Newton 1
Newton 1
2
two quadratic equations in C
viewed as a global dynamical system
John H. Hubbard and Peter Papadopol
In this paper, we will study Newton’s method for solving two simultaneous quadratic equations
in two variables.
In one dimension, if F is a polynomial, the Newton mapping is a rational function and we can
apply the now rather well developed theory of one-dimensional complex analytic dynamics. The
subject is far from completely understood, but much progress has been made, particularly by J.
Head, Tan Lei, and M. Shishikura. More recently, [HSS] give precise results on how to find all the
roots of a polynomial, based on the topology and complex analysis of the basins.
Presumably, there is no need to motivate a study of Newton’s method, in one or several variables.
The algorithm is of immense importance, and understanding its behavior is of obvious interest. It
is perhaps harder to motivate the case of two simultaneous quadratic equations in two variables,
but this is the simplest non-degenerate case.
The classical theory, due to Ostrowski and Kantorovitch, asserts that a non-degenerate zero of
a C 2 function f : Rn → Rn is a superattractive fixed point of the associated Newton map
Nf (x) = x − [Df (x)]−1 f (x).
For an elementary presentation of this theory, see [HBH], Section 2.7; we will follow the notation
used there. Shub and Smale [ShSm] have also made an extensive study of Newton’s method from
the point of view of complexity theory.
When f : C2 → C2 is analytic, a much more detailed description of the dynamics near a fixed
point is given in [HP], which gives a partial description of the basins of the roots. In that paper,
we show that associated to each root x0 there is a rational function gx0 : P1 → P1 constructed from
the quadratic terms of the Newton map at the root. This rational function (its Julia set, critical
orbits, . . . ) is the most important element in the description of the basin, and in the present paper
it will also play an important role.
What makes Nf such an intractable dynamical system is the simultaneous presence of critical
points and points of indeterminacy. There is an extensive theory for the dynamics of birational
maps P2 ∼∼> P2 [BS1-8],[FS1-3], [HO1-2],[Dil], [Br], [Jo], though in no sense do they give the sort
of information available for functions of one complex variable. Of particular relevance to this paper
is [HPV], where the technique of real oriented blow-ups is introduced. These play an important role
here too; in [HPV], we showed something of what complications can lurk in points of indeterminacy
(solenoids, etc.), but here the same sort of construction yields for more elaborate structures.
There is also a reasonable theory of endomorphisms of Pn [HP],[FS], [Ueda], though again the
descriptions provided are not nearly so precise as one might like, and there is a drastic shortage of
examples that are actually understood.
But next to nothing is known about mappings that have topological degree > 1 (and in par-
ticular, critical curves), and points of indeterminacy. One important result is due to Russakovskii
1
2 Hubbard and Papadopol
and Shiffman [RS]. There they prove that if f : Pn ∼∼> Pn is a rational mapping, and various
inequalities on the algebraic degree and topological degree are satisfied, then if you take a “generic”
point, take all its inverse images by f n , and give them all equal weight so as to form a probability
measure µn , then these measures converge as n → ∞ to a measure µ independent of the original
point. Here generic means “chosen off a pluripolar set.”
Newton’s method for solving two simultaneous quadratics does satisfy their conditions; in fact,
Russakovskii and Shiffman consider specifically the Newton map associated to
f (x, y) = (y − x2 , x − y 2 ).
This map, like all the Newton maps Nf associated to pairs of quadratics, has two points of indeter-
minacy at infinity. Russakovskii and Shiffman could not determine whether their measure charges
the points of indeterminacy at infinity; in this paper we prove that it does not (Corollary 3.3.3).
Chapter 2 is concerned with stable and unstable manifolds of invariant subsets. One important
result is Theorem 2.4.1. As stated above, each line joining a pair of roots is invariant, and the
restriction of Newton’s method to such a line has a Julia set, which is the real straight line
bissecting the two roots in the complex invariant line. The question whether this Julia set has
a stable manifold is quite subtle. It never has a genuine stable manifold, but sometimes there is
a 3-dimensional manifold associated to it which has some features of a stable manifold and some
features of a center manifold; this is the content of Theorem 2.4.1. Under more general conditions,
it is an ergodic attactor [PS], and has a measure-theoretic stable manifold, the topology of which
is still pretty mysterious.
In Chapter 3, we examine a case where the algebra simplifies: the intersection of the parabolas
y = x2 , x = y 2 . For this map, we will analyze the behavior at infinity, using infinitely many blow-
ups. One pay-off is that we are able to untangle the regions at infinity where Newton’s method
Newton in C2 3
is attracting and repelling, and show that the line at infinity is not charged by the Shiffman-
Russakovskii measure.
In Chapter 4, we analyze in the abstract the process of infinitely many blow-ups used in Chapter
3. This construction, which we call the Farey blow-up, is fundamental if one wants to understand
the behavior of mappings with critical points and points of indeterminacy. In some sense, the
idea of constructing this blow-up goes back to Max Noether, whose “monster” has intrigued many
mathematicians [Ma], [De], without ever really becoming an accepted citizen of algebraic geometry,
because the topology of the monster is so wild.
In contrast, we will show that the topology of Farey blow-ups is fairly tame, and quite easy to
describe in the language of real oriented blow-ups, as pioneered in [HPV]. In a rather surprising
coincidence, the topology of these real oriented blow-ups is reminiscent of the topology of non-
integrable Hamiltonian systems with two degrees of freedom, with zones corresponding to rational
“frequencies” separated by tori corresponding to irrational frequencies.
We also compute the second homology groups of Farey blow-ups. This is a space with a natural
quadratic form coming from the intersection product. This quadratic form is negative definite,
and the completion of the homology with respect to the quadratic form turns out to be a very
classical Sobolev space of functions with one derivative in L2 . We also compute the operation of
appropriate mappings on the homology.
In Chapter 5, we perform infinitely many Farey blow-ups to construct a compact space X∞ on
which the Newton map operates, and analyze in detail the structure at infinity.
The real oriented blow-up of X∞ has a 3-dimensional boundary, and now the resemblance with
the structures appearing in the KAM theorem [Ar] is uncanny. “First-order tori” corresponding
to irrational numbers, and carrying irrational foliations, separate zones corresponding to rational
numbers. Within these zones, further “second-order tori” again with irrational foliations, sepa-
rate zones corresponding to rationals, etc., the whole thing accumulating on some Cantor set of
solenoids.
Another way of thinking of this structure is in terms of Puiseux expansions, with the first-order
tori corresponding to certain one-term Puiseux expansions of transcendental curves, separating
zones corresponding to Puiseux exponents of algebraic curves. Within these zones, there are further
structures corresponding to two-term Puiseux expansions, some transcendental corresponding to
tori, and some algebraic, which can then be further enriched, etc. We do this in the case of the
same special Newton map as in Chapter 3, to simplify the computations.
In Chapter 6, we show that essentially all the results of Chapters 3 and 5 generalize to any
Newton map associated to a pair of quadratic equations in two variables (except some degenerate
cases).
Acknowledgments
We thank Jeff Diller, who helped with the computation of Section 2.1. Robert Strichartz saw
the connection of the homology with the Sobolev space H1 . Eric Bedford made many helpful
comments about an early version of the paper. Karl Papadantonakis contributed the computer
program which made the color pictures. Ramin Farzaneh wrote an earlier program, which raised
many of the problems we address in this paper. A big thanks to Barbara Burke Hubbard, for her
help with editing, and finding several mathematical mistakes.
4 Hubbard and Papadopol
Peter Papadopol thanks the Department of Mathematics of Cornell University, and the chair-
men Keith Dennis, Peter Kahn, Robert Connelly and especially John Smillie, for many years of
hospitality. From Grand Canyon University, he thanks former President Bill Williams for his
support for many years, President Gil Stafford and Chairperson Beth Dawkins for support and
understanding.
µ ¶
Figure 1. At left, we see the basins of the roots, marked with black dots: the basin of 00 in red,
µ ¶ µ ¶ µ ¶
the basin of 1 0 in green, the basin of 1 in blue, the basin of 2 in grey. The region displayed is
0 3
−4 ≤ x, y ≤ 4. Notice that the x-axis is entirely red and green, and the y-axis is entirely red and blue, in
fact all the lines joining pairs of roots contain only two colors. At right is a blow-up of the figure at left;
notice the blue fingers coming from above and below to touch the x-axis.
The main thing we see in these pictures is the importance of the polar locus, i.e., the preimage
of
³ the
´ line at infinity. In Figure 1, it is the hyperbola, one branch of which bounds the basin of
2 , drawn in grey in the picture. The other branch is a bit harder to see; one part bounds the
3
Newton in C2 5
basin of 1
³ ´
0 . The inverse image of this polar hyperbola is a curve of degree 6, which has three
branches in this picture. It is a bit difficult to see, since two of its branches are almost superposed
to the hyperbola.
Figure 2 represents
³ ´ ³ Newton’s
´ ³ ´ method ³ applied to two quadratic equations, in the case when the
0 1 0 −1
´
roots are at 0 , 0 , 1 , and −1 . The polar curve is the ellipse one sees emphasized
in the middle. More generally, if two real conics intersect in four real points that form a convex
quadrilateral, then the polar locus is a hyperbola, but if one intersection is in the convex hull of
the others, the polar locus is an ellipse. This time the real part of its inverse image is the 3-leafed
clover, a bit more lightly emphasized. Actually, this is only one branch of the inverse image; there
is another branch in C2 .
µ ¶ µ ¶ µ ¶ µ ¶
Figure 2. Left: the basins of the roots 0 (red basin), 1 (green), 0 (blue), and −1 (grey).
0 0 1 −1
The polar curve is an ellipse in this case, and it together with its first and second inverse images are drawn
on the right. They appear to form the boundary of the basins of the roots, or at least it seems that the
closure of the union of the iterated inverse images of the polar curve does.
It is not really surprising that the inverse images of the line at infinity appear to make up the
boundaries of the basins of the roots. After all, the line at infinity itself is never in the basin of
any root, so neither are its inverse images.
Since our objective is to understand the dynamics of Newton’s method in C2 , we need to find
a way of making complex pictures. We will also need to find the complex analog of the line at
infinity, and its successive inverse images. Of course, the line at infinity is still there, and so are
its inverse images, but they are now of codimension 2, and cannot form the boundary of anything.
2
The most obvious picture to make is a 1-dimensional complex ³ ´slice ´ ³ ´ C , in
³ through ³ which
´ we
0 1 0
draw the basins. If we use the same Newton map with roots at 0 , 0 , 1 , and −1 , and −1
slice by the line y = (1 + i)x, then the region 0 ≤ Re x ≤ 2, | Im x| < 1 is represented on the left
of Figure 3.
6 Hubbard and Papadopol
Figure 3. The same Newton’s map shown in Figure 2, sliced by the complex line y = (1 + i)x. The
different shades represent the basins of the four roots. On the right, a blow-up of the figure on the left.
What do we see? The basins of the roots are bounded by what seem to be smooth curves.
Inside some of the basins there seem to be archipelagoes of islands of a different basin, which
appear to converge to a point of a boundary with a third basin, where however the boundary
appears perfectly smooth (see Figure 4).
