CHAPTER 3: ROOT FINDING-b

– Roots of Polynomials:
– Müller’s Method
– Bairstow’s Method
– Matlab Functions for
Root Finding
Roots of Polynomials:
An order-n polynomial
f n ( x)  a0  a1 x  a2 x 2  ...  an x n

The roots of the polynomial follow these rules:

> For nth-order polynomial there are n real/complex roots.
> The roots are not necessarily distinct.
> If n is odd, there is at least one real root.
> If complex roots exist they are in conjugate pairs (a±bi).
Computing with Polynomials:
Consider a 3rd order polynomial:
f 3 ( x)  a3 x 3  a2 x 2  a1 x  a0 (Open form; 6 multiplications + 3 additions)
f 3 ( x)  (( a3 x  a2 ) x  a1 ) x  a0 (Nested form; 3 multiplications + 3 additions)
Nested form is preferred as it requires fewer operations:
> Lower cost of computation.
> Lower round-off errors.
Polynomial deflation:
 Once you computed a root of a polynomial, you need to remove
this root from the polynomial before proceeding to the next root
(so as not to trap at the same root again).
 This process is done by a process called polynomial deflation.
 Consider a third order polynomial. This polynomial can always be
written as
f 3 ( x)  ( x  r1 )( x  r2 )( x  r3 ) (factored form)

where r1, r2 and r3 are the roots. To remove the root x=r1, divide
the polynomial by (x-r1). The remaining polynomial will be a
second order polynomial:
f 2 ( x)  ( x  r2 )( x  r3 )

The remainder of this divison will be zero (R=0) since you divide by
its root. If it was not a root, R=consant (a 0th order polynomial).
 In general, if an n-th order polyomial (dividend) is divided by an
m-th order polynomail (divisor), the resulting polynomial
(quotient) is an (n-m)th order polynomail. The reminder will be
an (m-1)th order polynomial.

dividend x4 divisor

x  2 x  24
2

x6
Reminder (R) 0 quotient
 In computer the polynomial deflation is done by a process called
synthetic divison.
Synthetic division by a linear factor (x-t):
bn  an an .. a0 : coefficients of the dividend
bi  ai  bi 1t for i  n  1 to 0 bn .. b1 : coefficients of the quotient
b0 : coefficient of the reminder

For example above: a2=1, a1=2, a0=-24, t=4 > b2=1, b1=6, b0=0
 dividend divisor
x  2 x  2 x  24
4 3
x  2x  3
2

x2  3
quotient
Reminder (R) 6x  9

Synthetic division by a quadratic factor (x2-rx-s):

bn  an an .. a0 : coefficients of the dividend
bn 1  an 1  rbn bn .. b2 : coefficients of the quotient
b1 ,b0 : coefficients of the reminder
bi  ai  rbi 1  sbi  2 for i  n  2 to 0

 Synthetic division is subject to round-off errors. To minimize this

effect some strategies are used:
> root polishing: use root estimates as initial guess for successive
estimates.
> start from the root with smallest absolute value.
Conventional methods for finding roots of polynomials:
 In general, previously mentioned bracketing and open methods
can be used for finding the real roots of polynomials. However,
> Bracketing methods are difficult for finding good initial guesses.
> Open methods are problematic due to divergence problems.
 Special methods have been developed to locate both real and
complex roots of polynomials. Two conventional methods:
- Müller’s method: Fitting a parabola to the function
- Bairstow’s method: Dividing the polynomial by a guess factor
 There are other methods too,e.g.:
- Jenkins-Traub method
- Laguerre’s method
- ...
 Plenty of software library is available for locating roots of
polynomials.
 Müller’s Method
 This method is similar to Secant method. The difference is that
instead of drawing a straight line for two initial guesses, a
parabola is drawn with three initial guesses.

