Prof. A.

Meher Prasad
Department of Civil Engineering
Indian Institute of Technology Madras
email: prasadam@iitm.ac.in
Rayleigh’s Method

Background:
Consider that an undamped SDF mass-spring system is in free
harmonic motion, then
x = X sin (pt+α)
. (E1)
x = pX cos (pt+α)
The strain energy of the system, V, at any time t is given by
(E2)
V = ½ kx2 = ½ kX2 sin2(pt+α)
and its kinetic energy, T, is given by
.2
T = ½ mx = ½ mp2X2 cos2(pt+α) (E3)
The principle of conservation of energy requires that, the sum of V
and T be the same.
Note that when V = Vmax, T=0, and when T= Tmax, V =0. Hence
Vmax = Tmax (E4)
 or ½ kX2 = ½ mp2X2 (E5)
From which we conclude that p2 = k/m (E6)
This is identical to the expression obtained from the solution of the
governing equation of motion.
As a second example, consider the SDF system shown
2
1 a a  1 a 2
Vmax =[k   y 0 ]  y0  k   y0
2 L L  2 L
2
x 2
L
1 2 1 2 2 1
T= ∫ µ + = p y 0 [ µL + m]
2
max p [   0
y dx my 0 ]
2 0 L 2 3
μ m
Equating Vmax and Tmax, we obtain
2 y0
a k
k 
p2 = L (E7)
1 a
m + µL
3 L
Rayleigh Quotient

Consider now a MDF system in free vibration, such that

{x} = {X} sin(pt+α)
. (E8)
{x} = p {X} cos(pt+α)
The maximum strain energy of the system is
½ {F}T[d]{F}
Vmax = ½ {F}T{X} = (E9)
½ {X}T[k]{X}

In which {F} are the static modal forces corresponding to the

displacement amplitudes {X}, and [d] and [d] are the flexibility and
stiffness matrices of the system.
The maximum Kinetic energy of the system is given by
~
Tmax = ½ p2{X}T[m]{X} = p2 Tmax (E10)

~
In which Tmax = ½ {X}T[m]{X} (E11)
will be referred to as the maximum pseudo-kinetic energy of the
system.
The principle of conservation of energy requires that
~
Vmax =Tmax = p2 Tmax

Vmax
where p =
2
~ (E12)
T max
The above equation is known as Rayleigh’s quotient.
The Eq.(E12) could also be obtained from the equations of motion of
the system as follows:
..
[m] { x } + [k] { x } = 0
Making use of Eq.(E8) and pre multiplying the resulting equation by
{X}T, we obtain,
- p2 {X}T[m]{X} + {X}T[k]{X} = 0

~
Tmax Vmax

Vmax
where p2 = ~
T max
Properties of Rayleigh quotient

• If Vmax and T~ in Eq.(E12) are evaluated for {X} = to the jth mode {X },
max j
2 2
then the value of p will be precisely pj .
• If Eq.(E12) is evaluated for a vector {X} which does not correspond to
a natural mode, then the resulting value of p2 will not be a natural
frequency. Furthermore, to each assumption of {X} there corresponds
a different value of p2. In fact, if we recall that (n-1) displacement ratios
are necessary to define the configuration of a MDF system, we may
conclude that p2 in Eq.(E12) defines a “surface” in a space having (n-1)
dimensions.
• It can be shown that
(1) The natural frequencies of the system, i.e., the values of p2 obtained
when {X} is equal to any of the natural modes, correspond to
stationary (maximum, minimum or saddle) points of this surface.
It follows that an error in estimating the mode will produce an error
in frequency which is of smaller order (since the surface is ‘flat’ in
the neighbourhood of the stationary point).
2. The values of p2 obtained for an arbitrary {X} lies between the
lowest and highest natural frequencies of the system(i.e., p12 < p2 <
pn2).
It follows that if one assumes an {X} which approximates the
fundamental natural mode of the system, then the resulting value of
p will be greater than (will represent an upper bound estimate for)
p1. Similarly, if one assumes an {X} that approximates the highest
natural mode, the resulting value of p will be lower than (will
represent a low bound for) the true pn.
Because the fundamental natural mode of a system can normally be
estimated with reasonable accuracy, the procedure is ideally suited to
the computation of the fundamental natural frequency.Application of the
procedure requires the following steps:

1. Estimate the fundamental mode of vibration. This may be done

either by assuming directly the displacements of the nodes, or the
associated forces and computing the displacements.
~
2. Compute the values of Vmax and Tmax corresponding to the
estimated mode.
V
3. Evaluate p2 from, p = ~ max
2

T max

The value thus determined is higher than the true fundamental

natural frequency of the system, and, unless the assumption
regarding the mode was quite poor, it will be close to the actual
frequency.
If the frequency is computed for several different assumed
configurations, the smallest of the computed values will be closest to the
exact fundamental frequency, and the associated configuration is closest
to the actual fundamental mode.
The details of the procedure are illustrated by a series of examples.

Example #1

m 1 2 1.5
x2
k
m 1 1 1
x1
k

(a) (b) (c)

Assumption 1: Take x1=x2=1, as shown in Fig. (a)

Vmax may be evaluated from the deformations of the stories without

having to determine first the stiffness matrix of the system.