Figure 4. A sequence of “bubbles” of one domain inside another basin, converging to a boundary point
of a third basin. The left hand figure is the “top bubble” of the right-hand picture in Figure 3, and the
right-hand picture is a further blow-up of that. The reader might notice some similarity between the blue
“fingers” of Figure 1 and the blue bubbles in this figure; we will explore this similarity in Section 2.5.
Newton in C2 7
In other places, the picture appears to be self similar, but the curves forming the boundaries
undulate, and these undulations do not appear to become smaller as they get closer to the center
of self-similarity. These curves form a definite angle at the point of self-similarity.
Figure 5. A blow-up near the center of the left-hand side of Figure 3, and a further blow-up of it.
The pictures seen so far are misleadingly simple. As we will see in Chapter 2, things tend to
be a bit simpler when the coefficients of a Newton map are real. Figure 6 represents the line of
equation
³ ´ ³ ´y = ³ sx.
´ We³see´ the basins of the roots for Newton’s method, where the roots are at
0 , 1 , 0 and α , where α = 2 + 2i and β = −.3. Again, we understand roughly why
0 0 1 β
the drawing presents the sort of complication present, without understanding in detail just why
the structure is what one sees.
Some of the phenomena above will be explained in this paper: the apparent smoothness of
basin-boundaries, the self-similarity, the undulation of some boundary curves. There are many
things about these pictures which we don’t understand yet, more particularly the combinatorial
properties of all these pictures.
A question we have not been able to solve is: are the basins dense? This hides two quite different
questions:
– Are there any attracting cycles other than the basins?
– Are there wandering domains?
We conjecture that the answer to both these questions is no. But the questions are not of equal
importance: for the Newton map as applied to polynomials of degree greater than 2, there certainly
are other attracting cycles. But we conjecture that there are never any wandering domains for
Newton’s method as applied to pairs of polynomials in two variables. This stands out as a central
problem of dynamics in several variables: the tools used by Sullivan in the proof in one dimension
are simply not available.
Remark. The figures shown here are chosen essentially at random; practically any other param-
eters cut by any line are equally as interesting and complicated. A much larger gallery of pictures
can be seen at web site
http://www.math.cornell.edu/ dynamics/
The site also includes Macintosh programs that the reader can download. △
1
Generalities about Newton’s method
Given two vector spaces V and W of the same dimension, and a mapping F : V → W , the
associated Newton map NF : V → V is given by the formula
NF (x) = x − [Df (x)]−1 (F (x)). (1.1)
A well-known theory asserts that if x0 ∈ V satisfies F (x0 ) = 0 and if the derivative [DF (x0 )] is
invertible, then x0 is an attracting fixed point of the Newton map, and the rate of attraction is at
least quadratic.
9
10 Hubbard and Papadopol
Off the polar locus, the Newton map is differentiable; the derivative is a bit delicate to compute.
In one dimension, the derivative of the Newton map is given by
f (x)f ′′ (x)
(Nf )′ (x) = . (1.8)
f ′ (x)2
We want to extend this computation to several dimensions. This requires a bit of notation to
decide what the second derivative is. We will write the Taylor expansion of F in a neighborhood
of x as
1
F (x + u) = F (x) + [DF (x)](u) + [D2 F (x)](u, u) + . . . , (1.9)
2
so that
1 k
[D F (x)]F : V × · · · × V → W (1.10)
k!
denotes the symmetric k-linear mapping whose restriction to the diagonal is the terms of degree k
of the expansion.
Lemma 1.1.2. The derivative of the Newton map is given by the formula
³ ´
[DNF (x)](v) = [DF (x)]−1 [D2 F (x)] [DF (x)]−1 F (x), v . (1.11)
This formula reduces to the simple one above in dimension 1 (Equation (1.8)).
Proof. If M : X → Hom (V, W ) is a differentiable matrix-valued function, then
³ ´
[DM −1 (x)](y) = (M (x)−1 ) [DM (x)](y) (M (x)−1 ), (1.12)
where the middle term is naturally a linear transformation in Hom (V, W ), so that the composition
makes sense. Using this formula, we compute
³ ´
[DNF (x)](v) = v − [DF (x)]−1 [DF (x)](v) − [DF (x)]−1 [D2 F (x)] [DF (x)]−1 (F (x)), v , (1.13)
and the first two terms on the left-hand side cancel, to give the desired result. ¤
Of course, since roots of F are superattractive for Newton’s method, the derivative vanishes
there, as is clear from Lemma 1.1.2. It is often useful to know the quadratic terms also.
In Equation (1.15), the constant term is x0 , the linear terms cancel, and the quadratic terms give
µ ¶
1 ³ ´
−1 + [DF (x0 )]−1 [D2 F (x0 )](u, u) , (1.16)
2
which gives the desired result. ¤
Lemma 1.1.4. If the point x is not a critical point of F , then it is a critical point of the Newton
map NF if and only if the Newton vector [DF (x)]−1 F (x) points in a critical direction of the
quadratic function
u 7→ [D2 F (x)](u, u). (1.19)
Proof. This follows from Equation (1.14); the composition by [DF (x)]−1 on the left does not
affect which vectors are mapped to zero. ¤
Lemma 1.2.1. (a) The space of quadratic functions vanishing at four distinct points is 2-dimen-
sional, and GL2 acts transitively (by composition on the left) on pairs of such functions that are
linearly independent.
(b) Among these functions exactly two define parabolas (possibly degenerate), and exactly three
define pairs of lines.
From this and Lemma 1.1.1, we immediately see:
Corollary
in 1.2.2. Newton’s method to find the intersection of two conics depends only on the
intersection points and not on the choice of curves.
12 Hubbard and Papadopol
The family of curves given by a 2-dimensional vector space of functions is a 1-dimensional family
of curves, called a pencil of curves. Lemma 1.2.1 says that the conics passing through four distinct
points form a pencil. More generally, the curves of degree d passing through (d2 + 3d − 2)/2 points
form a pencil. When d = 3, we have (d2 +3d−2)/2 = 8, and given any eight points in C2 in general
position, we can consider the Newton map defined by any two cubics passing by these points. This
map is independent of the choice of cubics in the pencil. A well-known theorem asserts that if
three cubics are concurrent in eight points, then they are also concurrent in a ninth; the argument
above gives another proof of this fact. We can also see that the Newton maps applied to find the
intersections of two curves of degree d form a family of dimension d2 + 3d − 8.
Proof of Lemma 1.2.1. ³ By an ´affine
³ ´change
³ ´ of variables in the domain, the four points can
be moved, for instance to 0 , 0 , 1 , α
0 1 0
´ ³
β ; so up to conjugation by an affine mapping the
corresponding Newton maps depend on the two parameters α, β. We will find it convenient to set
1−α 1−β
A= and B = (1.20)
β α
as these quantities arise frequently in computation.
The general quadratic function vanishing at 00 , 10 , 01 , α
³ ´ ³ ´ ³ ´ ³ ´
β is
P (x2 − x) + Qxy + R(y 2 − y), P, Q, R ∈ C (1.21)
where Q = P A + RB. We can take as our equations any two linearly independent functions in this
family.
One possibility is to require that P x2 + Qxy + Ry 2 be a perfect square: this will happen when
Q2 = (P A + RB)2 = 4P R. (1.22)
Up to multiples, there are two such equations in general, defining parabolas.
We could also choose P = 1, R = 0, and P = 0, R = 1, to find
³ ´ ³ 2
F x x + Axy − x
´
y = 2
y + Bxy − y (1.23)
Most of this paper is devoted to understanding the dynamics of the Newton mapping given
by Equation (1.24) in terms of the parameters α and β. Sometimes, instead of studying that
mapping, we will switch to its conjugate (given by Equation (1.34)), which is sometimes simpler.
This conjugate is obtained by intersecting parabolas.
The following result is essential to many of our constructions.
Lemma 1.2.3. The lines joining the roots are invariant under Newton’s method, and in these lines
Newton’s method induces the 1-dimensional Newton’s method to find the roots of a quadratic
polynomial.
Newton in C2 13
Proof. By Lemma 1.1.1, it is enough to prove the invariance of the x-axis in the above normal-
ization. The Newton mapping becomes
1 x(Bx2 − x) = x2 /(2x − 1) ,
NF x
³ ´ ³ ´ ³ ´
0 = 0 0 (1.26)
2Bx2 − (2 + B)x + 1
The first coordinate is Newton’s method for the polynomial x2 − x with roots at 0 and 1. ¤
Lemma 1.2.3 explains why the lines joining pairs of roots in Figure 1 are colored in only two
colors.
Remark. Lemma 1.2.3 is a special case of the following more general result: if two curves of
degree d in the plane have d of their intersection points on a line, then this line is invariant under
the Newton map, and the restriction to the line is the Newton map for solving the one-variable
polynomial with its roots at those points.
Of course, if d > 2, it is exceptional for two curves of degree d to have d of their intersection
points aligned. △
We will refer to the lines joining pairs of roots as the invariant lines of the Newton map NF .
Their intersections also play an important role: they are points of indeterminacy of NF . This
should be clear: if the Newton map were defined there, the orbit would need to be in both lines,
so the point of intersection would have to be fixed, and it isn’t.
We will label these points
µ ¶
³
1/B
´ ³
0
´ α/(α + β)
q1 = 0 , q2 = 1/A , q3 = . (1.27)
β/(α + β)
In Figure 1, these points are the intersections of pairs of lines through disjoint pairs of roots. For
instance, the point q1 is the point on the x-axis at x = −1.
As points of indeterminacy, these points are fairly innocent. A standard result from algebraic
geometry says that NF will be well defined after an appropriate sequence of blow-ups in the domain;
in this case, it is enough to blow up once.
Lemma 1.2.4. The Newton map extends algebraically to the exceptional divisors of C2 blown up
at the points q1 , q2 , q3 , and the images of the exceptional divisors are straight lines.
We will see later just which straight lines the images of the exceptional lines are.
Proof. It is enough to prove this at q1 . We need to show that the limit
m 7→ lim NF 1/B +t
³ ´
mt (1.28)
t→0
exists. This is a straightforward computation using Equation (1.24); the constant terms in t vanish,
and after canceling a factor of t in the numerator and denominator of both coordinates, we find
1 2m Am
B + B2 − B
4
lim NF 1/B + t = 2−B+m( B −2−A) .
³ ´
mt m
(1.29)
t→0 B −m
4
2−B+m( B −2−A)
A degenerate case
The Newton map is special when two roots are on a line parallel to the line passing through the
other two roots. In the form above, this occurs when α = 1, or β = 1, or α = −β. These are
14 Hubbard and Papadopol
equivalent under an affine change of variables, so we can restrict ourselves to the case α = 1,
i.e., A = 0. A bit of computation shows that in this case, the first coordinate of NF is simply
x2 /(2x − 1), which is the one-variable Newton map for solving x2 − x = 0. Thus all points with
Re x < 1/2 are attracted to the line x = 0, and all points with Re x > 1/2 are attracted to the
line x = 1. Within each of these lines we are again faced with a one-variable Newton map, so
the dynamics are simple. Thus all the really interesting dynamics occurs in the real 3-dimensional
manifold Re x = 1/2.
In this 3-manifold, we have what might be called a fibered family of rational functions, i.e., a
map of the form
θ 7→ 2θ
³ ´ ³ ´
G : R/Z × P1 → R/Z × P1 , y g (y) (1.30)
θ
28 (a2 b2 + a3 + b3 ) + 25 32 ab = 33 ; (1.36)
it is a quartic of genus 0, with three cusps; the real locus of this quartic is shown in Figure 7.