Secant Müller

xr
x1 x0 xr x x1 x0
2

two initial guesses three initial guesses

 Parabolic fit function facilitates finding both real and complex roots.
 Define the parabola as
f 2 ( x )  a ( x  x2 ) 2  b ( x  x 2 )  c

 We need to find the coefficients (a,b,c) such that the parabola

passes from three initial guesses: [x0, f(x0)] , [x1, f(x1)] , [x2, f(x2)] .
We get these three equations with there unknowns (a,b,c):
f ( x0 )  a ( x0  x2 ) 2  b( x0  x2 )  c
f ( x1 )  a ( x1  x2 ) 2  b( x1  x2 )  c
f ( x2 )  a ( x 2  x 2 ) 2  b ( x 2  x 2 )  c

From the third equation => f ( x2 )  c

Then, we get two equations with two unknowns:
f ( x0 )  f ( x2 )  a ( x0  x2 ) 2  b( x0  x2 )

f ( x1 )  f ( x2 )  a ( x1  x2 ) 2  b( x1  x2 )
Define
h0  x1  x0 h1  x2  x1 f ( x1 )  f ( x0 )   f ( x2 )  f ( x1 )
0  1
x1  x0 x2  x1
Then
(h0  h1 )b  (h0  h1 ) 2 a  h0 0  h11

h1b  h1 a  h11
2
 The coefficients a,b,c are determined as

1   0 h1  x2  x1
a h0  x1  x0
h1  h0
f ( x1 )  f ( x0 ) f ( x2 )  f ( x1 )
b  ah1  1 0  1 
x1  x0 x2  x1
c  f ( x2 )

 To calculate the root, calculate the point where the parabola

intersects the x-axis, i.e. f ( x)  0 . Then, the roots (xr) are
 2c (alternative formula for
xr  x 2  roots of quadratic polyn.)
b  b 2  4ac

 2c (both real and complex

xr  x 2  roots are facilitated.)
b  b 2  4ac
 Root is estimated iteratively (i.e., xrx3)
 Error can be defined as the difference between the current (x3)
and the previous (x2) solution:
x3  x2  2c 1
a  100%  100%
x3 b  b  4ac x3
2

 A problem arises for which of the two estimated roots (xr) will be
used for the next iteration.
> choose the one that makes the error smaller.
 Once x3 is determined the process is repeated. For the three
points of the next iteration, two suggested strategies are:
> choose other two points that are nearest to the estimated root
(If estimated roots are real).
> just follow the method sequentially, i.e., drop x0, and take x1 , x2
, x3 (if a complex root is located).
 Once a root is located, it is removed from the polynomial by
synthetic divison.
EX: Use Müller’s method to determine a root of the equation
f ( x)  x 3  13 x  12

With initial guesses x0=4.5 , x1=5.5 , x2=5.0.

n=1:
f (4.5)  20.625 f (5.5)  82.875 f (5)  48

h0  5.5  4.5 h1  5  5.5

82.875  20.625 48  82.875
0   62.25 1   69.75
5.5  4.5 5  5 .5
Then,
69.75  62.25
a  15 b  15(0.5)  69.75  62.25 c  48
 0.5  1
To determine the estimated root:
b 2  4ac  62.252  4(15)48  31.54

since 62.25  31.54  62.25  31.54

The estimated root is:

 2(48) i Xr a(%)
x3  5   3.976
62.25  31.54 0 5
1 3.976 25.74
Error:
2 4.001 0.6139
3.976  5 3 4 0.0262
a  100%  25.74%
3.976 4 4 < 10-5

For the next iteration, assign new

method rapidly converges
guesses (sequential replacement): to the root xr=4.

x0=5.5 x1=5 x2=3.976

 Bairstow’s Method
 We know that dividing a polynomial by its root result in a
reminder of zero. If the divisor is not a root then the reminder
will not be zero.
 The basic approach of Bairstow’s method is starting a guess value
for root and dividing the polynomial by this value. Then the
estimated root is changed until the reminder of the division get
sufficiently small.
Consider the polynomial:
f n ( x)  a0  a1 x  a2 x 2  ...  an x n