2 Vmax= k [ (1)2+0] = k

~
2 Tmax= m [ (1)2+(1)2 ] = 2m

k k
=
p 2
= 0.5
2m m
Assumption 2: Take x1=1 and x2= 2, as shown in Fig.(b)
2 Vmax= k [ (1)2+(1)2 ] = 2k
~
2 Tmax= m [ (1)2+(2)2 ] = 5m

2k k
=
p2 = 0.4
5m m

Assumption 3: Take x1=1 and x2= 1.5, as shown in Fig.(c)

2 Vmax= k [ (1)2+(0.5)2 ] = 1.25k
~
2 Tmax= m [ (1)2+(1.5)2 ] = 3.25m

1.25k k
=p = 0.3846
2

3.25m m
Discussion

1. Assumption 3, which leads to the lowest frequency value, is the best

of the three approximations considered, and is only slightly off the
exact value of,
1 k k
p12 = [3- 5] = 0.38197
2 m m

2. That assumptions 1 and 2 would be poor, could have been anticipated

by considering the forces necessary to produce the assumed
configurations.
The deflection configuration in Fig.(a) is produced by a single
concentrated force applied at the first floor level, whereas the
configuration in Fig.(b) is produced by a single concentrated force
acting at the second floor level.
Clearly, neither of these force distributions is a reasonable
approximation to the inertia forces associated with the motion of the
system in its fundamental mode.
3. That assumptions 1 and 2 The deflection pattern considered in Fig.(c)
is produced by lateral forces which are proportional to the weights of
the floors. If these forces are denoted by F, the resulting displacements
are as shown.
2F/k + F/k
F

2F/k
F

Fig (2)

Subject to the justification noted later, this assumption generally leads

to an excellent approximation for the fundamental natural frequency of
the system.
4. In the following diagram, the value of p2 determined by application of
Rayleigh’s method is plotted as a function of the displacement ratio
x2/x1. Note that the two natural frequencies correspond to the
maximum and minimum points of the curve, and that in the vicinity
of these extremum points the frequency values are insensitive to
variations in the displacement ratio x2/x1.

3
2.618

p2
2
k m
1 0.38197
-0.618
1.618
-2 -1 0 1 2 3
X2/ X1
5. The first step in the solution that has been presented was to
estimate the mode of vibration of the system. Alternatively, we could
have assumed the distribution of the inertia forces associated with
the mode of interest and compute Vmax as the product of these
forces and the resulting displacements. For example, for the forces
considered in Fig.2,

2 Vmax = F∗ F/k ∗ [2+3] = 5 (F2/ k)

~
2 Tmax = m ∗ F/k ∗ F/k ∗ [22+32] = 13m (F2/ k2)

p2 = 0.3846 (k / m)

which is the same as the answer obtained above.

Example # 2

0.5m k x3 4+0.5 = 4.5 1.8

0.5k

m k x2 2.5+1.5 = 4 1.6
k

m k x1 k
2.5 1

Assume a deflection configuration equal to that produced by a set of

lateral forces equal to the weights of the system.
2Vmax = k [ 2.5+4+ ½(4.5) ] = 8.75 k
( This can also be evaluated from the story deformations as
k [ 2.52 + 1.52 + 0.52 ] = 8.75k )
 ~
2 Tmax = m [ 2.52 + 42 + ½(4.52) ] = 32.375m
8.75k k
=p2 = 0.27027
32.375m m
The exact value of p2 is 0.2680k/m, and the error in p is only 0.42%.
In the following figure, the value of p2 is plotted as a function of the
displacement ratios X2/ X1 and X3/ X1 in the range between –3 and 3.
As would be expected, there are three stationary points:
(1) a minimum point of p2=0.268k/m corresponding to the fundamental
frequency;
(2) a maximum value of p2=3.732k/m corresponding to the third natural
frequency;
(3) a saddle point of p2=2k/m corresponding to the second natural
frequency. The values X2/ X1 and X3/ X1 of the associated modes can
be read off the figure.
 **20 squares to the inch

Dimensional representation of Rayleigh’s quotient

for a 3 DOF system
Example # 3

Assumption # 1:
πx
y ( x ) = y0 sin m 1.5 m
l
l
π4 l
nx
= ∫0 ( y ) dx L4 y EI ∫0 sin L dx
'' 2 2 2
2Vmax EI o L/3 L/3 L/3

x
2Tmax m[[ y ( L )]2 + 1.5[ y (2 L )]2 ] p 2
3 3
3 3 2 2 3 9 2 15 2 2
= my [ + 1.5 ×=
2
0 ] p my0 [ + ] p = my0 p
4 4 4 8 8
π 4 EIy02 8 1 4π 4 EI EI
p2 = = 25.98
2 L3 15 my02 15mL3 mL3
The exact value of ,
EI
2
pexact = 25.8463
mL3
Assumption #2 1.5W
W

4 WL3 3.5 1.5WL3 9.25 WL3 4 1.5WL3 3.5 WL3 9.5 WL3
+ = + =
243 EI 243 EI 243 EI 243 EI 243 EI 243 EI

1 W 2 L3
=
2Vmax [1.0 × 9.25 + 1.5 × 9.5]
243 EI
2
W  1 WL  3
= × ( ) + × ( )
2 2
2Tmax  243 EI  [1.0 9.25 1.5 9.5 ]
g  
setting Vmax = Tmax ,
EI
p 2 = 25.8467
mL3

Which is in excellent agreement with the exact value

Application to systems with Lumped masses:

m1 m2 .... mj .... mn

Let F1,F2, Fj,…Fn be the inertia forces corresponding to the assumed

mode and y1,y2……yj,…..yn be the deflections induced by these
forces.
F1 F2 .... Fj .... Fn

yj

∑F y j j

Then, p2 = g j =1
n

∑W y
j =1
j
2
j
Selection of Fj

As already demonstrated, good accuracy is achieved by taking the

forces Fj to be equal to, or proportional to the weights Wj. It should be
realized, however that these forces are not the exact inertia
forces.Rather they represent the inertia forces associated with a uniform
(rigid body) motion of the system.