Figure 7. The locus of pairs (a, b), −4 ≤ a, b ≤ 4 for which the parabolas y = x2 + a and x = y 2 + b are
tangent.
In one dimension, a multiple root is an attracting fixed point of the Newton map, but not
superattractive. In two dimensions, the situation is a great deal more complicated. It is easier to
study if ³we ´go³back
´ ³to ´
a variant of the form (1.24), and consider the pencil of conics that pass
0 1 0
through 0 , 0 , 1 and have slope m at the origin. The space of quadratic functions defining
such conics is 2-dimensional, and we may take as basis the two functions
Lemma 1.3.1. For the mapping F of (1.33), the polar locus PF is the hyperbola of equation
4xy = 1.
Newton in C2 17
We simply cannot exclude PF from our dynamical system. We would then have to exclude its
inverse image, and the inverse image of that, etc., and we would lose all control over the domain
of our dynamical system. In addition, we saw in the introduction to this paper, in the computer
tour of Newton’s method, that at least in the real plane, the polar locus and its inverse images
form the most interesting part of all the dynamics.
There is an obvious cure: consider NF as a mapping in the projective plane.
Lemma 1.3.2. The map NF extends to the line at infinity, except at the points at infinity on the
axes of the parabolas (1.32). Moreover, the extension is the identity on the line at infinity (except
at these points), and the eigenvalues of the derivative are 1 and 2. The eigenvalue 1 corresponds
to the direction of the line at infinity.
Proof. In homogeneous coordinates, the Newton map becomes
NF : [x : y : z] 7→ [2x2 y + y 2 z − 2ayz − bz 3 : 2xy 2 + x2 z − 2bxz 2 − az 3 : 4xyz − z 3 ]. (1.40)
In particular, if x 6= 0, y 6= 0, then
NF : [x : y : 0] 7→ [2x2 y : 2xy 2 : 0] = [x : y : 0]. (1.41)
This shows that the map extends, and that the extension is the identity. We leave the computation
of the eigenvalues to the reader. ¤
The computation above shows that the two points at infinity on the axes of the parabolas
(which are also the points where the parabolas are tangent to the line at infinity) are points of
indeterminacy. We will label them p1 = [1 : 0 : 0] and p2 = [0 : 1 : 0], at least if we are writing
Newton’s method in the form (1.34); in the form (1.24), these points form a non-trivial double
cover of the parameter plane and cannot be unambiguously labeled.
As was the case for the points qi , the points pi are apparently rather innocent points of inde-
terminacy, and NF extends to the projective plane blown up at these points.
Lemma 1.3.3. Let Epi be the exceptional divisor of P2 blown up at pi . Then NF extends
algebraically to Epi , and maps Epi isomorphically to the line at infinity.
Proof. We will work at p1 , and use the coordinates (u = 1/x, y) in the domain, and (u, v = y/x)
in the range. Note that (u, y) are coordinates in the blown-up plane, whereas (u, v) are coordinates
in the projective plane itself. Then u = 0 defines the exceptional divisor of the blow-up at p1 , and
y parametrizes this divisor. In these coordinates, the Newton map is written
u(4y − u)
2 2 2 2
u 7→ 2y + y u − 2ayu − bu .
³ ´
y (1.42)
2y 2 u + 1 − 2bu − au2
2y + y 2 u2 − 2ayu2 − bu2
Clearly the exceptional divisor is mapped to the line at infinity, by the map
0 7→ 0
³ ´ ³ ´
y 1/(2y) . ¤ (1.43)
The points qi , pj give us five points of indeterminacy for NF . Let us see that that is all.
Proposition
in 1.3.4. The points p1 , p2 , q1 , q2 , q3 are points of indeterminacy of NF , and they
are the only ones.
18 Hubbard and Papadopol
Proof. This is a straightforward matter of writing the mapping in homogeneous coordinates and
seeing where all three coordinate values vanish. It is easier in the form (1.24); we need to find
points where
x(Bx2 + 2xy + Ay 2 − xz − Ayz) = 0 (1.44)
y(Bx2 + 2xy + Ay 2 − Bxz − yz) = 0 (1.45)
2 2 2
z(2Bx + 4xy + 2Ay − (2 + B)xz − (2 + A)yz + z ) = 0. (1.46)
To list the solutions to these equations, note that:
• The value z = 0 leads to Bx2 + 2xy + Ay 2 , giving the two points of indeterminacy at infinity
p1 , p2 ;
• The value x = 0 leads to z = Ay, the point q1 ;
• The value y = 0 leads to z = Bx, the point q2 ;
• The last point is found by setting Bx2 +2xy+Ay 2 −xz−Ayz = Bx2 +2xy+Ay 2 −Bxz−yz = 0,
which leads to x(1 − B) = y(1 − A), and the point q3 . ¤
Lemma 1.3.5. The polar curve PF passes by the following 11 distinguished points:
(a) The two points {p1 , p2 } at infinity on both parabolas;
(b) The six centers of the line segments joining the roots;
(c) The three intersections {q1 , q2 , q3 } of lines joining disjoint pairs of roots.
Proof. Part (a) is immediate, but (b) and (c) are rather hard to show in this form. If you switch
to the normalization (1.24), the polar locus is given by the equation
2Bx2 + 4xy + 2Ay 2 − (2 + B)x − (2 + A)y + 1. (1.47)
³ ´ ³ ´
It is straightforward to show that the points 1/2
0
and 1/B
0
satisfy this equation; this is enough.
¤
Proposition
in 1.4.1. The critical locus of NF is the union of the two cubic curves C1 and C2 of
equation
2xy 2 − x2 + 2xb − y + a = 0 and 2x2 y − y 2 + 2ay − x + b = 0. (1.50)
Newton in C2 19
Remark. The fact that the critical locus is reducible is rather surprising, and is a direct conse-
quence of the fact that [D2 F ] is constant. This will usually not happen when we intersect curves
of higher degree, and correspondingly the critical locus will then usually not be reducible. △
In dynamics, the critical values are often more important than the critical points; here they are
especially simple.
Lemma 1.4.2. (a) The images of the curves C1 and C2 are the parabolas Γ1 and Γ2 of equation
respectively
x = y2 + b and y = x2 + a. (1.51)
respectively.
In particular, we observe that the parabolas we chose to intersect are precisely the critical value
locus.
Proof. This is a straightforward computation from Equation (1.34). ¤
This now allows a fairly detailed description of C1 and C2 . They pass through the roots, and
they are non-singular if the roots are simple, i.e., if the two parabolas we are intersecting are not
tangent. They also go through the three points q1 , q2 , q3 of Lemma 2.4,(c), for the following reason.
They must intersect each invariant line in three points (counted with multiplicity, of course), and
two transverse intersections are at the roots. Where can the third point be? Not at infinity, since
this is a fixed point of NF and is not on Γ1 or Γ2 . Not at any point of the invariant line other than
the roots where NF is well defined. There is only the point of indeterminacy left.
We can use this information to describe the lines Li , images of the exceptional divisors obtained
by blowing up qi .
Lemma 1.4.3. From qi , draw the parallels to the axes of Γ1 and Γ2 . These lines intersect Γ1 and
Γ2 respectively at a single point; Li is the line joining these points. This line is tangent to both
parabolas.
Proof. The exceptional divisor Eqi intersects each critical cubic in a single point, so the image
Li of Eqi is a line that intersects each critical value parabola in a single point. But that means
that it is a line tangent to both Γ1 and Γ2 . There are exactly three of these double tangents. But
we know that C1 maps “vertically” to Γ1 , and C2 maps “horizontally” to Γ2 ; the description of Li
follows. ¤
The cubics C1 , C2 intersect the line at infinity as follows: C1 is tangent to the x-axis at ∞ and
tangent to the line at ∞ on the y-axis, and C2 is tangent to the y-axis at ∞ and tangent to the
line at ∞ on the x-axis. Note that this accounts for the three points at infinity of both C1 and
C2 , and also for the six intersections of each with PF : there is one double intersection at infinity,
one simple intersection at infinity, and three intersections at the points q1 , q2 and q3 , which must
therefore be transversal.
20 Hubbard and Papadopol
Figure 9. The two critical cubics and the critical value parabolas, when a = −4 and b = −3. We have
also drawn in the polar curve of equation 4xy = 1. The horizontal parabola and its inverse image are
drawn dark; the vertical parabola and its inverse image are drawn light.
Proposition
in 1.5.1. The inverse image of a line is a non-singular cubic in P2 , unless the line
is tangent to one or both of the critical value parabolas, or passes through a root. The five
exceptional cases are:
• Codimension 1 exceptional lines
If a line is tangent to exactly one of the critical value parabolas, not at one of the roots, then
its inverse image is reducible: it consists of a conic and the line through the point of tangency
parallel to the axis of the parabola. These two curves intersect transversally on the critical
cubic, at the two inverse images of the point of tangency.
If the line passes through a root, and it is not tangent to one of the critical value parabolas
at the root, then its inverse is a cubic of genus 0, with a double point at the root.
Newton in C2 21
• Codimension
in 2 exceptional lines
If a line is tangent to both critical value parabolas, its inverse image is the union of three
lines. Two of these lines are parallel to the axes of the parabolas. These two intersect at one of
the points of indeterminacy, and the third line joins the other two points of indeterminacy.
If a line is tangent to a critical parabola at a root, its inverse image is the union of a line
and a conic, both passing through the root and tangent there. The line is parallel to the axis
of the parabola.
If a line passes through two roots, its inverse image is reducible, consisting of itself (it is an
invariant line), and a conic that intersects it transversally at the roots.
The various possibilities are illustrated in Figures 10, 11, and 12.
Figure 10. Left: a generic line and its inverse image, which is a non-singular cubic. Right: the inverse
image of a line tangent to one of the critical value parabolas Γ1 ; its inverse image is the union of a line
parallel to the axis of the parabola (vertical in this case) through the point of tangency, and a conic, which
intersects the line at the points where the line intersects the critical curve C1 .
Figure 11. Left: a line through a root and its inverse image. Right: a line tangent to a critical value
parabola at a root, and its inverse image.
22 Hubbard and Papadopol
Figure 12. Left: a line through two roots and its inverse image; one component of its inverse image is
itself. Right: a line tangent to both critical value parabolas, and its inverse image, which consists of three
lines.
Proof. The inverse image of a line l can be singular only at critical points of NF or at points of
indeterminacy (which are transverse intersections of the critical curves C1 and C2 ). If l is transverse
to a critical value parabola, then a straightforward computation shows that the inverse images of
such a point are non-singular points of NF−1 (l), but are critical points of the restriction of NF to
NF−1 (l).
Each of the points qi “maps” to a line Li by Lemma 1.2.4, and the inverse image of l ∩ Li is at
qi . Unless l is not one of the Li , it intersects Li at a single point transversally, and we have that
NF−1 (l) passes through q1 , q2 , q3 and is smooth at these points. A similar argument shows that
NF−1 (l) is smooth at infinity. This proves the first part.
