Convergence of a Landen transformation

of degree six
Encarnaci
on Guti
errez

Miguel Rosario-Garc
a

University of California

University of Puerto Rico

Los Angeles

Mayag
uez

Mar
a de la Luz Torres
University of California
Riverside
August 3, 2000

Abstract

An integral I (a; b) is said to be invariant with respect to the transformation an+1 = f (an ; bn ); bn+1 = g(an ; bn ) if I (an+1 ; bn+1 ) = I (an ; bn ): A
rational Landen Transformation is a transformation that preserves the integral of a rational function over the positive real line. Boros and Moll have
shown that the integral of a rational function of degree six is invariant under the transformation an+1 = (an bn + 5an + 5bn + 9)=(an + bn + 2)4=3 and
1=3
bn+1 = (an + bn + 6)=(an + bn + 2)
. Furthermore, they show that the sequence (an+1 , bn+1 ) converges to the point (3,3) if a0 ; b0 > 0: In this paper we
show that the region of convergence for (an , bn ) can be extended from the rst
quadrant in the ab-plane to include the area in the second quadrant above the
parabola a2 =3 = b. The Boros-Moll conjecture states that the full region of convergence corresponding to the iteration of the sequence (an , bn ), corresponds
to the region of convergence for the integral I (a; b): This region is bounded by
the curve of the equation 4b3 a2 b2 18ab + 4a3 + 27 = 0.

1 Introduction
The transformation theory of elliptic integrals was initiated by Landen in 1771, by
proving the invariance of the function
G(a; b)

under the transformation

a1

Z =2
0

= a +2 b

d

a2 cos2  + b2 sin2 

and b1

85

ab:

(1.1)
(1.2)

Integrals similar to (1.1) appear in the calculation of the length of an ellipse and in
many other problems of Mathematics.
Gauss [2] rediscovered the invariance of G(a; b) while numerically calculating the
length of lemniscate. The Gauss-Landen transformation (1.2) can be iterated to
produce a double sequence fan; bng that satises 0 < an bn < 2 n. It follows
that fang and fbng converge to a common limit, the so-called arithmetic-geometric
mean of a and b, denoted by AGM (a; b). Passing to the limit in the invariance
G(a; b) = G(an ; bn ) produces

AGM (a; b)

Z =2

d

:
(1.3)
2
a cos  + b sin 
In this form, the elliptic integral G(a; b) can be evaluated as a limit of a sequence
dened recursively.
The existence of a Landen transformation for the integral
Z 1
cx + dx + e
U (a; b; c; d; e) =
dx
(1.4)
x + ax + bx + 1
was established in [1]. It was established that U is invariant under the transformation
0

an+1

an bn + 5an + 5bn + 9

(an + bn + 2) =
bn
= (aan++b bn++2)6 =
n
n
cn + dn + en
cn
= (a + b + 2) =
n
n
(
bn + 3)cn + 2dn + (an + 3)en
dn
=
an + bn + 2
cn + en
en
= (a + b + 2) = :
n
n
Moreover it was shown that if the initial data (a ; b ; c ; d ; e ) 2 R then
lim (an ; bn) = (3; 3)
n!1
4 3

+1

2 3

+1

2 3

+1

+1

1 3

and that there exists a number L such that

lim (cn; dn; en) = (1; 2; 1)L:
n!1
Passing to the limit in the invariance equation
U (an ; bn ; cn ; dn ; en ) = U (a ; b ; c ; d ; e )
yields the identity

U (a ; b ; c ; d ; e ) =  L:
2
86
6

5
+

(1.5)

The proof in [1] is elementary.

The goal of this project is to examine the transformation (1.5) for initial data
(a ; b ) outside the rst quadrant. We have examined conditions on the parameters
(a; b) such that the integral U (a; b; c; d; e) converges. An implicit equation for the
boundary of this region in the (a; b) plane has been obtained. We conjecture that
the iteration (1.5) converges if and only if the initial point (a; b) is inside the region
of convergence of the integral. Numerical evidence that supports this conjecture is
provided. We have also explored the planar dynamical system
+ 5bn + c
(1.6)
an
= anb(na ++5abn +
2) =
n
n
bn
= (aan++b bn++2)6 =
0

+1

4 3

+1

2 3

and examine its long time behavior as a function of c.

An interesting concept is to look at the variable c in (1.6). Varying the term by

dierent values produces extremely dierent graphs. Figure 1a, 1b, 2a, and 2b replaces c by 6.1, 6.45, 8.25, and 9 respectively. Replacing c by 9 yields the graph
representing the boundary curve which all points to the right of the curve will converge to (3,3).

Figure 1: a.

Figure 1b.