An improved approximation may be achieved by assuming a

configuration for the mode and taking Fj as the inertia forces
corresponding to the assumed configuration.It is important to note that yj
in Eq. E13 are the deflections produced by the forces Fj, not the
deflections assumed for the purpose of estimating Fj
Illustration:
•As a guide in the selection of the forces { f }, assume that the fundamental
mode varies as a sine curve , as shown in fig (a)
•The inertia forces corresponding to this assumption are shown in fig.(b).
For the example considered, these turn out to be the exact forces, and
hence the frequency computed from these forces will be exact in this case.

0.5m 1
x3 0.5F 0.5F
m √3/2
x2 √3/2F √3/2F
m
x1 0.5 0.5F 0.5F

(a) Assumed Mode (b) Inertia Forces

• If the mode were assumed to increase linearly along the height, the
forces and the deflections would be as follows:
1.5F 9.5 F/k

2F 8 F/k

F 4.5 F/k

• In this case
F2 F2
=
2V max [1*4.5 + 2*8 + 1.5*4.5
=] 34.75
K K
2 2
F F
=2Tmax m [ 20.25 + 64 + .5*90.25
=] m 129.375
K K
34.75 K K
=p 2
= 0.2686
129.375 m m
(The exact valve is 2- 3 = 0.26795)
Another example :
m1 m2

• Taking the {F} proportional to the masses will not lead to

satisfactory results. Directions important
• Take
F1 F2
Applications to continuous systems

EI, μ, L

πx
Let y ( x) = y0 sin
L
2Vmax may be evaluated by either of the following procedures
2 πx
L
= =
a) 2Vmax EI ∫ ( y ) dx " 2
π 4 EI 2
L4
y0∫ sin
L
dx
0 0
L
b) 2Vmax = ∫ q y dx
0

where q is the lateral force associated with the assumed deflection

 y(x) = yosin(πx/L)

This force may be determined from the differential equation for beams.
Recalling that , 4
d y
EI 4
= q( x )
dx
π4 nx
We obtain, q ( x ) = EI 4 yo sin
L L
Hence,
π4 L
nx
2Vmax = EI yo2 ∫ sin 2
L4 0
L
Which is the same as the result found by procedure (a)
 L L
2 nx
= 
∫0 y dx µ yo ∫0 sin L dx
2Tmax µ= 2 2

L
EI 2 2 nx
π44
y o∫ sin dx
L L 4 EI
p12 = 0
π
L
nx µ L4
µ yo2 ∫ sin 2 dx
0
L
π2 EI
p1 =
L2 µ
This is the exact frequency – Explain why?

If we had assumed as y (x) the deflection produced by a single

concentrated force at the center, we would have found that,

98.82 EI
p12 =
L4 µ
where
9.94 EI
p1 =
L2 µ

This is 0.8% too high

Dunkerley’s Method

{ x} = [ d ]{F }
Equation of motion : [ d ][ m ]{
x} + { x} = {0}
Let [m] be the diagonal matrix, [ d ][ m ]{
x} + { x} =
{0}
 1 
[ d ][ m ] − 2 [ I ] { x} = {0}
 p 
set d11m1 − 2
1
d12 m2 . . d1n mn
p
1
d 21m1 d 22 m2 − 2 . . d 2 n mn
Det p
=0
. . . . .
. . . . .
1
d n1 d n 2 m2 . . d nn mn −
p2
Given a nth order polynomial equation, (1/p2)
n n −1
 1   1 
 p 2  + ( d11m1 + d 22 m2 + .... + d nn mn )  p 2  + .... =
0
   
Sum of the roots of characteristic equation,
 1 1 1 
 2 + 2 + .... 2= ( d11m1 + d 22 m2 + .... + d nn mn )
 p1 p2 pn 
kii 1
let pii = where kii
mi dii

dii is the flexibility coefficient equal to deflection at i resulting from a

unit load of i, its reciprocal must be the stiffness coefficient kii,
equal to the force per unit deflection at i.
The estimate to the fundamental frequency is made by recognizing
p2 ,p3 etc are natural frequencies of higher modes and larger than
p1.

By neglecting these terms(1/p22, …1/pn2) ,1/p12 is larger than its true

value and there fore p1 is smaller than the exact value of the
fundamental frequency

1 1 1 1
≤ + + ....... +
p12 p112 p22
2 2
pnn

1 1 1 1
2
 2 + 2 + ....... + 2
p1 p11 p22 pnn
Dunkerley’s Approximation

It provides a lower bound estimate for the fundamental frequency.

