See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331344026

A modified three-point Secant method with improved rate and characteristics of

convergence

Preprint · February 2019

CITATIONS READS

0 189

1 author:

Ababu Teklemariam Tiruneh

University of Swaziland
27 PUBLICATIONS   75 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

variable chlorine decay rate modelling View project

All content following this page was uploaded by Ababu Teklemariam Tiruneh on 26 February 2019.

The user has requested enhancement of the downloaded file.

 A modified three-point Secant method with improved rate and characteristics of
convergence

Ababu T. Tiruneh1

1
University of Eswatini, Department of Environmental Health Science, P.O. Box 369, Mbabane, Eswatini

Abstract

This paper presents a modification of Secant method for finding roots of equations that uses
three points for iteration instead of just two. The development of the mathematical formula to be
used in the iteration process is provided together with the proof of the rate of convergence which
is 1.84 and is the same as the rate of convergence of Mueller’s method of root finding.
Application examples are given where it is demonstrated that for equations involving ill-
conditioned cases, the proposed method has better convergence characteristics compared to
Newton and Secant methods.

Key words: Root finding, Secant method, Mueller’s method, Newton method, rate of
convergence, iteration, numerical method

1. Introduction

Many problems in science and engineering require solving non-linear equations or systems of
equations requiring trial and error procedure because of the difficulty or impracticality of finding
direct analytical solutions. For example, solutions to problems that require solving polynomial
equations of degree five and higher are proven to be impossible to express in analytical terms
using radicals. Several trial and error methods and their variations have evolved over time in
solving such equations. Newton’s method is one such classical method that requires evaluation
of the function and its derivative to estimate a linear approximation to the root. The method has
convergence of order two near the root. Pathological cases may arise when the iteration
approaches local extremum point where the calculation displays erratic behavior or results in
singularity or divergence away from the desired root. The method also gets slow in equations
that have roots of multiplicities (Gerald and Wheatley, 1994).

1
Variations of Newton’s method are plenty. Examples include a third order method that involves
evaluation of a function and two derivatives (Weeraksoon and Fernando, 2000). Further
improvement (fourth order) has been suggested by Traub requiring the same number of function
and derivative evaluations (Traub, 1982). Similar fourth order convergence is claimed by
Sanchez and Barrero through composite function evaluations (Sanchez and Barrero, 2011).
More recently, sixth and higher order convergence have been achieved by a number of
researchers (Sharma and Guha, 2007; Chun, 2007; Kou and Wang, 2007; Kou, 2007; Kou and
Li, 2007; Kou, et al., 2009; Parhi and Gupta, 2008). The stability of Newton method in all these
improvements may be an issue still relevant to explore though higher order convergence is
undoubtedly a deserved merit of these methods.

The Secant method, also known as Regula falsi or the Method of cords, is another linear
approximation to the root that requires two points and does not require evaluating derivatives.
The order of convergence of Secant method is 1.618 and Secant method may also face
converegence problems similar to Newton method. A new class of Secant like methods have
recently been developed that employ more than one point of the iteration but also include
evaluation of the derivative. Examples of such methods include the methods developed by
Tiruneh et al. (Tiruneh et al., 2013), Fernandez-Torres (2015) and Tukral (2018a b).

Mueller’s method is a quadratic equation approximation to the root generally involving three
points of the iteration and is an extension of the Secant method as it also does not require
evaluating the derivative (Mueller, 1956). Mueller’s method starts with three function
evaluations to begin with but continues with one additonal evaluation as the iteration progresses.
The method of convergence of Mueller’s ,method is of order 1.84. Mueller’s method, however,
can possibly converge to non-real, complex roots unless the function value of one of the points
is opposite in sign to the other two ( Mekwi, 2011). Mueller’s method also faces degeneration
problems if the two points coincide where by the method reduces to the Secant method.

Problems of converegence of the traditional Newton and Secant methods have been tackled
through approaches using a hybrid of methods. Sidi (2008) used a method that involves a multi-
point Secant method whereby an n-degree polynomial is fitted using the previous points of
iteration and Newton method is applied in which the first derivative of the fitted polynomial
replaces the derivatiave of the actual function in the Newton formula. A method that combines
bisection with that of Secant method has been suggested by Dekker (Dekker, 1969). In this
method the function evaluations of bisection and Secant approaches are compared and the new
point resulting in estimate of function value that is closer to the root is chosen for the next trial
and error procedure. Brent (1973) suggested a procedure using root bracketing and inverse
quadratic extrapolation to the root. It is an improvement over Dekker’s method in terms of
improving the rate of convergence. The Leap-frogging method (Kasturiarachi, 2002) uses a

2
hybrid of Newton and Secant methods for iteration obviously resulting in improved order of
converegnce to cubic convergence.