If the line l is tangent to a critical value parabola, then its inverse image must have an ordinary
double point at both of the inverse images in the critical cubic. But an irreducible cubic cannot
have two double points, so the inverse image is necessarily a union of a line ml and a conic. The
conic is smooth, as otherwise it would degenerate to two lines, which would generate a third singular
point. The line must join the two double points, so it is the line through the point of tangency
parallel to the axis of the critical value parabola to which l is tangent, by Equation (1.52). The
conic must then pass through all of the points q1 , q2 , q3 .
If a line l passes through a root and is not tangent to a critical value parabola there, a local
computation shows that its inverse image has an ordinary double point there. If l does not pass
through any other root, this is the only singular point on NF−1 (l), which is then a cubic of genus
0. This proves the “codimension 1” part.
A line tangent to both parabolas is one of the Li . Passing to the limit in the argument about
lines tangent to a critical value parabola, we see that the inverse image of a double tangent must
contain two lines m′i and m′′i parallel to the axes of the parabolas through the points of tangency;
these intersect the point qi by Lemma 1.4.3. A cubic that contains two lines is the union of three
lines, so NF−1 (Li ) must contain a third line m′′′ i , which must go through the other two points of
indeterminacy. In fact, mi must intersect mi and m′′i on their respective critical cubics, so these
′′′ ′
The inverse image of an invariant line l is evidently itself and a conic, which must pass through
the roots on l and be transverse to l there. ¤
Remark. Proposition 1.5.1 applies to the cases where the roots are distinct, not on parallel lines.
The degenerate Newton maps are slightly different; we will not describe the differences here. △
We can now compute the degree of the Newton map.
Theorem
in 1.5.2. The mapping NF has topological degree 4.
Proof. There are several ways of seeing this. One is to take a generic line l, and its inverse image
Cl = NF−1 (l), which has genus 1. Then NF induces a mapping C → l, of some degree d, with
critical points at the inverse images of l ∩ (Γ1 ∪ Γ2 ), which consists of four points. Each point gives
rise to two critical points, so there are eight of them, and they are all ordinary critical points. Now
apply the Riemann-Hurwitz formula:
χ(Cl ) = 0 = dχ(l) − (number of critical points) = 2d − 8, i.e., d = 4. ¤ (1.55)
Remark. Another way to calculate the degree is to intersect the cubic curves C1,u and C2,v of
equation
(4xy − 1)(x − u) − 2y(x2 − y + a) − (y 2 − x + b) = 0 (1.56)
and
(4xy − 1)(y − v) − 2x(y 2 − x + b) − (x2 − y + a) = 0. (1.57)
These curves must intersect at nine points, and we know that they intersect at q1 , q2 , q3 , p1 ,
and p2 . It is not too hard to see that they
³ are
´ transverse at these points, and the other four points
u
of intersection are the inverse image of v . △
Theorem
in 1.5.3. For the Newton map NF , the basin of each root is connected.
Proof. The mapping is locally four-to-one near the roots. So we can choose a connected neighbor-
hood U0 of a root such that U1 = NF−1 (U0 ) is connected. Define Uk = NF−1 (Uk−1 ); we must prove
that all the Uk are connected. Suppose Uk is the first disconnected one, choose x ∈ Uk , and choose
a path γ in Uk−1 connecting NF (x) to a point of U0 . By a small perturbation, we may assume
that γ does not intersect the critical value locus Γ1 ∪ Γ2 , or the three double tangents L1 , L2 , L3 .
Then the inverse image of γ consists of four arcs, all ending at points of U1 . One arc must lead to
x in Uk . ¤
root at the origin they are the points of Pa corresponding to the axes, and to the line of equation
αy = βx.
In the representation (1.34), on the other hand, the points of Pa corresponding to vertical and
horizontal lines are the critical points of ga . Thus we will naturally find ga in the form
z2 + c
z 7→ , (1.58)
z2 + d
with critical points at 0 and ∞.
Comparing these two normalizations, we see that in the normalization (1.34), the rational func-
tion ga is never a polynomial, or conjugate to a polynomial. Indeed, polynomials have a fixed
critical point, and the vertical or horizontal lines are never invariant lines of NF in the form (1.34).
In the form (1.24), polynomials do arise, precisely when the roots are on pairs of parallel lines.
Since we know where the roots are, it is easiest to deal with the form (1.24). The quadratic
terms at the origin give the rational function
z 2 + Bz
(1.59)
Az + 1
with fixed points
0, with multiplier B;
∞, with multiplier A; (1.60)
β
, with multiplier α + β.
α
Note that these multipliers do satisfy the Fatou relation [Mi2]
X 1
= 1, (1.61)
mi − 1
where the sum is over the fixed points, and the mi are the multipliers of the fixed points. In the
case where one of the multipliers is 1, a more elaborate relation is required, but a look at equations
(1.60) shows that this case does not arise: it would require α = 0, or β = 0, or α + β = 1, i.e.,
that three roots be on a line, and we have seen that in that case, there is no Newton map, even
degenerate.
From this we can easily read off the multipliers of the other roots. The affine map that permutes
the roots as follows:
α 7→ 1 7→ 0 7→ 0 7→ α′
³ ´ ³ ´ ³ ´ ³ ´ ³ ´
β 0 0 1 β′ (1.62)
is
x 7→ y/β
³ ´ ³ ´
y −x − Ay + 1 (1.63)
³ ′´ ³
and in particular α 1/β
´
β ′ = (α + β − 1)/β .
Thus the multipliers of the fixed points previously at 10 , now at 00 , are
³ ´ ³ ´
β−1 α+β
A′ = , B′ = 1 − α , and α′ + β ′ = . (1.64)
α+β−1 β
One can of course continue this way for the other roots.
Figure 13 illustrates the multipliers.
Newton in C2 25
( 01 (
1-β
β−1
A
A-1 α+β
α+β
β
( αβ (
α
α+β−1 α−1
α
α+β
A α+β 1−α
β
B B
0
( (
0
B-1
( 10 (
Figure 13. Each invariant line corresponds to a fixed point of a rational function at each root it passes
through; this picture represents the corresponding multipliers.
There is a way to condense this information into a cleaner algebraic form (which still gives no
relation between the dynamics of the various rational maps). Note that one way of parametrizing
rational functions with a distinguished fixed point is by the multiplier m of that fixed point, and the
product µ of the multipliers at the other two fixed points. We can then summarize the information
in Figure 13 by the following statement.
Proposition
in 1.6.1. Let L be an invariant line of NF , and suppose m, µ and m′ , µ′ are the
coordinates of the two rational functions corresponding to roots on L, with m and m′ being the
multipliers of of the fixed points corresponding to L. Then
m
m′ = and µ = µ′ . (1.65)
m−1
2
Invariant 3-manifolds associated to invariant circles
The most striking feature of computer drawings for Newton’s method is the apparent existence
of smooth 3-manifolds in the dynamical plane, which make up most of the apparent boundary of
the basins. We will prove in Theorem 2.4.1 that under appropriate circumstances, there actually
are invariant 3-manifolds in P2 , which do belong to the boundaries of the basins.
In computer pictures, these manifolds look remarkably smooth. The pictures are misleading;
sections of these manifolds by complex lines are quasi-circles, or at least quasi-arcs, which are
non-differentiable on a dense subset (though they may be differentiable almost everywhere).
These manifolds are some sort of “stable manifolds” of invariant circles. But their existence
does not follow from any variant of the stable manifold theorem, as there are always parts of them
that are not attracted to the circles. In the final analysis, we control the topological structure of
these manifolds using a completely different tool, the λ-lemma of [MSS] concerning holomorphic
motions.
We also prove, in Theorem 2.3.4, that every point at infinity except p1 and p2 has an unstable
manifold. This is much less delicate than the previous result, and explains most of the self-similarity
one sees in computer pictures: the points that appear to be centers of self-similarity are simply
inverse images of points at infinity.
In Sections 2.1 and 2.2 we will write our Newton map in the form (1.24).
26
Newton in C2 27
Theorem
in 2.1.1. The Lyapunov exponents of the circle Re x = 1/2, y = 0 are λ1 = 2, corre-
sponding to the circle itself, and
|B| if the point of indeterminacy 1/B is in the basin of 1
(
λ2 = ¯¯ B ¯¯ (2.5)
¯ B−1 ¯ if the point of indeterminacy 1/B is in the basin of 0.
The rational functions gà ! and gà ! corresponding to the roots on the x-axis each have a fixed
0 1
0 0
Theorem
in 2.1.1’. The second Lyapunov exponent λ2 of the invariant circle in an invariant
line l is the absolute value of the multiplier of the fixed point corresponding to l of the rational
function at the root in l that does not attract the point of indeterminacy in l.
The second Lyapunov exponent λ2 above always satisfies |λ2 −1| ≤ 1, so we do have 0 ≤ λ2 ≤ λ1 .
If the point of indeterminacy is on the circle, then |B| = |B/(1 − B)|, and in that case this
number is λ2 .
Proof. By the ergodic theorem, the second Lyapunov exponent exists almost everywhere, and is
an invariant function on the circle. Angle doubling is an ergodic map of the circle to itself, hence
this function is constant (almost everywhere), and equal to its space average. In other words, we
have
µ Z ∞ ¯ ¯ ¶
1 ¯ B((1 + it)/2)((1 − it)/2) ¯ dt
λ2 = exp log ¯
¯ ¯
π −∞ it(1 − B(1 + it)/2) ¯ 1 + t2
µ Z ∞ (2.6)
1 + t2
¯ ¯ ¶
1 ¯ ¯ dt
= exp log ¯
¯ ¯ .
π −∞ 2it(2 − B(1 + it)) ¯ 1 + t2
Evaluating this integral is an entertaining exercise in the calculus of residues. Write it as the
sum of
1 ∞³ ´ dt
Z
log |B|− log 2 + log |t + i| + log |t − i|− log |t| = log |B| − log 2 + log 2 + log 2 + 0 (2.7)
π −∞ 1 + t2
and
∞ ¯ ¯ ½
− log |1 − B| − log 2 if Re (1/B) < 1/2
1 1 ¯ dt
Z
¯
log ¯
¯ ¯
2
= (2.8)
π −∞ 2 − B(1 + it) 1 + t
¯ − log 2 if Re (1/B) > 1/2.
The only terms requiring care in the first integral are the ones involving log |t ± i|. To evaluate
them we observe that the function log(z + i) has an analytic branch in the upper half-plane, so we
can close the interval [−R, R] with a half-circle in the upper half-plane, to form a closed curve ΓR
bounding a half-disc containing the pole of 1 + z 2 at i. Applying the residue formula, we find
log(z + i)dz log 2
Z
2
= 2πi = π log 2. (2.9)
ΓR z +1 2i
28 Hubbard and Papadopol
In the standard way, let R → ∞, so that the contribution from the half-circle tends to 0.
The third integral is evaluated by residues also; we need to avoid the zero of 1 − B(1 + iz)/2,
which is in the upper half-plane or the lower half-plane depending on the sign of Re (B − 1/2), so
we integrate around a semi-disc in the other half-plane.
In the final analysis, the average normal expansion on the circle is
( ¯¯ ¯
B ¯
¯ 1−B ¯ if Re (1/B) < 1/2
(2.10)
|B| if Re (1/B) > 1/2.
This makes sense: one of the two roots on that line is distinguished, the one that attracts the
point of indeterminacy 1/B. What we are finding depends on the non-distinguished point, and is
the multiplier of the fixed point of the rational function associated to the invariant line itself.
