Order Of Convergence Of An Iterative Scheme
Let the sequence of iterative values { xn } ¥n = 0   converges to 's'. Also let e n = s-xn and e n+1= s-xn+1 for n > 0 are the errors at nth and (n+1)th iterations respectively. If two positive constants A ¹ 0 and R > 0 exist, and
          lim
       n ®¥
 |s-xn+1|
    lim
   n ®¥
 |en+1 |

    = 
  =  A
  |s-xn|R
  |e n|R
then the sequence is said to converge to 's' with order of convergence R. The number A is called the asymptotic error constant.

This can be derived from Taylor series as follows

Let xi+1= g(xi) define an iterative method, and let 's' and xn  respectively are the exact and approximate solutions of        x = g(x). Then xn= s + e n ,        where en is the error in xn.  Suppose that g is differentiable any number of times, then from Taylor's formula
xn+1 = g(xn) = g(s + en)
        = g(s) + en g'(s) + 1 e n2 g''(s) + ... 
2
The exponent of en in the first non-vanishing term after g(s) is called the order of the iteration process defined by g. Since xn+1 - s = en+1 (and g(s) = s) the above equation can now written as
e n+1 = en g'(s) + 1 e n2 g''(s) + ... 
2
that is the error at (n+1)th iteration can be written as a function of error at the previous iteration. In the case of convergence en is small for large n and hence the order is a measure for the speed of convergence. For example if a scheme is second order, that is
e n+1   1 e n2 g''(s)
2
then the number of significant digits are approximately doubled in each step.
Order of Newton's Method: Since for Newton's method
    g(x) = x -   f(x)


 f '(x)

 
    g'(x)    f(x) * f ''(x)


     (f '(x))2
at x = s, g' = 0 since f(s) = 0  and 
      g''(x) 2f ''(x)


 f '(x)

 
 
at x = s, g''(s) need not be zero, hence Newton-Raphson method is of order two. That is for each iteration the scheme converges approximately to two significant digits.
Order of Fixed Point Iteration method : Since the convergence of this scheme depends on the choice of g(x) and the only information available about g'(x) is  |g'(x)| must be lessthan 1  in some interval which brackets the root. Hence g'(x) at x = s may or may not be zero. That is the order of fixed point iterative scheme is only one.
Order of Secant method : The g(x) for secant method is
xi+1= xi  (xi - xi-1) * f(xi

     f(xi) - f(xi-1)
If  xi+1 = s + ei+1,  xi = s + ei  and xi-1 = s + ei-1 then
s + ei+1 = (s + ei ) -  (ei - ei-1) * f(s + ei )

f(s + ei ) - f(s + ei-1)

 
® e i+1 = ei  -  (ei - ei-1) * f(s + ei )  

f(s + ei ) - f(s + ei-1)

 
              =   ei  -                  (ei - e i-1) * (eif '(s) + ( ½)ei 2f ''(s) + . . .)  

 (eif '(s) + ( ½)ei 2f ''(s) + . . .) - (ei-1 f '(s) + (½)e i-12 f ''(s) + . . .)

 
              =   ei  -     (eif '(s) + ( ½)ei 2f ''(s) + . . .)  

(f '(s) + (½)(e i+ ei-1 )f ''(s) + . . .)

 
                 =   ei - ( ei + 1  ei f ''(s) + . . .) (1 +  1 (ei-1 + ei) f ''(s)  + . . .)-1
2 f '(s) 2 f '(s)

 
      e i+1    =  1  ei ei-1 f ''(s)  + 0(ei 2ei-1 + e iei-12)
2 f '(s)

 
    ei+1 =  C eiei-1        where  C = 1 f ''(s)
2 f '(s)
and higher order terms are neglected.
since          ei = Aepi-1

           ei-1 = A-1/pe i1/p

=>           ei+1   =    Aei(A -1/pei 1/p)      =     A1-1/pe i1+1/p

=> order of the scheme p = 1 + 1/p


 
                                    p = 1 (1 ± 51/2)   = 1.618 
2

                                 and A =C p/(p+1)
 
 

 


Solution of Transcendental Equations | Solution of Linear System of Algebraic Equations | Interpolation & Curve Fitting
Numerical Differentiation & Integration | Numerical Solution of Ordinary Differential Equations
Numerical Solution of Partial Differential Equations