2. Method development

The new proposed method for root finding is an iterative technique that is based on applying
Secant method to the three most recent estimates of the root. Figure 1 shows the starting point of
the iteration involving the first three points. While points (x0, y0) and (x1, y1) may be chosen
arbitrarily to start the iteration, the third starting point (x2, y2) is better determined by application
of the traditional Secant method. Such an approach of estimating the additional starting point has
been proven to improve the iteration process as suggested by Thukral (1918b) in applying the
techniques to the method developed by Tiruneh et al (2013).

Figure 1: The starting point of the new method showing the first three points of the iteration

Referring to the x-y curve shown in Figure 1, the three distinct points are defined to be lying on
the curve with coordinates (x0, y0), (x1, y1) and (x2, y2). The initial point (x0, y0) is taken as the
reference point from which the angles of inclination αi of lines connecting this reference point to
the other two points, namely, (x1, y1) and (x2, y2) are measured. The variable m corresponding to
the tangent of this angle of inclination for any point (xi, yi) lying on the curve is thus defined as:

3
 ( )

Analogously, the m1 and m2 values corresponding to the points 1 and 2 are calculated as:

( ) ( )

Now a new curve (m, y) is drawn with the m values replacing the corresponding x values as the
independent variable. Figure 2 shows this curve.

Figure 2: A new m-y curve drawn using, as the new independent variable, the tangent (m) of
angle of inclination α1 and α2 of the points (x1, y1) and (x2, y2) measured from the reference point
(x0, y0)

Now the regular Secant method is applied to this new curve shown in Figure 2 to determine the
next approximation m3 (and hence by extension x3) to the root;

4
 ( )
( )

Replacing the m3 value with the corresponding x3 value in the Equation (1);

( )

since y3 = 0 as the next point of root approximation as shown in Figure 2;

( )

Reversing the above expression;

( )

( )

( )
( ) ( )

Defining y21 = y2 – y1 and m21 = m2 – m1 finally gives;

( )

5
Continuing the iteration likewise, the k+1th estimate of the root will be;

( )

In terms of the x and y values of the three most recent estimates of the root, the formula in
Equation (3) will eventually be written as;

( )
( )
( )( ) *( ) ( )+

Equation (4) will be used during the iteration process to estimate the next point of iteration from
the previous most successive points of the iteration. Now the proof of order of convergence of
the new method will be given.

2.1 Proof of convergence

Defining the error of estimate of the root r with respect to the ith estimate, xi, as;

Equation (4) can now be written as:

( )
( )
( )( ) *( ) ( )+

For points that are sufficiently close to the root xr, Taylor series expansion can be used to
estimate the y values in terms of the error terms;

( ) ( ) ( ) ( ) ( ) ( ) ( )

( ) ( ) ( ) ( ) ( ) ( )

( ) ( ) ( ) ( ) ( ) ( )

6
 ( ) ( ) ( ) ( )
( ) . …….. (9)

( ) ( ) ( ) ( )

( )
( )

( ) ( )
( ) ( )

Similarly;

( ) ( )

( ) ( )

In all of the Taylor series expansions above, the c terms in the series are given by;

( )

Where yn is the nth derivative of the function y with respect to x evaluated at the root x=r.

Substitution of the expressions given in Eq. (6) to Eq. (12) in Equation (5) above and
neglecting the fourth and higher order terms of the error will give after simplification;

( )
( ) ( ) ( )
( )

The terms in the denominator containing the error terms ek, ek-1 and ek-2 are all small enough to
be neglected compared to the term dominant term c12. This simplification gives;

7
 ( )

Defining positive real terms of the sequence Sk such that;

| | | | | |
| | | | | |

Expressing all the error terms in terms of ek-2;

| | | |

| | | | | |

| | | | | |

Finally, the error ratio below converges to the constant containing the c terms that contain the
derivatives of the y function evaluated at the root x= r;

| | | |

Substituting the equivalent expressions of errors containing the ek-2 term only;

| | | | | |

As the iteration approaches the root, the above ratio approaches the constant on the right hand
side of Equation containing the c terms. For this to be true unconditionally, the power of the
error term should approach zero, i.e.,

( )

The roots of the above third degree polynomial in Equation (13) are;

8
The order of convergence of the proposed method is therefore 1.839 which is similar to the rate
of convergence of Mueller’s method of root finding. However, the proposed method always
converges to the real root which is not necessarily the case with Mueller’s method which may
converge to imaginary roots unless one of the points has its y value that is opposite in sign to the
y values of the other two points. Can we say also that we have in the process discovered a family
member of Mueller’s method that is having linear forms?