Note that the multiplier is always a number in [0, 2]; the value 0 occurs when B = 0, i.e.,
β = 1, which means that the roots are on the two parallel lines y = 0 and y = 1. The value 2 is
realized exactly if Re (1/B) = 1/2, the case omitted above. This happens for instance for the case
α = β = −1, which also corresponds to solving the two equations
x2 = y, y 2 = x. (2.11)
conjugates u 7→ u2 to Newton’s method. In particular, it maps the unit circle to′ the invariant
m
2m 2m
circle in the x-axis. Set ζm,l = e2πil/(2 −1) , so that ζm,l = ζm,l . Suppose that ζm,l 6= ζm,l for all
m′ with 1 < m′ < m. Then the points
m−1
2 2
φ(ζm,l ), φ(ζm,l ), . . . , φ(ζm,l ) (2.13)
are precisely the cycles of length m on the invariant circle in the x-axis. The derivative of NF at
φ(ζ) has eigenvalues
Bζ
2 and ¡ ¢. (2.14)
(ζ + 1) 1 + B(ζ − 1)
Set ζ = ζm,l . The second eigenvalue of the corresponding cycle is
m−1 k
Y Bζ 2 Bm
Mm,l = = Qm−1 ¢. (2.15)
(ζ 2k + 1) 1 + B(ζ 2k − 1)
¡ ¢
2k
¡
k=1 k=0 1 + ζ (B − 1)
Newton in C2 29
We want to analyze for what m, l we have |Mm,l | < 1. The normal derivative at the point
x = z/(z − 1), y = 0 can be understood from Figure 14. The point 1 + z(B − 1) is the point z,
rotated by arg z around 1. Thus the ratio B/(1 + z(B − 1)) is easy to visualize: it is < 1 outside
the emphasized arc from B to B, but > 1 inside the arc. The arc disappears when B ∈ (0, 1), so in
that case (and, as we will see, nearby ones) all cycles are saddles. But when B moves away from
the real axis, some cycles become repelling.
B 1+z (B-1)
2
0 1
1+z1(B-1)
Figure 14. The multiplier along a cycle is the product of ratios B/(1 + z(B − 1)), i.e., ratios of the
magnitude of two points on the circle of radius |B − 1| centered at 1.
Proposition
in 2.2.1. If B is real and 0 < B < 1, we have |Mm,l | < 1 for all m ≥ 2, and all l.
Proof. A geometric picture of the product formula shows why this is obvious.
B
0 1
Figure 15. When 0 < B < 1, the derivative along every cycle is a product of ratios, every one of which
is < 1.
30 Hubbard and Papadopol
Remark. There is one cycle to which the argument above does not apply: the fixed point at
infinity. △
The argument above can be refined to show that there is a neighborhood of the interval (0, 1)
in C such that for B in that neighborhood, all the products have absolute value < 1.
Theorem
in 2.2.2. When B is in the region defined by the inequalities
¯ ¯ ¯ ¯
¯B − 1 ¯ 1 ¯ 1 ¯¯ 1
¯ B arg(1 − B)¯ < 3 , ¯B − 2 ¯ < 2 ,
¯ ¯ ¯
(2.17)
B2
¯ ¯ µ ¯ µ ¶¯¶
¯ exp 4 ¯ Im B − 1 arg(1 − B) ¯ < 1.
¯ ¯ ¯ ¯
¯
¯ (2 − B)(1 − |1 − B|) ¯ ¯ B ¯
all cycles on the invariant circle in the x-axis are saddles, except for the fixed point at infinity.
Remark. The second inequality is almost a consequence of the first: the relation of the two
regions is represented in Figure 16. △
1.0
-1.0 2.0
-1.0
Figure 16. The interior of the wing-shaped curve corresponds to the region given by the first of the
inequalities (2.17); it is almost contained in the region specified by the second inequality.
Proof. Choose B with |B| < 1, and |1 − B| < 1, and consider the circle C passing by B and
centered at 1. Call C0 the arc bounded by B and B, contained in the unit disc centered at the
Newton in C2 31
k
origin (see Figure 14). Call zk = 1 + ζ 2 (B − 1). Note that z ∈ C0 if and only if |B/z| > 1; if
|B/z| < 1, it is in the complementary arc. So the points zk ∈ C0 are the ones to worry about.
We will consider all the maximal subsets zp−1 , zp , . . . , zp+q with zp , . . . , zp+q ∈ C0 , so that
zp−1 ∈/ C0 .
What makes the theorem hard to prove is that there is no bound on q; there can be arbitrarily
long subchains of the cycle in C0 . But it turns out that
q ¯ ¯
Y ¯ B ¯
¯ ¯
¯ zp+k ¯ (2.18)
k=0
is convergent. We will want not just that it is convergent, we will want an explicit bound; Lemma
2.2.3 is a first step in that direction.
where we have used log(1 + x) ≥ 2x when −1/3 ≤ x ≤ 0. Exponentiation gives the result. ¤
The next step is to replace our product with one of the form studied in Lemma 2.2.3. In order
to do this, let us take u0 = zp+q , u1 = zp+q−1 and more generally
µ ¶1/2k
u0 − 1
uk = 1 + (B − 1) , (2.22)
B−1
where the power is the principal branch, well defined since B/u0 is close to 1. Unlike the z’s, the uk
are defined (by Equation (2.22)) for all k ≥ 0, and hence well adapted to taking infinite products.
Even the points uk are a bit awkward, and we will replace them by better points wk , whose
definition is suggested by Figure 17.
32 Hubbard and Papadopol
w2
B
w1 u2
u1
0 1
w0 u0
2-B
B
Figure 17. This illustrates how to find lower bounds for the mutipliers. The w’s are closer to 0 than are
the u’s.
z−1
Proof. This is easy but fussy plane geometry. It is easiest to make the transformations z 7→ B−1 ,
to obtain Figure 18.
1-B 1-w
1-B E
1-B 1-u
1-B F
1 θ φ O
1-B D
C
arg (1-u) β
arg 1/(1-B) α
1
Figure 18. We have made the change of variables z 7→ (z − 1)/(B − 1). On the right, we have written
arbitrary labels in standard “Euclidean geometry” style.
The first thing to observe is that the requirement |B − 1/2| < 1/2 translates into saying that
the point labeled 0 is in the half-plane where the real part is > 1. The requirement | arg B| < π/3
guarantees that the point E in the figure is above the point D. There are actually two cases to
Newton in C2 33
consider, arg(1 − u) > 0 and arg(1 − u) < 0. We will do only the first here, and leave the other,
which is easier, to the reader. The object is to show |Ow| < |Ou|.
Consider the two triangles (ODE) and (OCF ). They satisfy the following properties:
|EO|2 < |ED|2 + |OD|2 < |OC|2 + |F C|2 < |OF |2 , (2.27)
where the first and last inequalities come from the law of cosines and the fact that θ is acute and
φ is obtuse. ¤
It follows that
∞ ¯ ¯ ∞ ¯ ¯
Y ¯B¯ Y ¯B¯
¯ ¯≤ ¯ ¯. (2.28)
¯ uk ¯ ¯ wk ¯
k=0 k=0
Remark. There are definitely values of B where |B| < 1 and |1 − B| < 1, so that the invariant
circle in the x-axis is on average attracting, but for which this circle contains infinitely many
repelling cycles. We will see examples of such things in Figures 29 and 30. △
Theorem
in 2.3.1. There exists a neighborhood U of l∞ − {p1 , p2 } in P2 , a neighborhood V of
the y-axis with 0 removed in C2 , and an analytic isomorphism Φ = (φ1 , φ2 ) : U → V such that
· ¸
2 0
Φ = Φ ◦ NF (2.33)
0 1
on U ∩ NF−1 (U ).
Any Ψ satisfying the same property on a neighborhood of any open subset of l∞ − {p1 , p2 }
is of the form Ψ = (ψ1 , ψ2 ), where
ψ2 = α(φ2 ), and ψ1 = β(φ2 )φ1 (2.34)
for appropriate analytic functions α, β, where β does not vanish.
In particular, we see that the curves φ1 = constant are mapped to themselves by Newton’s
method, providing a dynamically natural foliation of a neighborhood of l∞ − {p1 , p2 } by invariant
Riemann surfaces. How these can be extended is an obviously interesting problem, which will be
examined in Section 2.5.
Proof. The first step is to find an appropriate neighborhood of l∞ − {p1 , p2 }. Let
(µ ¶ ¯ r )
x ¯ |x|
¯
2
UC = < |y| < C|x| ∪ (l∞ − {p1 , p2 }), (2.35)
y ¯ C
¯
Newton in C2 35
|y|
UC
|x|
Figure 19. The region UC is a neighborhood of the line at infinity with the points p1 , p2 removed.
Proof. We will show that the restriction of NF to UC ∩ N −1 (UC ) is proper and a local homeo-
morphism. It is then a finite sheeted covering map, of some degree. But the points at infinity have
only themselves as inverse images in UC , so the degree is 1, and the map is a homeomorphism.
Clearly when C is sufficiently small UC will not intersect either of the critical value parabolas
or contain any of q1 , q2 , q3 ; the statement about NF being a local homeomorphism follows.
It remains to show that the restriction of NF is proper. Let us denote P̃2I the projective plane
blown up at the five points of indeterminacy, and ÑF : P̃2I → P2 the lifted Newton map, which is
now well defined by Lemmas 1.2.4 and 1.3.3, and is evidently proper. Let xn be a sequence in UC
converging to a ∈ UC , and x′n be a sequence of preimages in UC ∩ NF−1 (UC ); we need to show that
we can extract a subsequence of the x′n that converges in UC ∩ NF−1 (UC ). Since ÑF is proper, we
may assume that x′n converges to a′ ∈ P̃2I ; of course we have ÑF (a′ ) = a.
Let us see where a′ can be. One possibility is that it might be at infinity. In that case, it is
fixed, contradicting a ∈ UC and ÑF (a′ ) = a. This is true even if a′ ∈ Ep1 ∪ Ep2 , as the sequence
x′n would approach p1 or p2 tangentially to the line at infinity, and the corresponding points of
the exceptional divisors are fixed.
Another possibility is that a′ ∈ ∂UC . That is excluded by Lemma 2.3.3; points in the boundary
of UC are not mapped to points in the closure of UC .
when (x, y) ∈ UC ∩ C2 , and the indices 1 and 2 indicate the first and second coordinates.
Proof. This is a straightforward matter of seeing what terms dominate in the formula for NF .
Recall that
µ 2
2x y + y 2 − 2ay − b
¶
x
³ ´ 1
NF y = (2.38)
4xy − 1 2xy 2 + x2 − 2xb − a
36 Hubbard and Papadopol
and we have
¯ 2
¯ 2x y + y 2 − 2ay − b x ¯ ¯ 2xy 2 + x2 − 2ax − b y ¯
¯ ¯ ¯
¯ − ¯¯ ∈ o(|x| + |y|), ¯ − ¯¯ ∈ o(|x| + |y|). ¤ (2.39)
¯ 4xy − 1 2 ¯ 4xy − 1 2
A third thing that could happen is that a′ ∈ UC ∩ ∂NF−1 (UC ). But at such a point NF is
a local homeomorphism, mapping boundary to boundary, so that is excluded also. Thus a′ ∈
UC ∩ NF−1 (UC ). ¤ Lemma 2.3.2
Next, we give a general “unstable manifold theorem” under a relaxation of the present hy-
potheses, where we assume that one eigenvalue is 1, and there is a corresponding analytic “center
manifold”.