Let p = natural frequency of system
pA, pB, pC, …….. pN = exact frequencies of component systems

Then 1 1 1 1 1
 + + +...........+
p2 p2A pB2 pC2 pN2

or T 2  TA2 +TB2 +TC2 +.................+TN2

The frequency so determined can be shown to be lower than the

exact.
Example # 1

0.5m 0.5m

k
m m
k
m m
k
k k k
p2A = pB2 = pC2 =
m 2m 3m/2

1 m m
= [1+2+1.5]=4.5
p2 k k
k k
p2 ≥ =0.2222
4.5m m
If natural modes of component systems A, B, C are close of each other,
then the value of p determined by this procedure can be shown to be
close to the exact.
Example # 2

Consider the cantilever beam shown for which

the component systems A,B,C are indicated .

m m m/2
Since the natural modes of the
system are in closer agreement in this L/3 L/3 L/3
case than for the system of the shear
beam type considered in the previous
m
example, the natural frequency A
computed by Dunkereley’s method L/3 2L/3
can be expected to be closer to the
exact value than with case before. m
B
2L/3 L/3

m/2
C
L
 EI EI
p 2A = 3 3
= 81
(L 3) m mL3
EI 81 EI
p 2B = 3 3
= 3
(2L 3) m 8 mL
EI EI
p C2 = 3 3 =6
L ( m 2) mL3
1 1 1 1 1 8 1  mL3
≈ 2 + 2 + 2 = + + 
p 2
p A p B p C  81 81 6  EI
1  1 1  mL3 15 mL3
≈ +  =
p 2  9 6  EI 54 EI
EI
p 2 ≈ 3.6 3
( Low bound )
mL
EI
2
( pexact = 3.7312 3
)
mL
As expected the agreement is excellent in this case.
Example # 3

Upper bound: Determined by Rayleigh’s method with y(x) = y0 sin(πx/L) is,

π 4 (EI L3 ) μ
p ≤
2 m
µ L + 2m
Lower bound: Determined by Dunkerley’s approximation
If we consider one mode,
π 4 EI 48EI
=p = 2 2
; pB
µL 4 A
mL3
1 µ L4 mL3  π 4 m  L3
= 4 + = µ L + 
p 2
π EI 48EI  48  π 4 EI
(EI L3 )
where p ≥
2

µ L π 4m
( + )
90 48
For m = µ L, we find
π4 EI π 4
EI
≤ p 2
≤
3.029 µ L4 3 µ L4
 x μ dx
m

l
µ x 2 (l − x) 2 dx ml 3 µ l 5
ml 3

∫
0
3lEI
+ =
48EI 3lEI
[l 2l 3 − 2ll 4 + ] +
5 48 EI
µl 4 1 1 1 ml 3 µl 4 ml 3
= [ − + ]+ = +
3EI 3 2 5 48 EI 90 EI 48 EI
π 4 EI π 4
EI
≤ p1 ≤
2

3.117 µ L4
3 µ L4
µ l 4
 ∞
1  ml 3
p12 ≈ 4  ∑ 4  +
Consider all modes,
π EI  n =1 n  48EI
µ l 4
 π 4
 ml 3
p12 ≈ 4   +
π EI  90  48 EI
Limitation of procedures:

• One cannot improve the accuracy of the solution (depends on the

deflected shape of structure) in a systematic manner.

• Extension of procedure : Rayleigh - Ritz

Stodola Method – (Method of Iteration)

• Applicable to systems governed by [H]{X} = λ{X} ,where [H] is not

necessarily symmetric.
• Meaning to a solution : Finding an {X} which when operated by [H]
will give a vector is proportional to itself.Then {X}=characteristic
vector and λ = the associated characteristic value.

Procedure:

1. Assume an {X}.
2. Compute [H]{X}
3. If step 2 results in a vector which is proportional to {x}, then {X} is a
characteristic vector, and the factor of proportionality is the
associated characteristic value.
4. In general, the vector computed in 2 will not be proportional to
{X}. Now if we take as our next assumption the result of step 2
and repeat the process, the procedure will converge to the
characteristic vector associated with the largest characteristic
value λ .

5. It can further be shown that if, at the end of a cycle, we compute

the values of λ which will make the elements or components of
the derived and assumed vectors equal, the highest characteristic
value lies between the largest and smallest of these values.
 Stiffness Matrix Flexibility Matrix

{0}
x} + [k ]{x} =
[m]{ [d ][m]{ {0}
x} + {x} =

{x} = {X }sin( pt + α )
{0}
- p 2 [d ][m]{ X } + {x} =
2
{0}
- p [m]{ X } + [k ]{ X } =
1
[d ][m]{ X } = 2 { X }
[m]-1[k ]{ X } = p 2 { X } p
1
[G ]{ X } = p 2 { X } [ H ]{ X } = 2 { X }
p
Converges to highest natural Converges to fundamental
frequency and mode natural frequency and mode
 Example Frame

