = x2

The Analytical Reactions are:

d d 2w d 2w
R =
1
e
EI 2 R = − EI 2
e
2
dx dx x = x1
dx x = x1

2 2
(3.13)
d d w d w
R3e = − EI 2 R4e = EI
dx dx x = x2
dx 2 x = x2
69
The weak form for FEA a ( wh , wh ) e = (u h , f ) e + δ e R he { }{ }
h
d 2 wh d 2 wh h he dw he
2
EI dx = w h
qdx + w1 R1 ( x = x1 ) + R2 ( x = x1 )
e dx dx 2 e dx x = x1

h (3.14)
h he dw he
+ w2 R3 ( x = x2 ) + R4 ( x = x2 )
dx x = x2

The FEA computed nodal reactions can be expressed as

R4 ]T = [ K e ]{δ e } − {F e }
e e e e
{R h ,e } = [ R1 R2 R3 (3.15)

where
[ K e ] = [ B ]T EI [ B ]dx {F e } = [ N ]T q ( x)dx
e e

70
 3.3.1 The B Subspace and Strain Projection for the
Simple Euler Beam Element (Cubic)
Element formulation with Hermite Cubics
T
dwh dwh
Disp : wh = [N1 N2 N3 N4 ] {δ }, {δ } =
e e
w1 w2
dx 1
dx 2

Strain : ε h = d 2 wh / dx 2 = [B ]{δ e }

Le
D = EI , Inner Pr oduct < a, b >= {a} EI {b}. dξ
1 T
Element Rigidity
−1 2
The dimension of the subspace B is also the rank of the stiffness matrix,
m=dim (B)=N-R (m=4-2=2 for the Euler beam element).

< ε ,vj >

Projection rule: {ε } = {ε } =
h m

j =1 < v , v >
{v j }, < vi , v j >= 0 for i ≠ j
j j

Here ε is the analytical strain (a stiffened one from nodal reaction error, if any)

71
 Examples: Cantilever beam analysis with single Euler element
(a) (b

qo
qo

I = I o (2 + ξ ) Io 2Io 3Io
L w2 = 0
θ2 = 0
1 2
L w2 = 0
θ2 = 0
ξ = -1 ξ=0 ξ= 1

∗
2 2
5q o L 1
° I1
I2
∗ 12 EI
ξ = -1 ξ =1
ξ=0
° qo L2
0 ∗ 2 EI

qo L
2
° °
12 EI BENDING STRAIN (CURVATURE)
° 7 qO L2 q O L2
∗ 44 EI o 6 EI o
°
Analytical strain
0 ∗
∗
Strain by projection (FEM) qO L2
°
44EI o
°
,
BENDING STRAIN (CURVATURE)

72
 Examples: Cantilever beam analysis
(..Continued)
(b) Linearly varying
(a) Constant
sectional moment of inertia.
sectional moment of
( I1=Io, I2=3Io)
inertia.

Rigidity matrix EI ( 2 + ξ )
EI o
[D(ξ )]

Analytical strain 2 2
vector
2
qo L ( 1 +
2
) qo L ( 1 + )
{ε } 8EI 8EI ( 2 + )
o
Orthogonal basis
vectors,
{u1},= {1 } { u } = { 1,}
1

{u2 }= { } { u } = {( 6
2
− 1) }
FEM strain = strain 2
by projection
2
qo L ( 2 + 3ξ ) qo L (3 + 4 )
{ε }= {ε}=
h
m=2 < ε,u j >
< uj ,uj >
{ uj} 12EI 44EI
o

{ }
j =1
2 5
The energy error rule 2 5 qo L 12
2
qo L ln(3) −
- h
=
2
− h
2
= 128EI 11
720EI o 73
CANTILEVER BEAM ANALYSIS USING TWO EULER BEAM ELEMENTS
Uniformly distributed loading is q per unit length.
q is such that fixed end curvature (bending strain) is qL2/2EI=4 (m-1)
q

Le=L/2 Le=L/2

q ( Le ) 2 Bending strain
12 EI Analytical (Curvature)
(quadratic)
FEM q ( Le ) 2
(linear) 12 EI
4

Shear Force
qL
74
 Lecture 3

Chapter 4

Consequence of the best-fit nature in FEA.

