Iterative Methods and Complex Dynamics

Iterative Methods and Complex Dynamics

by

Lily Yau, student, Operations Research & Computer Science, Columbia University ty35@columbia.edu

Adi Ben-Israel, RUTCOR-Rutgers Center for Operations Research, Rutgers University
bisrael@rutcor.rutgers.edu


Research supported by NSF, Research Experience for Undergraduates (REU) Program, at DIMACS (Center for Discrete Mathematics and Theoretical Computer Science), Rutgers University.


How does the Newton method work for complex roots?

Consider the equation

f(z)=0

where f is analytic, and the Newton iteration

z := N(z)

used for solving it. In the real case, the geometric interpretation of the Newton iteration is well known: at a point x_, the next iterate N(x_) is the zero of the line

y = f(x_) + f'(x_)(x-x_)

tangent to the graph of f at (x_,f(x_)).

What is the geometric interpretation of Newton's iteration in the complex case?

For convenience we identify the complex number z = x+iy with the 2-dimensional vector (x,y). Let F(x,y) := |f(x+iy)| be the absolute value of f.

At a point z_ = x_ + i y_, consider: Then the next Newton iterate N(z_) corresponds to the point on the line L which is closest to (x_,y_), i.e. the perpendicular projection of (x_,y_) on the line L.

We prove this in Ref. [2], Theorem 1. See illustration in Fig. 2.


Fig. 1: Level sets of |z^4-1|
Fig. 2: Several Newton iterations

Same for the Halley method

Consider again the equation f(z) = 0, and the Halley iteration

z := H(z)

for its solution. The geometric interpretation of the Halley iteration is given as follows.

Let w = h(z) be the Möbius transformation which is osculating to f at z_, i.e.

h(z_) = f(z_)
h'(z_) = f'(z_)
h''(z_) = f''(z_)

Equivalently, let w = h(z) be the Möbius transformation whose level set {z: |h(z)| = |h(z_)|} is the osculating circle of the level set of |f| at z_.
Then the next Halley iterate H(z_) is the zero of h.

This is proved in Ref. [2], Theorem 2.


Fig. 3: Two Halley iterations, level sets and osculating circles

Same for the quasi-Halley method

The quasi-Halley method is a method of order 2.41, which uses the same information as Newton's method (except for an additional point z- and derivative f'(z-) at the initial iteration), see Ref. [1].

The geometric interpretation of the quasi-Halley iteration
z := Q(z)
is: the new iterate Q(z) is the zero of the Möbius transformation w = h(z) which is tangent ("1.5 order") of f at z, in the sense that

h(z) = f(z)
h'(z) = f'(z)
h''(z) = (f'(z) - f'(z-))/(z - z-)

The level set {z: |h(z)| = |h(z_)|} is no longer the osculating circle of {z: |f(z)| = |f(z_)|}, but approximates it near a root to which the method converges.

The attraction basins of the Newton and Halley methods

Consider a polynomial equation

f(z) = 0

and an iterative method for its solution. Given a root of f, its basin (or domain of attraction) is the set of points from where the iterative method eventually converges to that root.

The global behavior of the Newton method is chaotic: the (common) boundary of the different basins is a Julia set of intricate shape, see Fig. 4.

The level sets of |f(z)| shed some light on the Julia set of the Newton basins, as can be seen in Fig. 5, obtained by superimposing the level sets (Fig. 1) on the Newton basins (Fig. 4).

Fig. 4 Fig. 4: Newton basins for f(z) = z^4-1

Fig. 5Fig. 5: Same with superimposed level sets of |z^4-1|

The global behavior of the Halley method is "less" chaotic, compare Fig. 4 and Fig. 6.

The Julia set is on the boundary of circular "half-moons" (comets?), see Fig. 6 and Fig. 7.

Fig. 6 Fig. 6: Halley basins for f(z) = z^4-1

Fig. 7 Fig. 7: Detail of the Halley comets for f(z) = z^4-1

Fig. 8 Fig. 8: Halley basins for f(z) = z^3-1

Fig. 9 Fig. 9: Halley basins for f(z) = z^3-z

Fig. 10: Superimposing the level sets of f(z) = z^4-1 on its Halley basins.


References

Ref. [1] Adi Ben-Israel, Newton's method with modified functions, Contemporary Mathematics 204(1997), 39-50

Ref. [2] Lily Yau and Adi Ben-Israel, The Newton and Halley methods for complex roots, Amer. Math. Monthly 105(1998), 806-818

Related links


You are the th visitor to our page
since Fri Aug 16 14:25:52 EST 1996
Thank you for visiting us! Please E-mail us at

ty35@columbia.edu or bisrael@rutcor.rutgers.edu

if you have comments, suggestions, or just to say hello.


Pictures will be added from time to time.