 2 -1 0  1 1 1 0.5m k x3
[ k ] = k -1 2 1  [ ]
d =
1
1 2 2 
 0 -1 1  k  

1 2 3 
m k x2

m k x1
1 0 0 
[ m] = m 
 0 1 0 


 0 0 0.5 

Using Flexibility formulation first,we obtain

1 1 1  1 0 0     
m  0 1 0   X  = 1  X 
1 2 2      2  
k   p  
   
1 2 3   0 0 0.5    
1 1 1 2     
1 2 1   X  = k  X 
    mp 2  
     
1 2 1 2    
k
Let 2
=λ
mp
1  2.5 
   
 
1 ⇒ λ  4.0  2.5 ≤ λ ≤ 4.5
1  
   4.5 
1  3.5  3.5
   
1.6  ⇒ λ 6.0  3.75 3.5 ≤ λ ≤ 3.83
1.8  6.9 
    3.83
1  3.7  3.7
   
1 .71 4  ⇒ λ  6.399  3.75 3.70 ≤ λ ≤ 3.75
1.914  7.385 
    3.83
1  3.729  3.729
   
1.729  ⇒ λ  6.454  3.733 3.727 ≤ λ ≤ 3.733
1.996  7.452 
    3.733
⇓
1.0  1 
   
1.732  exact is  3 
1.999   2 
   
1 0 0   2 − 1 0     
 0 1 0   -1 2 -1  X  = p m  
2

    K
X 

 0 0 2   0 -1 1      
 
ξ
 2 -1 0     
 -1 2 -1   X  = p m  
2

   K
X 
 
 0 - 2 2     
 
1 1 
    Obviously it is diverging from
 
1 ⇒ 0  ⇒ the fundamental mode
1 0 
   
1  3 
   
 
-1 ⇒ ξ -4  3≤ξ ≤ 4
1  4 
   
1  10 3 3.33
   
 −4 3  ⇒ ξ  −5  3.75 3.33 ≤ ξ ≤ 4
4 3  16 3
    4
1  3.5  3.5
   
 − 3 2  ⇒ξ  -5.6  3.73
8 5  6.2 
    3.88
1  3.6  3.6
   
 −1.6  ⇒ ξ -5.971 3.732 3.6 ≤ ξ ≤ 3.87
1.771 6.742 
    3.87

3.6 - 5.971- 6.742

ξ = 3.7321
1- 2.60 -1.771

ξ exact = 3.7321
Combining iteration with Rayleigh Method

{X }
T
[ k ]{ X }
Rayleigh Quotient ⇒ p2 =
{ X }T [ m]{ X }

1
Stodola Method: [ H ]{ X } =
2
{X }
p
Where [ H ] = [d ][m]
Let [ H ] { X j } = {V j +1}

From Rayleigh’s Quotient,

{V }
T
j +1 [k ]{V j +1}
p1
2
=
{V j +1}T [m]{V j +1}
 But= =
{V j +1} [ H ]{ X j } [d ] [m] { X j }

{V }
T
j +1 [k ][d ][m]{ X j }
i.e. p 1
2
=
{V j +1}T [m]{V j +1}

{V }
T
j +1 [m]{ X j }
p 1
2
=
{V j +1}T [m]{V j +1}

If convergence is incomplete, the Rayleigh Quotient gives the better

approximation.Any error in the first mode frequency computed by the
Rayleigh Quotient is always on high side
Stodola convergence

[ H ]{ X } = λ { X }
{ X 1}= α1 {Φ1} + α 2 {Φ 2 } + ....... + α n {Φ n }
Assumed mode can be expressed as,

[ H ]{=
X} α1 [ H ]{Φ1} + α 2 [ H ]{Φ 2 } + ........... + α n [ H ]{Φ n }

{ X 2=
} α1λ1 {Φ1} + α 2λ2 {Φ 2 } + ....... + α n λn {Φ n }
λ n = Highest characteristic value
On Iteration,

{ X=
3 } α λ
1 1
2
{Φ 1 } + α λ
2 2
2
{Φ 2 } + ....... + α n n {Φ n }
λ 2
 After ‘s’ Iterations,
  λ s  λ2 
s

{X s} λn α1   {Φ1} + α 2   {Φ 2 } + ..............α n {Φ n }
s 1

  λn   λn  

 λi 
 for i ≤ n  < 1
 λn 

[
{ X S +1} m ]{ X S }
T

p 2
=
{ S +1} [ m ]{ X S +1}
1 T
X
Stodola process for the second mode

Any Vector, {X } = α 1 {Φ 1 } + α 2 {Φ 2 } + ....... + α n {Φ n }

{Φ1} α1Φ1T [ m ]{Φ1} + α 2 Φ1T [ m ]{Φ 2 } + ........ + α n Φ1T [ m ]{Φ n }
[ m]{ X } =
T

Φ1T [ m ]{ X }
i.e. α1 =
Φ1T [ m ]{Φ1}

The vector, {X }− α1{Φ1} has a zero first mode component

T


 Φ 

 
 m   X 

 X  −α Φ =







 X  − Φ




 

 1

 1 1

 
  
 1 
 T


Φ 

 m  Φ 



 1 
 
 1
 
 T 
Φ  Φ 

 
  

  m 
 
=   I  -  1  1  
 X 
 T  

Φ   m  Φ  
 
 
 
 
 


 
 1    1 

Define sweeping matrix [S1] to be,
 
 T 
Φ Φ






m

 

 


S = I - 1 1


 
 

 
 


1  T
  
Φ  Φ

  
  

 m  

 1    1
 
 

Then premultiplying any arbitrary vector {X} by the sweeping

matrix [S1] removes the first mode component

[ m ]2 = [ H ] [ S ]1

Matrix iteration is carried out for, [H] 2{X}=λ2 [X]

 1 1 1  1 0 0 
1
[d ] = 1 2 2  [m ] m 0 1 0 
k    

1 2 3  
 0 0 0.5

 1 
1 1
2 
m  1
[d ][m ] = 1 2 1  2 { }
x
k p
 3  0.5m
1 2  x3
 2 k
 1  m x2
1 1
2  k
 