Consider an open subset U ⊂ C, a neighborhood V ⊂ C of 0, and a mapping U × V → C2 of
the form
F : u u + vf (u, v)
³ ´ ³ ´
v →
7 v(λ(u) + vg(u, v)) , (2.40)
Proposition
in 2.3.4. There exist local coordinates (ũ, ṽ) on a neighborhood of U × {0} , with
u = ũ on U × {0}, such that in these coordinates the mapping F is given by
ũ ũ
³ ´ ³ ´
F : ṽ 7→ λ(ũ)ṽ . (2.41)
Proof. This is a matter of putting parameters into the standard proof of the linearization theorem
for saddles with 1-dimensional stable manifolds, or more precisely, the existence of a parametriza-
tion of the stable manifold that linearizes the mapping in the stable manifold.
Choose R0 > 0, and M such that
|f (u, v)| ≤ M , |g(u, v)| < M when |u|, |v| ≤ R0 . (2.42)
Set
Λ= sup |λ(u)| and choose A so that Λ < A < 1. (2.43)
|u−u0 |≤R0
and
|vn+1 | = |vn ||λ(un ) + vn g(un , vn )| ≤ An |v0 |(Λ + M R) ≤ An+1 |v0 | (2.48)
It follows immediately from these inequalities that the sequence un converges; we can now define
ũ(u0 , v0 ) = lim un . (2.49)
n→∞
Clearly ũ is an analytic function on some neighborhood of U ×{0}, and ũ(u, 0) = u and ũ(F (u, v)) =
ũ(u, v).
Now define ṽ by the formula
vn
ṽ(u0 , v0 ) = lim . (2.50)
n→∞ (λ(ũ(u0 , v0 )))n
We will show that when v0 is small enough, this is a Cauchy sequence. To simplify notation, we
will write λ(ũ(u0 , v0 )) = λ̃. We find
¯ ¯ ¯ ¯
¯ vn+1 vn ¯¯ |vn | ¯¯ λ(un ) + vn g(un , vn ) ¯
¯ λ̃n+1 − λ̃n ¯ = |λ̃|n ¯ − 1¯¯ . (2.51)
¯
λ̃
Now to bound these quantities:
|un − ũ| ≤ C1 An gives |λ(un ) − λ̃| ≤ C2 An , (2.52)
n
which together with |vn g(un , vn )| ≤ M RA gives
¯ ¯ µ 2 ¶n
¯ vn+1 vn ¯¯ A
¯ λ̃n+1 − λ̃n ¯ ≤ C3 |λ̃| (2.53)
¯
for some constants C1 , C2 , C3 which could be computed from R, M , and the error in the linear
approximation to λ around ũ.
Unfortunately, A2 /|λ̃| is not smaller than 1 unless we do something to make this happen. If
we make R very small, then λ(un ) is very nearly constant, and we can choose A so that A2 <
|λ(un )| < A. With R this small, the series defining ṽ converges. ¤
To prove Theorem 2.3.1, apply Proposition 2.3.4 to the mapping NF−1 : UC → UC , which is
guaranteed to exist by Lemma 2.3.2. ¤
In Section 2.2, we found conditions under which all the cycles in Sl are saddles, repelling in the
direction of l, of course, and attracting in the “normal” direction.
In this section, we will examine whether Sl has a stable manifold. This, if true, is much better
than just knowing that the cycles are individually saddles; it is saying that the stable manifolds of
these cycles fit together so that their closure is a 3-dimensional manifold.
The problem with this program is that it obviously does not work. At l ∩ l∞ , we can clearly
see the splitting of the tangent space, into the direct sum of the tangent space to l (the direction
of the base), and the tangent space to l∞ (the direction of the fiber). These are the only invariant
subspaces. But the Newton map is the identity on l∞ , (except at the two points of indeterminacy,
but l ∩ l∞ is not a point of indeterminacy). Thus no point of l∞ can be attracted to Sl . Neither
can any point of the polar curve, or any point of its inverse images, etc. The inverse images of
l ∩ l∞ are dense in Sl , so there is a dense set of points that do not have a stable manifold in any
classical sense of the word.
Despite this, we will show that a sort of “center-stable” manifold of Sl exists under appropriate
conditions. It does not have the smoothness one would expect if Sl were hyperbolic, and the
techniques used to show that it exists at all are quite different from the standard tools: one key
ingredient is the theory of holomorphic motions.
Theorem
in 2.4.1. Suppose the point (α, β) is chosen so that B = (1 − β)/α is real and satisfies
0 < B < 1. Then there exists a neighborhood Ω ⊂ C2 of (α, β) such that if (α′ , β ′ ) ∈ Ω, then
there exists a 3-dimensional topological manifold Vl ⊂ P2 , together with an inclusion Sl ⊂ Vl
and a projection p : Vl → Sl whose fibers are Riemann surfaces homeomorphic to discs, such
that
p|Sl = id and Nα′ ,β ′ (Vl ) ⊂ Vl , (2.55)
and such that
p ◦ Nα′ ,β ′ = Nα′ ,β ′ ◦ p (2.56)
on Vl .
A section of Vl by a complex line sufficiently close to l is a topological circle. Our proof will
do better; it will show that such a section is a quasi-circle. This is probably optimal. Although
computer pictures show 3-dimensional manifolds separating basins that appear to be smooth, this
is probably false; these apparently smooth manifolds are probably not in fact C 1 on any open
subset. Following the proof, we will give examples to demonstrate both why they look smooth,
and why they are not.
Proof. The proof of Theorem 2.4.1 is long (more than four pages) and requires a number of
intermediate results. The beginning of the proof looks rather like the graph transform argument,
except that although a graph transform map is well defined, it is not contracting.
Without loss of generality, we will assume throughout that the invariant line we are considering
is the x-axis. The first thing to do is to write the derivative of N along l.
The first step in constructing the graph transform map is to find an invariant family of cones.
Lemma 2.4.4. There exist a neighborhood W̃ of Sl and a (small) constant Θ such that if CÃ !
z
v
is the subset
· ¸
ξ
2
∈T Ã ! P | |η| > Θ|ξ| (2.61)
η
z
v
then
³ ´´−1
DN vz
³
C z ⊂ CÃ z ! (2.62)
à !
N
v v
z
³ ´
for all v in W̃ .
Proof. First, let us see this when |z| = 1 and v = 0. There it follows from the fact that the second
diagonal term of the matrix (2.59) is smaller than the first, and the off-diagonal term is bounded.
We can choose a single Θ that works for all points, since Sl is compact. By continuity, the result
will hold in a neighborhood of Sl . ¤
We will consider the variable z as “horizontal” and the variable v as “vertical,” and correspond-
ingly we will set
∂H W1 = the part of the boundary where |v| = ǫ1
∂V W1 = the part of the boundary where |z| = e±δ
(2.64)
∂H W2 = unstable manifolds of points where |v| = ǫ2
∂V W2 = the part of the boundary where |z| = e±δ .
We will also consider the part of the boundary where arg z = δ as part of the horizontal boundary.
Figure 20 shows these regions, with z drawn as 2-dimensional and v drawn as 1-dimensional; Figure
21 is an attempt to draw their images (drawn separately for W1 on the left and W2 on the right.)
W2
e−δ1
δ2
ε1 δ ε2
e 1
ε1 ε2
W1
The disc of radius
The point at infinity
on the invariant circle ε2 in the line at ∞
Figure 20. A sketch of the regions W1 and W2 . Of course, these are really 4-dimensional; we have
represented the coordinate v as 1-dimensional, whereas z is represented as a thickened circle in a plane,
which it indeed really is.
Figure 21. The images of W1 and W2 , with the image of W1 on the left and W2 on the right. The space
W1 really maps “hyperbolically,” contracting in the vertical direction and expanding in the horizontal
direction. But W2 does not: the map does expand in the horizontal direction, but the horizontal boundary
is contained in the image of the horizontal boundary, since it is made up of unstable manifolds. With the
convention of the drawing, i.e., that v is 1-dimensional, one should imagine the top and bottom of W2 as
each a single unstable manifold, which is simply “spread out” over itself by z 7→ z 2 .
Newton in C2 41
We are now in a position to state our principal lemma, which allows us to set up a sort of “graph
transform.”
is not on the boundary of W2 . The case where the point is in W2 and its image is in W1 is slightly
more complicated: it is on the manifold of a point w with |w| < ǫ2 , and so is its image. Moreover,
| log z| < δ1 /2, so log z1 | < δ1 , and
|σw (z1 )| < |w| + Kδ < ǫ2 + Kδ < ǫ1 . (2.69)
The hardest case is if a point (z, v) ∈ W1 is mapped to a point (z1 , v1 ) with | arg z1 | < δ and
e−δ < |z| < eδ ; in this case, (z1 , v1 ) ∈ W2 , and is not on the horizontal boundary of W2 . The
hypothesis implies that z1 is close to −1, and the normal multiplier in the entire left half-plane is
at most |B|. So |v1 | < |Bv|, and the unstable manifold through (z1 , v1 ) will intersect the line at
infinity at a point v2 with |v2 − v1 | < Kδ, hence with |v2 | < Bǫ1 + Kδ < ǫ2 .
The conditions we have set up will still be satisfied if we replace Nα,β by Nα′ ,β ′ , with the point
(α′ , β ′ ) chosen sufficiently close to (α, β); W1 is kept precisely as it was; and W2 is chosen as the
union of the unstable manifolds of the same points z = 1, |v| < ǫ2 for the new map.
Now suppose that z12 = z2 , and that D crosses the circle at z2 . There is then a unique component
D of N −1 (D) ∩ W that contains z1 , and D′ is an analytic disc in P2 . Our first requirement
′
guarantees that D intersects only the horizontal boundary of W , so the thing to check is that no
part of ∂D′ is in the interior of W . If a sequence in D′ converges to a point of the boundary of
D′ that is an interior point of W , its image also converges to an interior point of W , and this is a
contradiction. ¤
The two lemmas together show that the graph transform map is well defined.
Set X1 to be the polar locus, and Xn+1 = NF−1 (Xn ). Let Yn,k be the component of Xn ∩ W
n
that contains the dyadic root of unity ζn,k = e2πik/2 . Moreover, let ǫ = ǫ2 − Kδ.
Corollary
in 2.4.7. Each Yn,k is the graph Γfn,k for some fn,k : Dǫ → C expressing z as a
function of v ∈ Dǫ , with fn,k (0) = ζn,k .
Corollary
in 2.4.8. The map h : ∆ × Dǫ → C given by
h : (ζn,k , v) 7→ fn,k (v) (2.70)
is a holomorphic motion parametrized by Dǫ .
Proof. All the conditions for a holomorphic motion are clearly satisfied: the map is clearly
injective and analytic with respect to v. ¤
We can now prove Theorem 2.4.1. By the λ-lemma ([MSS]), the map h extends to a map
h̃ : Sl × Dǫ → C, quasiconformal with respect to z and analytic with respect to v. The set Vl is the
image of h̃, parametrized by Sl × Dǫ where u = h(ζ, v) is a 3-dimensional manifold, and N maps
Vl to itself. Since the diagram
Vl −→ Vl
↓ ↓ (2.71)
Sl −→ Sl
commutes on the set parametrized by ∆ × Dǫ , it commutes on Sl × Dǫ . Since the fibers of the
projection pr1 are graphs of analytic functions, they are analytic discs. ¤ Theorem 2.4.1
Newton in C2 43
In Figures 22–24 we give examples to demonstrate both why the boundaries of the basins look
smooth, and why they are not. If a bundle Vl as above exists, we should see it as a line separating
the red and green basins.