The proposed method has, therefore, better order of convergence compared to the regular Secant
method which has order of convergence of approximately 1.68. It will now be shown in the
following section that the proposed method, in addition, displays better convergence
characteristics for ill-conditioned cases in which either Newton or Secant method (and in some
cases both) may fail to converge to the root.

3. Application Examples

Examples of application of the proposed three point Secant method for locating roots of several
equations are discussed below. The equations tested included those that regularly converge to the
roots; equations that have roots of multiplicities in which the rate of convergence is slowed down;
and equations that display pathological behavior during iteration whereby application of Newton
or Secant method may fail to reach convergence.

For the purpose of determining the number of iterations, a stopping criterion is used which uses
the following rule:

| | | | ( )

The rate of convergence of the methods to the roots is calculated with the following formula:

(| |) (| |)
( )
(| |) (| |)

In Equation (15), αk is the order of convergence of the iterative process at the kth iteration step,
ek and ek+1 are the errors of estimate of the root at the kth and k+1th iteration steps respectively
and r is the desired root of the equation.

9
3.1 Test of the proposed method for regular cases

Table 1 shows summary of the results of the iteration processes among the three methods for
equations in which the iterations regularly convergence to the root. Examination of the rate of
convergence of the proposed method for regular cases show that the method generally converges
at a rate of 1.84 as the iteration approaches the root value as predicted theoretically, a rate that is
between Secant and Newton methods. For the equation having roots of multiplicity, namely,

( )( )

all the three methods display slow and almost linear convergence requiring greater number of
iterations with Newton method relatively faster than the other two methods followed by the
proposed three point Secant method. This order of rate of convergence is also similar to the order
for the regular cases except for the greater number of iterations required.

It is also interesting to note that for another function that has root of multiplicity, namely,

( )

All the three methods show regular rate of convergence rate as the iteration approaches
convergence near the root x= 2. However, depending on the starting point, the total number of
iterations required varies differently among the different methods. The geometric mean of the
rate of convergence is higher (1.64) for the proposed three point Secant method compared with
Newton method (1.20) and Secant method (1.14). The total number of iterations required is also
proportional to this geometric mean of convergence. The three point Secant method in this case
displays superior overall rate of convergence.

In a similar pattern, for the equation:

The geometric mean of rate of convergence for the three point Secant method is the highest
(1.24) compared to Newton method (1.18) and Secant method (1.06). The total number of
iterations is proportional to this geometric mean with the proposed three point Secant method
requiring marginally lower number of iterations compared to Newton and Secant methods. As

10
Table 1 shows, for this particular equation, the total number of iterations required is 26 for the
three point Secant method compared to 28 for Newton method and 39 for Secant method.

Figure 3 shows comparison of the number of iterations required for finding roots of several
equations among the different methods where by the proposed three point Secant method shows
convergence rate between Secant and Newton methods and better overall rate of convergence for
cases displaying characteristics similar to equations having roots of multiplicities as the
examples above have shown.

Table 1. Comparison of result of iterations of the Three point Secant method with Newton
and traditional Secant Methods.

Number of iterations required

Function Root Starting points Secant Newton Three point
Method Method Secant
method
0.5, 0.55, 0.6 10 8 9
1.365230013414100
0.9, 0.95, 1.0 10 6 8
-1.0, -0.975, -0.95 10 7 8
[ ( )] -1.404491648215340
-3.5, -3.25, -3.0 10 7 9
-2.0000000000000 -3.1, -3.05, -3.0 169 117 121
( )( )
2.0000000000000 1.4, 1.45, 1.5 117 32 87
1.5, 1.55, 1.6 25 17 9
( ) 2.00000000000000 2.5, 2.55, 2.6 12 8 8
3.5, 3.55, 3.6 15 11 11
-0.9, -0.85, -0.8 10 7 7
( ) ( ) -0.603231971557215
-0.7, -0.65, -0.6 8 5 5
4.0, 4.05, 4.1 29 20 20
3.000000000000000
4.4, 4.45, 4.5 39 28 26
2.0, 2.05, 2.1 9 5 7
( ) 1.857183860207840
0.4, 0.45, 0.5 11 8 8

11
Figure 3: Comparison of number of iterations required for convergence among the different
methods

3.2 Test of the proposed method for divergent cases

The proposed new method has been examined for its convergence for pathological cases
whereby the iteration process displays divergence or failure when applying either the Newton or
Secant methods. Table 2 summarizes the results of the iteration for these pathological cases in
which comparison is method between the proposed three point Secant method and the traditional
Newton and Secant methods. The Table shows that, whereas either Newton or Secant method
(and in some cases both) fail to converge to the root; the proposed method almost always
converges to the root. This result shows that the proposed three point Secant methods has a
relative advantage for ill-conditioned cases where application of either Newton or Secant or both
methods may not lead to convergence.