If you divide the polynomial by its factor (x-t), you get a new
polynomial of the form

f n 1 ( x)  b1  b2 x  b3 x 2  ...  bn x n 1 and a reminder R  b0

This process can be done by using synthetic division:
bn  an
bi  ai  bi 1t for i  n  1 to 0
 In order to facilitate the complex root Bairstow’s method divides
the polynomial by a quadratic factor: (x2-rx-s). Then the result
will be in the form of
f n  2 ( x)  b2  b3 x  ...  bn 1 x n 3  bn x n  2

with a reminder R  b1 ( x  r )  b0

The process can be done by synthetic division by a quadratic factor:

bn  an
bn 1  an 1  rbn
bi  ai  rbi 1  sbi  2 for i  n  2 to 0
 Then, the problem is to find (r,s) values that make the reminder
term zero (i.e. b1=b0=0).
 We need a systematic way to modify (r,s) values to reach the
correct root. In order to this, we use an approach that is similar
to Newton-Raphson method.
Considering b0 and b1 are both functions of (r,s), and expand them
using Taylor series:
b1 b
b1 (r  r , s  s )  b1  r  1 s
r s
(no higher order terms)
b0 b0
b0 (r  r , s  s )  b0  r  s
r s
 We force b0 and b1 to be zero to reach the root; then we get
b1 b1
r  s  b1
r s solve these two equations
b0 b0 for two unknowns: r and
r  s  b0 s
r s
Define
b0 b1 b0 b1
c1  c2   c3 
r r s s
Then, we have the following simultaneous equations:
c2 r  c3 s  b1
c1r  c2 s  b0
 We previously applied synthetic division to derive values of “b”
from values of ”a”. Bairstow showed that we can evaluate values
of “c” from values of “b” in a way similar to synthetic division:
cn  bn
cn 1  bn 1  rcn
ci  bi  rci 1  sci  2 for i  n  2 to 1
 Once r and s are calculated using the simultaneous equations,
Then, for the next iteration (r,s) values are improved as:
r  r  r s  s  s
 Error at each iteration can be calculated for r and s seperately:
r s
 a ,r  100% and  a,s  100%
r s
 When both values fall below a stopping criterion, the roots can
be determined as the roots of x  rx  s :
2

r  r 2  4s
x1, 2 
2
 Both real and complex roots are obtained.
 Once the roots are located, these roots are removed, and the
same procedure is applied to locate other roots.

EX: Use Bairstow’s method to determine the roots of the equation

f 5 ( x)  x 5  3.5 x 4  2.75 x 3  2.125 x 2  3.875 x  1.25
using initial guess of r=s=-1, and stopping criterion s=1%
First iteration:
Apply synthetic division to obtain b values:
b5=1 ; b4=-4.5 ; b3=6.25 ; b2=0.375 ; b1=-10.5 ; b0=11.375
Apply synthetic division to obtain c values:
c5=1 ; c4=-5.5 ; c3=10.75 ; c2=-4.875 ; c1=-16.375
Solve the simultaneous equations to get: r=0.3558 and s=1.1381.
Revise (r, s) values: r+r=-0.6442 and s+s=0.1381
Error: a,r=55.23% and a,s=824.1%

Second iteration:
b5=1 ; b4=-4.1442 ; b3=5.5578 ; b2=-2.0276 ; b1=-1.8013 ; b0=2.1304
c5=1 ; c4=-4.7884 ; c3=8.7806 ; c2=-8.3454 ; c1=4.7874
Solution of the simultaneous equations: r=0.1331 and s=0.3316.
Revise (r, s) values: r+r=-0.5111 and s+s=0.4697
Error : a,r=26.0% and a,s=70.6%

Third iteration:
Forth iteration:
r=-0.5 with a,r=0.063%
s=0.5 with a,s=0.040%

Errors satisfy the error criterion. Then, we obtain first two roots as:
 0.5  (0.5) 2  4(0.5)  0.5 
r1, 2   
2  1.0

For the other roots, we use the deflated polynomial:

f ( x)  x 3  4 x 2  5.25 x  2.5
Apply Bairstow’s method with starting guesses r=-0.5 and s=0.5
(solution of the previous step) .
After five iterations, we get r=2 and s=-1.249. Then, the roots are
2  2 2  4(1.249)
r3, 4   1  0.499i
2
We can easily obtain the fifth root by deflating the polynomial into
the first order. The fifth root will be r5=2.
 Matlab functions for root finding
function Method
fzero function is  find the real root of a single non-linear
commonly used to find equation
root of non-linear or  A combination of bisection method + secant
complex equations (Müller’s) methods
 It first search for sign change, then applies
secant algorithm.
 If the root falls outside the interval, bisection
method is applied until an acceptable result is
obtained with faster methods.
 Roots wıth sıgn changes are located.

 determines all real and complex roots of

polynomials.
 uses the method of eigenvalues.