[ ] 
H = 1 2 1 
m x1
 3 k
1 2 
 2
 k
λ =
mp 2
1.0 
 
= 1.732 
1.999 
 
{φ1} [m ] = [1 1.732 0.9995]
T

{φ1} [m ]{φ1} = 5.997825

T

 0.1667 0.2888 0.1666

{φ1}{φ1} [m ] = 0.2888 0.5002 0.2886
T

{φ1} [m ]{φ1} 0.3333 0.5773 0.3331 

T

 
 0.8333 − 0.2888 − 0.1666 
{φ1}{φ1} [ m ] 
T

[ S ]1 =
[I ] − = −0.2888 0.4999 − 0.2886
{φ1} [ m]{φ1}  −0.3333 − 0.5773 0.6669 
T

 

0.37795 − 0.0776 − 0.12185

 −0.0776 0.1337 − 0.070 
[ H ]1 =
[ H ][ S ]1 =
 
 −0.2442 − 0.1550 0.2564 

Iteration [ H ]1 { x} = λ2 { x}
 1  0.4221 
   
{X }
= 1  ⇒  0.1331 
 −1  −0.6536 
   
1  0.5709 
   
0.3154  ⇒ 0.0354 
 −1.554   −0.6353
   
1  0.5085 
   
0.062  ⇒ 0.0163 
 −1.113  −0.5394 
   
1  0.5045 
   
 0.032 ⇒
  0.0083 
−1.061 −0.5214 
   
⇓
1 
 
 0.016 
-1.034 
 
m
(0.5045)(1)m + (0.032)(m)(0.0083) + (−0.061) (−0.5214)
1 2
= p= 2

λ2 2
m
(0.5045) 2 m + (0.0083) 2 m + (0.5214) 2
2
0.78137
= = 2.000
0.39052
Application of Rayleigh Ritz Procedure

Execute the first few natural frequencies and the associated modes
of the beam shown , and study the rate of convergence of the results
as a function of the stiffness of the spacing ,i.e. or an appropriate
dimensionless measure of it.

L/4 3L/4
 πx 2π x 3π x
y ( x ) =a1 sin + a2 sin+ a3 sin + ...........
L L L
x
When expressed in terms of the dimensionless distance ξ =
L
The expression for y(x) becomes,
y (ξ ) =a1 sin πξ + a2 sin 2πξ + a3 sin 3πξ + ...........
The coefficients ‘a’ in this expression must be such that the value of
Vmax
p2,determined from p =
2
is stationary.
Tmax

∂Vmax 
2 ∂Tmax
This requires that, = p
∂a j ∂a j
 Consider first only Two Terms in the Series,

∂2 y 2
L
1 1
=Vmax EI ∫ ( 2 ) ∂x + k[ y ( L / 4)]2
2 0 ∂x 2
1 EI ∂ 2 y 2
1
1
2 L3 ∫0 ∂ξ 2
=Vmax ( ) ∂ξ + k [ y ( L / 4)]2

2
π π
2
1 
1
1 EI
3 ∫
[π a1 sin πξ + 4π a2 sin 2πξ ] ∂ξ + k  a1 sin + a2 sin 
2 2 2

2 L 0 2  4 2
1 EI π 2
1

∫ πξ + πξ πξ + 2πξ ] ∂ξ
2 2 2 2
3
[a1 sin 8a a
1 2 sin sin 2 16 a 2 sin
2 L 0
1  2 2π π π π
+ K  a1 sin + 2a1a2 sin sin + a2 2 sin 2 
2  4 4 2 2
  1 1 1

1 π EI  a1 ∫ sin πξ + 8a1a2 ∫ sin πξ sin 2πξ + 16 a2 ∫ sin 2πξ ] ∂ξ 
2 2 2 2 2

=
2 L3  0 0 0 
 
1  2 2 2
+ K  a1 + a1a2 + a2 
2  2 
1 π 2 EI 2  1 2 
=
VMAX  a + 16 a 2
 + ρ a + a 2
+ 2 a a
2 L3 
1 2  0 1 2 1 2
 2 
KL3
where ρ0 = 2
π EI
L 1
1 1
TMAX µ ∫ y=2
( x )∂x µ L ∫ y 2 (ξ )∂ξ
2 0 2 0
 1 2
1
or, T=
max µ L ∫ [a1 sin πξ + a2 sin 2πξ ] ∂ξ
2 0

1
=Tmax µ L [a12 ∫ sin 2 πξ ∂ξ + 2a1a2 ∫ sin πξ sin 2πξ ∂ξ + a2 2 ∫ sin 2 2πξ
2
1 1 1
= µ L [ a12 + 0 + a2 2 ]
2 2 2

Application of the above Equation for aj =a1 and aj =a2 leads to ,

1 π 4 EI 1 2
[a + ρ ( a + 2a ] = p µ L[a ]
2 L 3 1 0 1 2 2 1

1 π 4 EI 1 2
[16a + ρ (2a + 2a ] = p µ L[a ]
2 L 3 2 0 2 1 2 2
Canceling the factor ½ on the two sides of these equations , and
Introducing the dimensionless frequency parameter
p2 p2
λ0
= =
p0 2 Π 4 EI

µ L2
We obtain after rearrangement of term:

(1 + ρ0 ) 2 ρ0  a1 
    = λ a1 
  a  0
 a2 
 2 ρ0 (16 + 2 ρ0   2 

(1 + ρ 0 - λ0 ) 2 ρ0
=0
2 ρ0 (16 + 2 ρ 0 − λ0 )