Optimal points (Gauss’s and Prathap’s points)
for exact strain recovery

75
 4.1 The Gauss points as the optimal points
(For statically determinate problems)
Review of cases where the analytical strains that are
polynomials just one order higher than the polynomial
representing the approximate strains.
N +1
Exact (analytical) ε= ai Pi (ξ ) −1 ≤ ξ ≤ 1
i =1
N
Approximate as best-fits εh = ai Pi (ξ )
i =1
At the optimal points, the exact strains are captured.
εh =ε
Thus the optimal points are determined from the roots of
the highest Legendre Polynomial PN+1( ). These are called
the Gauss points.
PN +1 (ξ ) = 0
ξ 0 = ξGauss 76
 Example 1. Analysis of a bar under u.d.l. q using FEM
Fixed end; L
u1=0
q
Le=L/3 Le Le Discretization
R 1 = −q(3Le ) 1 2 3 4
= −qL
e
F1( applied ) = qL / 2 F2 ( applied ) = qLe F3( applied ) = qLe F e
4 ( applied ) = qL / 2

THE STRAIN VARIATION DIAGRAM

Analytical h
Strain ε FEA Strain ε
(linear) (constant)
qL/(EA) Optimal points
P2=0 qLe/(2EA)
0=0

FEA gives the best-fit to the analytical strain. Why ? 77

 Example 2. Cantilever beam analysis using a single euler beam
element of length L
Uniformly distributed loading is q per unit length.
q is such that fixed end curvature (bending strain) is qL2/2EI=4 (m-1)

Optimal points 0 = ±1/ 3 q

P3=3 2-1=0
= ±1/ 3

78
Example 3. Cantilever beam analysis using two euler beam elements
Uniformly distributed loading is q per unit length.
q is such that fixed end curvature (bending strain) is qL2/2EI=4 (m-1)
q

Le=L/2 Le=L/2

q ( Le ) 2 Bending strain
12 EI Analytical (Curvature)
(quadratic)
FEM q ( Le ) 2
P3=3 2-1=0 (linear) 12 EI
0 = ±1/ 3
4

Shear Force
qL
79
 4.2 Prathap points as general optimal points
(for statically determinate problems)
For cases where the polynomial of the analytical strain
is more than one order higher than the polynomial for
the approximate strain. N +r
Exact (analytical) ε= ai Pi (ξ ) −1 ≤ ξ ≤ 1
i =1
N
Approximate as best-fits εh = ai Pi (ξ )
i =1
At the optimal points, the exact strains are captured.
εh =ε
Thus the optimal points are determined from the roots of
the following expressions. These are the Prathap points.
a N +1PN +1 (ξ ) + a N + 2 PN + 2 (ξ )..... + a N + r PN + r (ξ ) = 0
ξ 0 = ξ Prathap
Gauss points are thus special cases of Prathap points 80
 Example 4. Cantilever bar analysis using one linear element
under linearly varying axial load distribution
q=-q0 (x/L)
x q q0

L
L/2

q0 L
εh = Analytical strain
6 EA (quadratic)
q0 L
FEM: constant strain (by Rayleigh-Ritz)
2 EA
L Optimal
x0 = point
3
L 2 x0 2
Optimal Point εh =ε x0 = ξ0 = −1 = − 1−
3 L 3

d du x q0 x 2
Analytical: EA = −q = −(−q0 ) ε=
dx dx L 2 EA L
du h q0 L
Rayleigh-Ritz: h
u ( x ) = −a ( L − x) u ( L) = 0 h
ε = = 81
dx 6 EA
 Rayleigh –Ritz calculations for Example 4
Cantilever bar analysis using one element
under linearly varying axial load distribution
q=-q0 (x/L)
x q q0

Using a linear displacement function (or a constant strain)

u h ( x ) = −a ( L − x) u ( L) = 0
h du h
ε = =a
dx
2
( )
L h L L L
1 du x h 1 x
Π= EA dx − q0 u dx = EAa 2 dx − q0 {− a( L − x)}dx
20 dx 0
L 20 0
L

∂Π q0 L h du h q L
=0 a= , ε = =a= 0
∂a 6 EA dx 6 EA

82
 Example 5. Cantilever bar analysis using two linear elements
under linearly varying axial load distribution
q=-q0 (x/L)

q q0

L
L/2
q0 ( L / 2)
6 EA
Analytical strain
(quadratic)
FEM: constant strain q0 L
(by Rayleigh-Ritz)
2 EA
(L / 2)
3
(L / 2)
3

83
 4.3 Prathap points as general optimal points
(for statically indeterminate problems)
For cases where the system suffers from spurious
stiffening of the system due to nodal reaction errors.
The FE strain is the orthogonal projection (best-fit) of
the stiffened analytical strain (from nodal reaction errors)

he
m < ε e stiff , v j >
{ε } = {v j }, vi v j for i ≠ j; span( B) = (v1 , v2, ...vm )