2 A decomposition of the iteration

In this short section we just observe that the Landen iteration (1.5) consists of a
nonlinear map from the plane to itself
an+1
bn+1

an bn + 5an + 5bn + 9

(an + bn + 2) =
= (aan++b bn++2)6 =
n
n

4 3

2 3

87

(2.1)

Figure 2: a.

Figure 2b.

coupled with a map from R to R that, given the value of (an ; bn), it is a linear map
in (cn; dn; en): dene n = (an + bn + 2) = , then we have
3

1 3

cn+1
dn+1
en+1

=
=
=

2n (cn + dn + en )

n ((bn + 3)cn + 2dn + (an + 3)en )

3

n (cn + en ):

The consequences of this splitting need to be explored.

3 The convergence of the integral

In this section we discuss the region
Z 1

+ dx + e dx < 1g:

:= f(a; b) 2 R :
(3.1)
x + ax + bx + 1
We begin by characterizing this region in terms of the roots of x +ax +bx +1 = 0.
The method of partial fractions shows that these roots play a crucial role in the
evaluation of this integral. We make this role more precise.
Theorem 3.1 The region
corresponds to the points (a; b) for which the cubic
equation t + at + bt + 1 = 0 does not have a positive real root.
Proof. Suppose that
2

cx4

t3 + at2

+ bt + 1 = (t

t1 )(t

t2 )(t

t3 )

(3.2)

where t ; t ; t are the complex roots of the cubic. The method of partial fractions
now shows that the denominators of the decomposition of
cx + dx + e
x + ax + bx + 1
1

88

are of the form x xi , where xi = t ; t or t . It is now clear that the presence of a

positiveproot ti yields a divergent integral. The term that produces the divergence is
1=(x ti ). The converse can be established by a similar argument. The nature of
the roots ti has to be analyzed, the details are left to the reader.
We now provide a complete description of the region
.
We introduce the notation P (t) = t + at + bt + 1.
p
Lemma 1. The critical points of P are at t = ( a 
a
3b).
0
Proof. Solve the equation P (t) = 0.
Lemma 2. The region f(a; b) 2 R : a < 3bg is contained in
.
Proof. If a < 3b then P (t) is increasing and so it has no positive real roots since
P (0) = 1.
The points (a; b) such that a < 3b corresponds to the case where the roots x are
not real. From now on we assume that these roots are real.
Lemma 3. Suppose (a; b) is such that x < 0. Then P has no positive real roots.
Proof. Simply observe that P is increasing to the right of x .
Lemma 4. Suppose a  3b and x  0. Then (a; b) 2
if and only if P (x ) > 0.
Proof. This is clear.
A direct calculation yields
27P (x ) = (27 + 2a 9ab) 2(a 3b) =
(3.3)
2

1
3

3 2

The identity (3.3) and Lemma 4 shows that the region 27 + 2a 9ab < 0 is outside

. Therefore
consists of the region fa < 3bg plus a part of the region
that
is bounded by the two branches of the function B of the curve b = (27 + 2a )=9a
corresponding to a > 0 and a < 0 respectively.
Lemma 5. The part of
inside
is given by the inequality
R(a; b) := 4a
18ab a b + 4b + 27 > 0
(3.4)
3

1
3

2 2

. This is a simplication of (3.3).

The zero set of the function R(a; b) has two connected components. A convex
curve C outside the rst quadrant and a cusp contained in the rst quadrant. The
boundary of the region
\
is bounded by the branch B of b = (27 + 2a )=9a and
the curve C . We now eliminate the possibility that these two curves intersect.
Lemma 6. The branch B
and the curve C do not intersect. Moreover the curve C
is strictly above B .
Proof. The expression


27
+
2
a
4(a 3) (a + 3a + 9)
R a;
=
(3.5)
9a
729a
Proof

89

has a xed sign.

Lemma 7. The branch P of the parabola f(a; b) : a = 3bg and the curve C do not
intersect either. Moreover the curve C is strictly below P .
Proof. The expression


(a 3) (a + 3a + 9)
(3.6)
R a; a =3 =
27
has a xed sign.
We conclude that the curve C ts between the branch B and the parabola P .
We summarize the previous discussion in a theorem.
Theorem 3.2. The region
Z 1
cx + dx + e

:= f(a; b) 2 R :
dx < 1
(3.7)
x + ax + bx + 1
is described by the points
2

f(a; b) 2 R : R(a; b) := 4a
2

18ab

a2 b2

+ 4b + 27 > 0g:
3

(3.8)

4 The dynamical behavior

In this section we show examples of the long time behavior of the dynamical system
+ 5bn + 9
(4.1)
an
= anb(na ++5abn +
2) =
n
n
bn
= (aan++b bn++2)6 =
+1