Figure 22. This represents the line y = .1x, as parametrized by x, where we have −1 ≤ Re x, Im x ≤ 2,
for the Newton map where α = β = −1. The circle in the x-axis (which is the straight line Re x = 1/2) has
multiplier B/(B − 1) = 2/3, so it is attracting, and in fact has a stable manifold by Theorem 2.4.1. You
would expect this stable manifold to cut the line y = .1x along a quasi-circle close to the line Re x = 1/2.
Indeed, this is the case. Note the complicated decorations in the basin of the origin, red in the picture; one
bubble sequence is blown up at right. This is also as one would expect: the bubbles arise because there is
a point of indeterminacy on the x-axis, and they are in the red basin because the point of indeterminacy
is in the red basin. They present a dyadic structure, as they should.
Figure 23. This represents the line of equation y = .6x, for the same Newton’s method as above. This
time, we are plotting 1/x, i.e., near the point at infinity. The undulations of the quasi-circles are very
apparent.
44 Hubbard and Papadopol
Figure 24. This represents the line y = .1x, for the Newton’s method with α = −1, β = 3. This is
actually conjugate to the one above, but this time the circle in the x-axis is repelling (with multiplier 2,
since the point of indeterminacy is on it). When we perturb the slope away from 0, we would not expect
a curve to separate the red and green basins, and indeed there is no such curve; instead, a zone of blue
appears. On the right, we see the same line, but this time near infinity. Since the red and green basins
touch along the line at infinity, they must touch along its inverse images, which have a dyadic structure.
That is what one observes.
Proposition
in 2.5.1. For each m ∈ C∗ , the mapping ψm extends to an analytic mapping ψm :
2
C→P .
Proof. The extension is defined by the formula ψm (2w) = NF (ψm (w)). The only difficulty one
can run into is when ψm (w) is a point of indeterminacy, and even then there is no real difficulty.
On any disc D ⊂ C, the map ψm : D → P2 lifts to a mapping ψ̃m : D → P̃2I , and NF : P̃2I → P2 is
well defined. ¤
As we will see, the extended ψm is no longer injective; in fact, we expect the image to be in
general very wild, perhaps dense in all of C2 .
Newton in C2 45
Another unpleasant feature is that the extended ψm does not depend continuously on m, and
this gives rise to one of the characteristic features of drawings in complex slices: bubble sequences.
We will now explore this phenomenon.
There are some values of m for which ψm can be completely understood, and in fact an explicit
formula can be given, namely those that are the slopes of the invariant lines. Within these invariant
lines the point at infinity is a repelling fixed point, and it is well known that at a repelling fixed
point of a rational function, the linearizing map extends to an entire meromorphic function.
For the map z 7→ z 2 and the fixed point 1, this extended linearizing map is w 7→ ew . In the
form (1.24), the x-axis is an invariant line, mapped to itself by NF by the formula
³ ´ ³ 2
NF x x /(2x − 1) .
´
0 = 0 (2.73)
0
0
Figure 25. The dyadic structure of inverse images of points of indeterminacy. If the point generating
the structure is real, the dyadic pattern of points is as on the left, but if it is not real the pattern will look
skewed, as on the right. The right-hand side of Figure 29 shows such a distorted dyadic pattern.
Proof. There is no difficulty in showing that Φ−1 can be extended analytically to the region
|Re w| < Re w0 + ǫ, |m| < ǫ for ǫ > 0 sufficiently small. If r < ǫ, and m 6= 0 satisfies |m| < ǫ, then
the points ψm (w) form a disc that passes very near the point of indeterminacy q1 . If we blow this
point up, the lift of this disc will come very close to the entire exceptional divisor, and its image
will come very close to the image of the exceptional divisor, which is a straight line (a line tangent
to both parabolas in the normalization (1.34).) But the original invariant line does not come close
to the image of the exceptional divisor. ¤
Thus we see something like the Picard-Lefschetz formula appearing: there is a “vanishing cycle”
on the unstable manifolds ψm , whose images, as m tends to 0, tend to the union of the invariant line
and the image of the exceptional divisor. In the disc Dr above, and at all its forward and inverse
images, i.e., at all the points 2n (w0 + 2kπi), n, k ∈ Z, we should see all of the dynamics of the
exceptional divisor. But for m sufficiently small, we should see them only in a small neighborhood
of the dyadic grid; at every other point w, ψm (w) is analytic as a function of m for m sufficiently
small. This is the source of the bubbles.
What should one see in a bubble? Mainly a third basin. The point of indeterminacy is the
intersection of two invariant lines. The disc ψm (Dr ) above will intersect the other invariant line
somewhere, and a neighborhood of that point will be in the basin of the root to which the point of
indeterminacy is attracted in the other invariant line. Usually, the bubble is primarily made up of
that basin, but with complications around the edges, which can range from other bubble sequences
in the simplest cases to horribly complicated things in other cases.
Interestingly enough, it appears that there are values of the parameters such that the images of
the ψm enter only three basins, not all four, for m sufficiently
³ ´ small. For instance, when α = β = 2,
we have been unable to find any points of the basin of 22 in the unstable manifolds parametrized
by ψm for |m| < .1. This has consequences for the Russakovskii-Shiffman measure (discussed in
Section 3.3): it is then not true that the entire line at infinity is in the support of the measure.
Figure 26 represents the unstable manifold ψm (w) for m = .05(1 + i), and α = β = 2.
µ ¶
Figure 26. Left: a sequence of blue bubbles (blue = basin of 0 ) in the red basin (red = basin of
1
µ ¶
0 ). Right: a blow-up shows what the green dot off the left-most blue bubble really looks like.
0
Newton in C2 47
In this case, the mapping is real, and we can easily visualize the unstable manifold and the
bubbles.
Figure 27. The real picture of Newton’s method when α = β = 2. A line y = mx with m small will be
mainly red and green, with blue interruptions in the red, near the point of indeterminacy and its inverse
images. The blow-up at right shows red and green stripes near the point of indeterminacy; these stripes
correspond to the green balls in the red basin in Figure 26, right. There does not appear to be any gray
in either the real or the complex picture.
The examples above are misleadingly simple. Figure 28 shows what happens when α = i, β =
(1 + i)/2, in the line y = .02x. The decorations around the boundaries of the bubbles are now
much more complicated, and all four colors are present in the picture.
This picture brings out several interesting features. One is the undulations of the quasi-circle
making the boundary between the red and the green basin. Another is the dissymmetry between
the part above the origin (which corresponds to the point at infinity) and the part beneath. The
discussion in Section 2.2 shows that such lack of symmetry is to be expected when B is not real,
and indeed here B = (1 + i)/2 is not real.
Figure 29. On the left, we see what looks like a stable manifold, separating the green and red basins, with
bubbles in the green basin exhibiting the normal sort of dyadic behavior. The vertical line has equation
Re z = 1 − cos 2π/7 ≈ .3765, and should intersect the “circle” near the points ζ3,1 (at the top) and ζ3,6 (at
the bottom), since angles are computed on the circle viewed as the boundary of the outside (the basin of
the origin). The cycle containing ζ3,1 is repelling, as we saw, and the cycle containing ζ3,6 is attracting,
with multiplier of absolute value ≈ .4103. On the right, we see a first blow-up at the top, with many more
bubble sequences, but the circle still looks smooth. This region is blown up further in Figure 30.
where ω = e2πi/7 . For the √value B = .8 + .4i, the invariant circle in the x-axis is attracting (with
average multiplier |B| = .89), but the 3-cycle φ(ζ3,k ) is repelling, with multiplier of absolute
value ≈ 2.556. We will look in the line of equation y = .05x, and in the local coordinate z = 1/x.
Moreover, we will draw the region −1 ≤ Re z ≤ 3, | Im z| < 2. In the x-axis itself, the invariant
circle
³ ´ is the circle of radius 1 centered at 1, with the basin of the origin outside and the basin of
1 inside. At that scale, in the line y = .05x we see the “circle” shown in the left-hand side of
0
Figure 29. However, the apparent circle does not really exist. If we perform three further blow-ups,
we find the structure on the right of Figure 30. Similar complications are dense on the entire circle,
but at very small scales at most places.
Figure 30. On the left, a further blow-up of the right-hand side of Figure 29 shows problems with the
proposed “circle,” presumably the appearance of a “tongue” emanating from the repelling cycle. A further
blow-up gives some details of what one should see within such a tongue.
3
The behavior at infinity when a = b =
In this chapter we will study Newton’s method as applied to the two equations
y = x2 , x = y2 . (3.1)
The associated Newton’s method is
1 ³ 2
2x y + y 2 .
N: x
³ ´ ´
y 7→ 2
4xy − 1 2xy + x
2 (3.2)
The dynamics of this map N = N0,0 is probably not substantially simpler than any other Na,b ,
but the formulas are simpler. In Chapter 6 we will show that all the same phenomena occur for
all Na,b .
The specific mapping N was studied by Russakovskii and Shiffman. In [RS], they study rational
mappings P : Pn ∼∼> Pn , whose topological degree is greater than the degree of P −1 (W ), where
W is a generic hyperplane. This is the case for the mapping above, with topological degree 4 and
algebraic degree 3. They prove that in this case there is an invariant measure; but in the case of
the mapping N they could not decide whether this measure charges the points of indeterminacy
at infinity. As an example of the untangling allowed by Farey blow-ups, we answer this in Section
3, showing that the points of indeterminacy are not charged, and in fact giving a fairly detailed
description of the measure.
50
Newton in C2 51
Proposition
in 3.1.1. The irreducible
³ ´ components of Dn correspond to the elements of the se-
quence Pn , in the sense that if pq ∈ Pn , then the curve in C2 parametrized by
³ p´
t 7→ ttq (3.4)
We will prove the stronger result (Proposition 3.1.2) below, providing curves that tend to all the
points of Dn , allowing us to parametrize these components. The underlying reason our construction
works is that the function xq y −p extends to LÃ ! ; this extension provides an isomorphism LÃ ! →
p p
q q
P1 . We will use this as follows: since p and q are coprime, there exist integers k and l such that
lq − kp = 1.
Proposition
in 3.1.2. The curve xq y −p = m is parametrized by
³ l p´
t 7→ mmt , (3.5)
k q
t
and the mapping
ml t p
³ ´
m 7→ lim mk tq (3.6)
t→∞
This shows that we can use (m1 , m2 ) as coordinates. The blow-up is being performed at the point
m1 = ∞, m2 = 0, so coordinates on the blow-up are provided by
m3 = m1 m2 and m2 . (3.9)
In terms of x and y, we have m3 = xq1 +q2 y −(p1 +p2 ) , and the curve
³ p1 +p2 ´
t 7→ ttq1 +q2 is the curve m3 = 1, m2 = 1/t, (3.10)
In Chapter 6, we will see how the primitive space must be modified when a and b are arbitrary; in
that case the axes have no particular dynamical significance, and we will need to use other curves
in their place.