Figures 4-6 show the characteristics of convergence for the three different methods when applied
to the equation:

Where the root x=0 is a straightforward solution. It is seen in Figure 4 that applying Newton
method results in oscillatory divergence to infinity. Similarly, application of Secant method to
the same equation leads to oscillation between finite points in which the iteration fails to

12
converge to the root as shown in Figure 5. By contrast the proposed three point Secant method
shows oscillatory convergence towards the root as Figure 6 shows.

Figure 4: Divergence of iteration by Newton method for the function y = x1/3

Figure 5: Oscillation of the iteration by Secant method for the function y = x1/3

13
 Figure 6: Oscillating convergence of the iteration by three point secant method for y= x 1/3

Another example of the advantage of the proposed method for ill-conditioned cases is the five
degree polynomial equation:

Application of Newton method results in oscillation where by the iteration fails to converge to
the root as shown graphically in Figure 7. In a similar pattern, application of the regular Secant
method results in a gradually diverging oscillation where the method fails to converge to the root
as shown in Figure 8. By contrast, application of the proposed three point Secant methods shows
convergence with a rate of 1.84 as proven theoretically near the root and with an overall rate of
convergence (geometric mean) of 1.15 for the starting points given in Table 2. The process of
convergence for this polynomial equation as the iteration proceeds is shown graphically in
Figure 9.

Figure 7: Constant oscillation of the iteration by Newton method for y= x5- x +1

14
 Figure 8: Oscillating divergence of the iteration by Secant method for y= x5- x +1

Figure 9: Oscillating convergence of the iteration by three point Secant method for y= x5- x +1

4. Conclusion

A mathematical formula involving iteration for finding roots of non-linear equations has been
developed and presented in this paper. The proposed formula is a modification of the Secant
method employing three successive points of the iteration instead of just two. The method
applies the Secant method to the tangent of the relative angle of inclination of the lines joining
the furthest point with the two most recent points of iteration. The rate of convergence of the
proposed method is of the order 1.83929 which is equivalent to the rate of convergence of
Mueller’s method of root finding. In fact it can be said that this method is a discovery of a
variant or a family of Mueller’s method that is linear in form as it is recalled that Mueller’s
method is also an iterative procedure that is based on three successive points of the iteration.

15
Table 2. Results of application of the proposed three point Secant method for cases in which
either Newton or Secant method (or both) fail to converge.

Comparison of number of iterations required

Three
Function Root point
Secant Newton Secant
Starting points Method Method method

1, 1.5, 2.0 11 Oscillates 10

1.887207676120680
0.5, 0.55, 0.6 24 Oscillates 23
( ) 1.000000000000000 3.0, 3.25, 3.5 Fails Fails 9
3.0, 3.25, 3.5 Diverges Diverges 9
( ) 0.0000000000000
-3.0, -3.25, -3.5 Diverges Diverges 10
2.0, 2.5, 3.0 Oscillates Oscillates 27
-1.167303978261420
7.0, 7.5, 8.0 112 Oscillates 44
1.7639320225002100 2.0, 2.5, 3.0 10 Oscillates 9
6.236067977499790 5.0, 5.5, 6.0 10 Oscillates 11
1.0, 1.25, 1.5 Oscillates Diverges 95
0.00000000000000
-1.0, -1.25, -1.5 Oscillates Diverges 93
1.679630610428450 3.0, 3.25, 3.5 Diverges Diverges 14
0.101025848315685 -1.0, -1.5, -2.0 Diverges Diverges 19

However, the proposed method is linear in nature being based on Secant method of root finding
and as such does not lead to imaginary roots unlike Mueller’s method which is based on
quadratic solution to the root approximation. Stated in another way, the method does not require
that the function values of one of the iteration points be opposite in sign to the others in order to
avoid imaginary root values during the iteration.