4 3

+1

2 3

starting at the initial data (a ; b ). The main result of [1] is the convergence of (2.1)
to the xed point (3; 3) if the initial data is in the rst quadrant. Figure 3 illustrates
this convergence with a = 1:0 and b = 10. The graph contains the rst 50000
points of the iteration.
As one would expect, the graph consists of a few scattered points centered around
and the point (3,3). The small number of points that are visible lead to the conclusion
that the sequence converges rather quickly.
Figure 4 shows the rst 50000 points of the dynamics but with initial conditons
a = 10:0 and b = 10:0.
The behavior of the iteration is now completely dierent. Immediately, one notices
a few distinct features:
 The points are scattered mainly throughout quadrant two and quadrant three,
but they have a tendancy to accumulate on the right hand side of the graph.
0

90

5
4.5
4
3.5
3
2.5
-3

-2

-1

Figure 3: The iteration starting at a = 1:0 and b = 10:0

0

20

10

-50

-40

-30

-20

-10

10
-10

Figure 4: The iteration starting at a = 10:0 and b = 10:0

0

 The graph has a distinct upper boundary which starts in the upper second

quadrant, drops down into the third quadrant, and then begins to level o as
it enters the fourth quadrant. This boundary resembles the graph of a cubic
equation with a negative leading coecient, shifted to the left.
 The graph in gure 3 contains a large number of features that remain unsolved.
The dynamics forms several "lines", the horizontal line in the middle being the
most visible. These questions will be explored in the future.
 The graph also has other interesting features that have not yet been found to
have great signicance when evaluating the integral, but nonetheless will be
addressed later.

91

an

bn

an

-1
10
1.798 3.032
2.978 3.008
2.999 3
3
3
3
3
3
3
3
3
3
3
3
3
3
3

-10
-10
0.190 -2.03835
-7.696 14.551
-3.751 3.004
-4.454 4.521
-4.105 3.738
-4.252 4.061
-4.185 3.912
-4.214 3.978
-4.201 3.948
-4.207 3.961

Table of the rst ten iterations of (a ; b )

for a = 1; b = 1
0

bn

Table of the rst ten iterations of (a ; b )

for a = 10; b = 10

5 The convergence conjecture

In this section we investigate the set of points in the ab plane for which
+ 5bn + 9
an
= anb(na ++5abn +
2) =
n
n
bn
= (aan++b bn++2)6 =
+1

+1

4 3

(5.1)

2 3

converges. It is a standard result of dynamics that, if a sequence like (5.1) converges

to a point P = (a ; b), then P is an attracting xed point of (5.1). This means that
the right hand side of (5.1) reproduces P and the jacobian at P has eigenvalues with
negative real part.
We have evaluated the xed points numerically: these are solutions of the equation
xy + 5x + 5y + 9
(5.2)
(x + y + 2) = = x
x+y+6
(x + y + 2) = = y:
4 3

2 3

We have obtained two real solutions: the important point (3; 3) and the approximate
point P = ( 4:20557; 3:95774). The work in [1] proves that (3; 3) is an attractor and
the linearization of (4.1) at P has eigenvlaues f 0:445734; +7:01992g. This suggests
that the xed point P is a saddle point.
. The curve R(a; b) := 4a 18ab a b + 4b + 27 = 0 is the
stable manifold for the xed point P .
The main conjecture states that every point on the curve R(a; b) = 0 is attracted to
the point P but the dynamics on dierent sides of the curve is completely dierent.
3

Main Conjecture

92

2 2

Those points to the right converge to (3; 3) while the points to the left have the
complicated dynamics indicated in gure 3.
Example. Look at the curve R(a; b) = 4a
18ab a b + 4b + 27 = 0 (Fig. 5) and
3

2 2

25
20
15
10
5

-10

-5

10

-5

Figure 5: Graph of the equation 4a

18ab

a2 b2

+ 4b + 27 = 0
3

x a to any arbritary constant and nd the zeros of b.

i:e:

Letting a = 3 and calculating the zeros of b yields b = 1:62237.