Expanding,we obtain the following quadratic equation in λ0

 λ0²-(17+3ρ0)λ0+16+18ρ0=0

λ0=1/2 {17+3ρ0 ± 225 + 30 ρ0 + 9 ρ0 2 }

The modes are defined by the ratio a2/a1 this is given by,

1 + ρ 0 − λ0
a2 / a1 = -
2 ρ0

a2 2 ρ0
=
a1 16 + 2 ρ 0 − λ0
Consider Three Terms in Series

1
EI 4
2=
Vmax 3 ∫
π [1
a sin πξ + 4 a 2 sin 2πξ + 9 a 3 sin 3πξ ] dξ
2

L 0
π π 3π 
2

+ K a1 sin + a2 sin + a3 sin 
 4 2 4

π 4 EI  2
3  1 ∫
a sin 2
πξ d ξ + 16 a 2 ∫
2
sin 2
2πξ d ξ + 81a 3 ∫
2
sin 2
3πξ dξ + 0 + 0 + 0
L 
 1 1 1 1 1 1 
+ K a12 + a2 2 *1 + a32 * + 2a1a2 + 2a1a3 + 2 a 2 a3 
 2 2 2 2 2 2
or,
π 4 EI  1 2 1 2 1 2 1 2 1 2 
=
2Vmax a + 16* a + 81* a + ρ ( a + a 2
+ a + 2 a a + a a + 2 a a
L3  2
1 2 3 0 1 2 3 1 2 1 3 2 3
2 2 2 2 
 2Tmax { }
µ L a12 ∫ sin 2 πξ dξ + a2 2 ∫ sin 2 2πξ dξ + a32 ∫ sin 2 3πξ dξ + 0 + 0 + 0

or,

 1 2 1 2 1 2 
= µ L  a1 + a2 + a3 
2Tmax
2 2 2 

Application of the Equation for aj =a1 and aj =a2 leads to,

(1 + ρ0 ) 2 ρo ρ0  a1  a1 

    
 2 ρ0 (16 + 2 ρ0 ) 2 ρ 0   a2  =
λ0 a2 
  a  a 
ρ
 0 2 ρ 0 81 + ρ 0
  3   3
This leads to the determinantal equation,
1 + ρo − λo 2 ρo ρo
2 ρo 16 + 2 ρo − λo 2 ρo =
0
(15)
ρo 2 ρo 81 + ρo − λo

Expanding we obtain the following cubic equation in λ:

λ 3 − (98 + 4 ρo )λo2 + (1393 + 278 ρo )λo − (1296 + 1474 ρo ) =
0
Modes: These are defined by the ratios a2 / a1 and a3 / a1 .

Considering the first two equations (13) and eliminating a3 we obtain:

a2 2(1 − λo )
= (16)
a1 16 − λo
Considering the first and third of equation (13) , and eliminating a2 we obtain:
a3 (1 − λo )
= (17)
a1 81 − λo
Note that Equations (16) and (17) are independent of ρo. However ρo
enters in these equations indirectly through λo .The equations are valid
irrespective of the order of λo considered (i.e. for all three modes)

In considering the second mode ,it is more convenient to express it in

terms of the ratios a1/ a2 and a3 / a2 (i.e normalize it with respect to
a2).These ratios are given by,

a1 16 − λo
= (18)
a2 2(1 − λo )

and
a3 (16 − λo )
=
a2 2 (81 − λo ) (19)
Convergence of natural frequencies and modes

No of Frequency coefficient Fundamental mode Second mode Third mode

terms
used
λ1 λ1 λ1 a1 a2 a3 a1 a2 a3 a1 a2 a3

(a) For ρo=1

1 2 1

2 1.8760 18.124 1 -0.0877 0.0877 1

3 1.8665 18.089 82.044 1 -0.0867 -0.0109 0.0864 1 -0.0235 0.0129 0.0224 1

(b) For ρo=5

1 6 1

2 3.7526 28.247 1 -0.3178 0.3178 1

3 3.6661 27.097 87.237 1 -0.3050 -0.0344 0.3007 1 -0.1456 0.0723 0.1238 1

Mass Condensation or Guyan Reduction

• Extensively used to reduce the number of D.O.F for eigen value

extraction.
• Unless properly used it is detrimental to accuracy
• This method is never used when optimal damping is used for mass
matrix

{[ k ] − λ [ m]}{u} =
0
Let 'm' represent those to be restrained
Let 's' represent those to be condensed
 [ kmm ] [ kms ] [ mmm ] [ mms ]  {um } =
 −λ     0
  [ k sm ] [ k ss ]   [ msm ] [ mss ]   {us } 
{um } = Master d .o. f
{us } = Slave d .o. f
• Assumption: Slave d.o.f do not have masses – only elastic forces are
important

[ m= ms ] [ m= sm ] [=mss ] 0

{ s } [ ss ] [ ms ] {um }
u = − k
−1
k
T

( s ×1) ( s × s ) ( s × m ) ( m ×1) --- Guass Elimination Scheme

{um }  [I ] 
=   [T= ]{um } [T ]  T
{us }   − [ k ss ] [ kms ] 
−1

[ r ] [ ] [ ][ ]
= [ r ] [ ] [ m][T ]
T T
k T k T , m T
 Reduced eigen problem
[ kr ] {u=
m} [Ω] [ mr ][um ] master d.o.f
( m × m )( m ×1) ( m × m ) ( m × m ) ( m ×1)
Slave d.o.f