{} {}
j =1 < vj,v j >
he e e
{ε } = ε + ∆ ε

{ε }=
e

j =1
m < ε e,vj >
< vj,vj >
{v j }; { }=
∆ε
e
m

j =1
< ∆ε e , v j >
< vj ,vj >
{v j }

Best-fit of additional analytical

Best-fit of original analytical strain
strain from nodal reaction errors

For indeterminate cases, Prathap points are determined from the

84
best-fit strain of the stiffened analytical strain from reaction errors
 Example 6.
Fixed-fixed tapered bar analysis with two linear bar elements.
Case : Uniform loading q = 1

E = 1, l = 1, A2 = 1 and A1 = 0.01, u(x=0)=u(x=l)=0

The exact (analytical) solutions for this case :

Displacement : u = -1.010101x + 0.219341 ln (1 + 99x)

Strain: = -1.010101 + 21.714759 / (1 + 99x)

Stress resultant: Q = 0.2070 – x

A1

A2
85
 X

A
1

Case 1. Fixed-fixed indeterminateA2

bar with u.d.l. q=1
Element e 1 2
FEM Strain
0.4950 -0.4950
{ε he }
Node 1 2 3 Best-fit Strain
e
{ε } -0.1671 -0.7216
Position x 0. 0.5 1.0
Sectional Area, 0.01 0.505 1.0 e
{ε he } − {ε } 0.6621 0.2266
A 0
Exact 0. 0.355 0. FEM -0.3775 0.1225
displacement u 2 Reaction -0.1225 -0.6225
FEM 0. 0.247 0. {Rhe}
displacement uh 5 Analytical -0.2070 0.2930
du / dx 20.7 -0.58 - Reaction -0.2930 -0.7930
0 0.792 {Re}
Exact strain
9 Reaction -0.1705 -0.1705
Error 0.1705 0.1705
{Rhe}–{Re} 86
 0.3775

A1

Case 2. Free-fixed bar with u.d.l. q=1, and A2

Load P=0.3775 ( ) at the free left end

Element e 1 2
{ε he }
0.4950 -0.4950
FEM Strain

Node 1 2 3 e
{ε }
0.4950 -0.4950
Position x 0. 0.5 1.0 Best-fit Strain

Sectional Area, 0.010 0.505 1.0 {ε he } − {ε }

e

A 0 0

Exact 0.
displacement u -0.7931 0.2376 FEM Reaction -0.3775* 0.1225
{Rhe} -0.1225 -0.6225
FEM 0. 0.2475 0. Analytical -0.3775* 0.1225
displacement uh Reaction -0.1225 -0.6225
{Re}

*The externally applied force at the Reaction

Error
0
0
0
0
free end appears as the end reaction. {Rhe}–{Re}

87
 0.3775
X
X

A
1 A1

Case 1. Fixed-fixed indeterminateA2 Case 2. Free-fixed bar with u.d.l. q=1, and
A2
bar with u.d.l. q=1 Load P=0.3775 ( ) at the free left end

88
 Summary
A review of linear algebra has been made. The concept of vector spaces and
projections are introduced. Orthogonal projections of vectors onto vector spaces
(including polynomial spaces) have been demonstrated.

The weak forms that naturally arise in finite elements (from virtual work principles)
have been shown in terms of inner products and norms of strain vectors in function
spaces. The special cases of one dimensional elements (bar and beam elements)
have been clearly demonstrated with suitable examples.

It has been shown how FE formulations actually lead to the ‘best-fit paradigm’ of
strains (and therefore stresses) in the mathematically abstract language of linear
algebra. Element strains are best-fits (or orthogonal projections) of analytical strains
onto function subspaces B that are generated from the strain-displacement matrices.

For statically determinate systems, the FE strain is the best-fit of the original
analytical strain vector. For statically indeterminate systems, displacement (and
strain) approximations can lead to errors in nodal reactions, that make the FE strain
deviate from the best-fit. However, even for such cases, the FE strain stands as best-
fit to the element strain from stiffened analytical solution.

Optimal strain/stress recovery points (Prathap Points) result from the best-fit
paradigm. It has been shown that the well known Gauss points for strain recovery
are actually special cases of the Prathap points. For finer meshing, Prathap points
progressively march towards Gauss points. 89
 When Arts and Science met at the crossroads…

An extract from “ Sanchaita ” by Rabindranath Tagore.

A
A−A
900
v

u P
A
Subspace

90