Choosing a = 3 and b = 1:62236 as initial points which are slightly below the
curve yields a divergent graph after several iterations (Fig. 6 on left). However, it is
clearly visible that choosing a = 3 and b = 1:62238 which are slightly above the
curve converges to the point (3; 3) (Fig. 6 on right).
0

20

10

5
4

-50

-40

-30

-20

-10
3
-10
-4

-2

Figure 6: Graphs of the sequence an; bn

93

6 A larger region of convergence

The integral (1.4) will converge only when the denominator does not have a positive
real root. The roots of x + ax + bx + 1 = 0 correspond to the roots of the equation
t + at + bt + 1 if we make the substitution x = t. If we graph this cubic function,
we know that the local minimum, tmin  0. Taking the derivative of this function
and applying the quadratic formula to nd the zeros leads to the solution:
6

3b

a2

3
From the discriminant, we set the2 condition that a 3b < 0. Graphing this inequality,
we obatain the parabola b = a in the ab-plane. We conjecture that if a and b are
chosen so that they lie in the region above the graph of this parabola, the sequence
(an , bn ) will converge to the point (3,3).
In order to prove2 that the sequence will converge for any a and b in Q1 and above
the parabola b = a , we will show that
2

+1

+1

(a 3) + (b 3)  2 n[(a 3) + (b 3) ]
(6.1)
then we will have a geometric convergence to (3,3). We will use the same method
used by Boros and Moll in [1] except we will extend the region of convergence used
in their proof. We will refer the reader to [1] as we will use some of the results of this
proof, and omit some of the details.
Expanding the inequality in (6.1) and moving all terms to one side of the inequality
leaves us with a function f (a; b)  0 that has an absolute minimum of 0 at (3,3). Now
we will restrict our analysis to the region
2

:= f(a; b) 2 R2 : a3  b;

 0g

(6.2)

We introduce the new variables x = (a + b + 2) = and y = ab in order to simplify

the inequality f (a; b), and write h(x; y) for F (a; b). The region
is then transformed
into the region
1 3

2=3

 = f(x; y) 2 R : x  2 = and x  ((3y) = + y3 = + 2) = g

Using our new variables, we need to show that
2

1 3

1 3

1 3

1 3

h(x; y )

=x

10x

14

11

2x + 12x
10

2x (y + 1) + 44x

2x + 4x (3y 11) 20x (y 1) 2(y 1)2  0

for (x; y) 2
 : The following lemma will be useful.
6

Consider the polynomial (6.4).

x  1; k > 0; then h (x ; y (x ))  0:
Lemma 8.

94

Let y (x ) = (x 1)k for

(6.3)
(6.4)

If we expand h(x; y) in terms of x 1 and replace y by y(x) = (x 1)k ,

then it is easy to see that

Proof.

1)k+1 +517(x 1)2 +44(x 1)k+2 +714(x 1)3 +84(x 1)k+3 +119(x 1)4 +128(x 1)k+4 810(x
+5 779(x 1)6 +56(x 1)k+6 +352(x 1)7 +16(x 1)k+7 +1369(x 1)8 +2(x 1)k+8 +1444(x 1)9 +889(x
+354(x 1)11 +91(x 1)12 +14(x 1)13 +(x 1)14

15+156(x
+112(x

1)+28(x

1)k

is positive for x  1. Although there are two terms with negative coecients in
this expansion, it is easy to majorize each of them by a higher-power term so that
h(x; y (x))  0.
Now let us dene the region

< = f(a; b) j (a  0 and b  0) or (a < 0

Lemma 9. Assume (a; b) 2 <, then (a ; b ) 2 <; where
1

a1

 a3 )g

+ 5b + 9
= ab(a++5ab +
2) =

b1

Proof.

and b

(6.5)

4 3

= (aa++b b++2)6 =

2 3

We don't have a proof, but we have checked this for many cases.

Theorem 6.1.

Any point in the region dened in (6.2) converges to the point (3,3).

We have evaluated theorem 6.1 numerically for a large value of numbers for which it
has shown to be true. After a close examination of y (x ) = (x 1)k , we notice that
8 k > 0, h (x ; y (x )) is positive.
The region dened in (6.2) under the transformation x = (a + b + 2) =
and y = ab is mapped onto the region
 and it is contained in the region
f(x; y) j y  0 and x > 1g
and by Lemma 8, in this region h(x; y)  0.

1 3

Proof.

7 Conclusion
According to the Boros-Moll conjecture the region of convergence of the sequence
fan; bng is the same as the region of convergence of the integral U . We now have
an equation of the boundary curve of this region and we can support this conjecture
sequence fan; bng with numerical evidence. The EMM "Theorem" extends the region
of2 convergence of the sequence fan; bng to include the region bounded by the parabola
a < b. However, the EMM Theorem is incomplete and needs to be revised.
6

95

5
10

1)
1)

References
[1]
[2]

Boros, G. - Moll, V.: A rational Landen transformation. The case of degree

six. Contemporary Math., 251, 2000.
Gauss, K. F.: Arithmetische Geometrisches Mittel, 1799. In Werke, 3, 361432. Konigliche Gesellschaft der Wissenschaft, Gottingen. Reprinted by Olms,
Hildescheim, 1981.

96