Examples of application of the proposed method of root finding to a variety of equations has
been presented and compared with the results of iterations of Newton and Secant methods. For
equations leading to regular convergence, it is shown that the proposed method has rate of
convergence that lies between Secant and Newton methods as is also supported by the
mathematical proof of the rate of convergence presented in this paper. For equations that display
roots of multiplicities or characteristics similar to roots of multiplicities, the proposed method
generally displays better overall rate of convergence compared to Secant and Newton methods.

For ill-conditioned cases in which Newton and Secant methods may fail to converge displaying
oscillation, divergence to infinity or off-shooting to undesirable or invalid domain, the proposed
three point Secant method almost always leads to convergence as is demonstrated by a number

16
of examples presented in this paper. This is the inherent advantage of the proposed method over
the traditional Newton and Secant methods that display pathological behavior for ill-conditioned
cases. The application examples demonstrate that the proposed method has better convergence
characteristics for such ill-conditioned cases compared to Newton and Secant methods.

References

1 Gerald, C.F. and Wheatley, P.O. (1994) Applied numerical analysis. Fifth Edition, 40-75.

2 Weerakoon, S. and Fernando, T.G.I. (2000) A variant of Newton’s method with

accelerated third order convergence, Applied Mathematics Letters, 13, 87-93.

3 Traub, J.F. (1982) Iterative methods for the solution of equations, Prentice-Hall,
Englewood.

4 Sanchez, M.G. and Barrero, J.L.D. (2011) A Technique to composite a modified

Newton's method for solving nonlinear Equations, arXiv: 1106.0996v1, Cornell
University Library.

5 Sharma, J.R. and Guha, R.K. (2007) A family of modified Ostrowski methods with
accelerated sixth order convergence, Appl. Math. Comput. 190, 111-115

6 Chun, C. (2007) Some improvements of Jarratt’s method with sixth-order convergence,

Appl. Math. Comput. 190, 1432- 1437.

7 Kou, J. and Wang, X (2007) Sixth-order variants of Chebyshev- Halley methods for
solving non-linear equations, Appl. Math. Comput. , 190, 1839-1843.

8 Kou, J. (2007) On Chebyshev-Halley methods with sixth-order convergence for solving

non-linear equations, Appl. Math. Comput., 190, 126-131.

9 Kou J., and Li, Y. (2007) An improvement of Jarratt method, Appl. Math. Comput. 189,
1816-1821.

10 Kou, J. Li, Y.and Wang, X. (2009) Some modifications of Newton’s method with fifth
order convergence, Journal of Computational and Applied Mathematics, 209, 146-152.

11 Parhi, S.K. and Gupta, D.K. (2008) A sixth order method for nonlinear equations,
Applied Mathematics and Computation 203, 50-55.

17
 12 David E. Muller, (1956) A Method for Solving Algebraic Equations Using an Automatic
Computer, Mathematical Tables and Other Aids to Computation, 10, 208-215.

13 Mekwi, W.R. (2001) Iterative methods for roots of polynomials, University of Oxford.
M.Sc. thesis,

14 Dekker, T.J. (1969) Finding a zero by means of successive linear interpolation, In B.

Dejon and P.Henrici (eds), Constructive Aspects of the Fundamental Theorem of
Algebra, Wiley-Interscience, London, SBN 471-28300-9.

15 Brent, R.P. (1973) Algorithms for Minimization without Derivatives, Chapter 4. Prentice-
Hall, Englewood Cliffs, NJ. ISBN 0-13-022335-2.

16 Kasturiarachi, A. B. (2002) A Leapfrogging Newton's method, International Journal of

Mathematical Education in Science and Technology, 33(4), 521-527.

17 Sidi, A. (2008). Generalization of the secant method for nonlinear equation. Applied
Mathematics E-Notes, 8, 115-123.

18 Tiruneh, A.T., Ndlela W. N. and Nkambule S. J. (2013) A two-point Newton method

suitable for non-convergent cases and with super-quadratic convergence, Adv. Numer.
Anal. art. ID687382.

19 Fernandez-Torres, G.A (2015). A novel geometric modification of the Newton-secant

method to achieve convergence of order 1+ √ and its dynamics, Mod. Sim. Eng., art. ID
502854.

20 Thukral, R. (2018a). A New Secant-type method for solving nonlinear equations.

American Journal of Computational and Applied Mathematics, 8(2): 32-36.

21 Thukral, R. (2018b) Further development of secant-type methods for solving nonlinear

equations. International Journal of Advances in Mathematics, 38(5), 45-53.

18

View publication stats