{u }
−1
− [ kss ] − λi [ mss ]i  [ kms ] − λi  mms
=   {um }i
T T
s i  i

• Choice of Slave d.o.f

– All rotational d.o.f

kii
– Find m ratio, neglect those having large values for this ratio
ii

– If [ Mss ] = 0, diagonal, [Kr] = same as static condensation then

there is no loss of accuracy
Subspace Iteration Method
• Most powerful method for obtaining first few Eigen values/Eigen
vectors

• Minimum storage is necessary as the subroutine can be

implemented as out-of core solver

• Basic Steps

– Establish p starting vectors, where p is the number of Eigen

values/vectors required P<<n

– Use simultaneous inverse iteration on ‘p’ vectors and Ritz

analysis to extract best Eigen values/vectors

– After iteration converges, use STRUM sequence check to verify

on missing Eigen values

[ k − µ m] =
[ L ][ D ][ L ]
T
• Method is called “Subspace” iteration because it is equivalent
to iterating on whole of ‘p’ dimension (rather that n) and not
as simultaneous iteration of “p’ individual vectors

• Starting vectors

• Strum sequence property

For better convergence of initial lower eigen values ,it is better

if subspace is increased to q > p such that,
q = min( 2p , p+8)
Smallest eigen value is best approximated than largest value in
subspace q.
Starting Vectors

(1) When some masses are zero, for non zero d.o.f have one
as vector entry.
0  0 0
 2  1 0 
=m =  , {X } 
 0  0 0
   
 1  0 1

(2) Take ku / mu ratio .The element that has minimum value
will have 1 and rest zero in the starting vector.

3  2 
 2   0 
Diagonal [ k ] =
=  , [ m]  
 4   4 
   
 8   1 
 ku /=
mu 3/ 2, ∞,1,8
0 1
0 0
{X } =  
1 0
 
0 0
• Starting vectors can be generated by Lanczos algorithm-
converges fast.
• In dynamic optimisation , where structure is modified previous
vectors could be good starting values.
Eigen value problem

[k ][φ ] = [Ω][m][φ ] (1)

[k ]n× p , [φ ]n× p
[φ ]T [k ][φ ] = [Ω] p× p (2)

[φ ]T [m][φ ] = [ I ] (3)
Eqn. 2 are not true. Eigen values unless P = n

If [φ] satisfies (2) and (3),they cannot be said that they are true
Eigen vectors. If [φ] satisfies (1),then they are true Eigen vectors.

Since we have reduced the space from n to p. It is only

necessary that subspace of ‘P’ as a whole converge and not
individual vectors.
Algorithm:

Pick starting vector XR of size n x p

For k=1,2,…..
[k ][ X k +1 ] = [m]{ X k } static

[k ]k +1 = { X }Tk +1[k ]{ X k +1} pxp

Smaller eigen value [m]k +1 = { X }Tk +1[m]{ X k +1} pxp

problem, Jacobi
[k ]k +1{Q}k +1 = [m]k +1{Qk +1}[ Λ]k +1

[ X ]k +1 = { X }k +1[Q ]k +1

Λk+1 Λ { X }k+1 - φ k ∞
Factorization [k ] = [ L][ D ][ L]T (1/2)nm2 + (3/2)nm
Subspace Iteration [k ][ X ]k +1 = [Y ]k nq(2m+1)

[k ]k +1 = [ X ]Tk +1 [Yk ] (nq/2)(q+1)

(nq/2)(q+1)
[ M k +1 ] = [ X k +1 ] [Yk +1 ]
T

[k ]k +1[Q ]k +1 = [ M ]k +1[Q ]k +1[ ]k +1

[Y ]k +1 = [Y ]k +1[Q ]k +1 nq2
Sturm sequence check ] [k ] − µ [ M ]
[k= n(m+1)
[k ] = [ L][ D ][ L]T (1/2)nm2 + (3/2)nm

[k ][φ ]ik +1 − λik +1[ M ][φi ]k +1

4nm + 5n
[k ][φi ] k +1
Total for p lowest vector.
@ 10 iteration with nm2 + nm(4+4p)+5np
q = min(2p , p+8) is 20np(2m+q+3/2)

This factor increases as that iteration increases.

N = 70000,b = 1000, p = 100, q = 108 Time = 17 hours
Example

Use the subspace Iteration to calculate the eigen pairs (λ1,φ1) and
(λ2,φ2) of the problem Kφ = λM φ,where

 2 −1 0 0 0 
 −1 2 −1 0   2 
=K = ; M  
 0 −1 2 −1  0 
   
0 0 −1 1   1

2 −1 0 0 0 0
 −1 2 −1 0 2 0
  X2 =  
0 −1 2 −1 0 0
   
0 0 −1 1 0 1

2 1
4 2
X2 =   and
4 3
 
4 4
 2 1 6 4
=K 2 4=
 1 1 ; M 8 4 3
2
  
  1 2 
 −  0   1 1 
 2 4   8 + 4 2 4 2 −8
Α2  =  ; Q2  
 1 2   1 1 
 +    4 + 4 2 4 2 − 4 
0
 2 4
  
 1 1 
 4 −
4 
 
 1
−
1 
 2 2 
and X 2 =  
 1 + 2 − 1 + 2 
 4 4 
 
 2 2 
 2 2 