Proposition
in 3.2.1. The Newton mapping extends as a rational map to the primitive space,
giving NP : XP ∼∼> XP , with four points of indeterminacy:
• the point of coordinate −2 in LÃ ! and the point of coordinate −1/2 in LÃ !;
2 2
1 1
Note that the points of the first pair are at the intersections of D∞ with the parabolas x+2y 2 = 0
and y + 2x2 = 0; each of these parabolas is one component of the inverse image of the x-axis and
the y-axis (the axes are tangent to a critical value parabola at a root, so the inverse images are
in agreement with Proposition 1.5.1). The points of the second pair are at the intersection of D∞
with the polar hyperbola of equation 4xy = 1.
Proof. The curve parametrized by
m l tp
³ ´
t 7→ mk t q , (3.12)
integer coordinates, and at least one coordinate positive, represent components of D∞ ), broken up
as shown in Figure 31 into five zones.
4 3
Figure 31. The marked points of the grid correspond to components of D∞ ; the five zones are distin-
guished by the shading.
In the different zones, different terms are dominant, and the four points on the boundaries of
zones correspond to the components of D∞ where there are competing terms.
The mapping in zone 1: p + q < 0 and p > 0
In zone 1, the dominant terms are
2m2l+k t2p+q
µ ¶
1
t 7→ . (3.14)
−1 m2l t2p
This breaks up into two cases, depending on whether q is even or odd.
If q is even, say q = 2q1 , then the image component is LÃ ! , and the coordinate in that
p+q1
p
component is
xp y −(p+q1 ) = (−1)q1 2p m−1 . (3.15)
Thus we see that the component LÃ ! is mapped by an isomorphism to the component LÃ !.
p p+q1
q p
However, the even power of t in both coordinates means that this component is a critical curve: a
small circle around such a component is mapped to a circle that wraps twice around the image com-
ponent, much closer than the original circle. This will be a key consideration when understanding
invariant measures.
54 Hubbard and Papadopol
If q is odd, the situation is quite different. Then the image component is LÃ !, and the
2p+q
2p
coordinate within this component is
Thus we see that the domain component is mapped by a rational map of degree 2 to its image,
with ramification at m = 0 and m = ∞.
1 ³
2t/m + t−2
´
−1 −2 2 2 . (3.17)
4/m − 1 2m t + t /m
Clearly the first term dominates in the first coordinate and the second term dominates in the
second coordinate, so LÃ ! maps to LÃ ! , with the point of coordinate m mapping to the point
1 1
−1 2
of coordinate
1 4 4
x2 y −1 = 4
(m2 ) = , (3.18)
4/m − 1 m m(4 − m)
which is a rational function of degree 2.
Let us look a bit more carefully at this rational function. The point of coordinate m = 4 in
LÃ ! is one of the two points where the polar curve of equation 4xy = 1 intersects D
∞ (the
1
−1
other is in LÃ ! ). Since the polar curve maps to the line at infinity, which is LÃ ! in the present
−1 1
1 1
notation, and LÃ ! does not intersect LÃ !, this will create a problem, and this point m = 4 will
1 1
2 1
be a point of indeterminacy of the extended Newton map.
4
How about critical points? The rational function m(4−m) has two critical points, at ∞ and
at m = 1/2. The latter is one of the intersections of the critical cubic C2 with D∞ , and maps
to the point where the critical value parabola of equation y = x2 intersects D∞ in LÃ ! . The
1
2
other critical point is accounted for, like all the critical points at 0 and ∞, by the fact that many
components of D∞ are themselves critical.
In the denominator the term tp+q dominates over the constant −1, otherwise again the first term
dominates in the first coordinate function and the second dominates in the second coordinate
function, so altogether we find as principal terms
ml tp /2
µ ¶
. (3.21)
ml−k tp−q /4
Thus the image of LÃ ! is the line LÃ !, and the point with coordinate m is mapped to the
p p
q p−q
point of coordinate
2
xp−q y −p = 2p+q m(p−q)l tp(p−q) m−p(l−k) t−p +pq
= 2p+q m−1 . (3.22)
We see that the line LÃ ! is mapped to the line LÃ ! by an isomorphism.
p p
q p−q
A computation just like the ones above shows that the parametrized curve
³ 2´
t 7→ mtt , (3.23)
leading to the point with coordinate m on LÃ !, maps to the curve with principal terms
2
1
2m2 t5 mt2 /2
µ ¶ µ ¶
1
t 7→ = , (3.24)
4mt3 2mt4 + m2 t4 (2 + m)t/4
with the two terms in the second coordinate in competition. Thus this line maps to itself, with
the point of coordinate m mapping to the point of coordinate
¶−2
mt2
µ
−2 4 8m
xy = = . (3.25)
2 (2 + m)t (2 + m)2
This rational fraction plays an important role in the dynamics of Newton’s method at infinity.
It is dynamically quite simple, similar to the Chebychev polynomial z 2 − 2. The Julia set is a
line segment, but instead of the complement being attracted to a superattractive fixed point, it is
attracted to a linearly attracting fixed point.
More specifically, the critical points are at m = ±2. The one at −2 maps to infinity, which maps
to 0, √
which is a repelling fixed point. The one at m = 2 is attracted to the attracting fixed point
at 2( 2 − 1). The Julia set is the negative real axis, and its complement is entirely attracted to
the attracting fixed point.
This is the dynamics of the rational map, but doesn’t quite jibe with the 2-dimensional dynamics
of N . More specifically, the critical point m = 2 is on the critical cubic C1 , and it maps to the
point m = 1 on the critical value parabola of equation x = y 2 . But the other critical point is much
more troublesome. It is on the parabola of equation x = −2y 2 , which maps to the x-axis, and
the point m = 0 on LÃ ! does not intersect the x-axis. As such, this point m = −2 is a point of
2
1
indeterminacy of the extended Newton map. We will need to blow it up again, and a lot of the
complications of the Newton map are hidden in this point.
56 Hubbard and Papadopol
The case of zone 3. In zone 3, the map is especially simple. There the first term dominates in
each´ coordinate, ³and the
´ 4xy dominates in the coefficient, so the principal terms of the image of
x are simply x/2 . Thus every line in zone 3 is mapped to itself by the identity. But notice
³
y y/2
the 2′ s in the denominator; they indicate that in XP , all these lines are repelling. Of course, we
knew all about that in the case of LÃ ! , from Section 2.3.
1
1
1 and −1 , are completely parallel to those
³ ´ ³ ´
The cases of zones 4 and 5 and of the points 2 1
already covered, and we only give the results.
The case of the line LÃ !. This line is mapped to itself, by the rational function m 7→
2
1
(2m + 1)2 /(8m), obtained from the case of LÃ ! by conjugating by m 7→ 1/m.
1
2
The case of zone 4. The line LÃ ! is mapped to the line LÃ ! by the isomorphism m 7→
p q−p
q q
2−(p+q) m.
The case of the line LÃ !. This line is mapped to LÃ !, by the rational map of degree 2
−1 2
1 1
m 7→ (4m − 1)/(4m2 ).
The case of zone 5. Again there are two cases. If p is even, say p = 2p1 , then the image
of LÃ ! is LÃ ! , and the mapping is given by m 7→ (−1)q1 2−p m−1 . Again in this case the
p p
q q+p1
p -component is critical. If p is odd, the image of LÃ ! is LÃ
³ ´
!, and the mapping within
q p 2p
q 2q+p
this component is m 7→ −2−2p m−2 .
will allow us to tease these various behaviors apart. Guided by the blow-ups, we will chop up a
neighborhood of infinity in C2 as follows. Let UK be the locus where sup{|x|, |y|} ≥ K.
The two lines LÃ ! and LÃ ! are both in zone 3, hence fixed, and normally repelling with
3 2
2 3
multiplier 2. As such, every point has an unstable manifold, by Proposition 2.3.4. The adherence
of the 3-dimensional locus of equation |x|2 = |y|3 in C2 is a circle in LÃ ! , which can also be
3
2
described as the locus |m| = 1, where m is the parameter on LÃ ! defined in Proposition 3.1.2.
3
2
There is a similar circle in LÃ !.
2
3
We will consider the unstable manifolds of these circles, i.e., the union of the unstable manifolds
of their points. These form two 3-dimensional manifolds, at least near infinity, and we can consider
the components M1 , M−1 of these manifolds in UK containing the circles, which form bundles of
discs over the circles.
The components LÃ ! and LÃ ! have a sequence of inverse images, alternately near the point
3 2
2 3
at infinity on the x-axis and y-axis. Let us call the Mi the intersection of UK with N −1 (M1−i ), so
that the Mi march monotonically from the point at infinity on the x-axis to the point at infinity
on the y-axis. These 3-manifolds cut up UK into regions Ak , k ∈ Z, as represented in Figure 32.
-14 -3 -4 -1 1 2
-6 3 -4 1 -2 3 -1 2 0 3 1 3
1 1 1 1 1 2 1
y 1
M3 2
polar -1
A3 1
hyperbola
-1
3
parabola M4 -4
x=-2y2,
inverse image 1
A4
of y-axis -2
M5 1
-3
A5 1
-4
3
M6
-14
1
-6
K x
Clearly, we can choose K large enough that any finite number of these inverse images are
well-behaved, but knowing that all the inverse images are well-behaved requires proof.
Lemma 3.3.1. For K sufficiently large, each Mi cuts UK into two pieces, one containing all the
Mj , j < i, and the other containing all Mj , j > i.
Proof. Choose K large enough that the statement is true for i = −2, −1, 1, 2. For all the others,
near the component of D0 adherent to Mi , the Newton mapping is either of the form
x 7→ ∗ x 7→ −y 2 .
³ ´ ³ ´ ³ ´ ³ ´
y −x 2 or y ∗ (3.27)
Thus in this region, UK maps well inside itself, and is locally separated by M1−i . It follows that
the inverse image Mi separates UK . ¤
The manifolds Mi now allow us to build a “Markov partition” of UK ; we label Ak the closures
of the components of the complement of the Mi , ordered so that
Ai is bounded by Mi and Mi+1 when i > 0,
Ai is bounded by Mi and Mi−1 when i < 0,
A0 is bounded by M1 and M−1 .
Moreover, we define Ak (C) ⊂ Ak to be the subset where k x
³ ´
y k ≤ C; in particular, if C < K,
then Ak (C) = Ak (K) = Ak .
for any C > K, the probability measure 14 N ∗ ν also satisfies this inequality.
Corollary
in 3.3.3. The Russakovskii-Shiffman measure σ does not charge the points p1 and p2 .
Proof of Corollary 3.3.3. The measure σ is obtained by giving weight 1 to a generic point (off
an exceptional pluripolar set), then giving weight 1/4 to its inverse images, then weight 1/42 to
its inverse images, etc. This defines a sequence of measures σ0 , σ1 , . . . ; Shiffman and Russakovskii
prove that the sequence converges.
Clearly we can choose σ0 so that it satisfies Equation (3.29), for instance by choosing the initial
point outside of UK . Then all σn also satisfy (3.29), and so does σ. But for any ǫ > 0, there exists
K ′ such that any measure satisfying (3.29) assigns measure < ǫ to the region UK ′ . ¤
Proof of Theorem 3.3.2. The proof will take six pages, so before giving it we will outline
what is really happening. The main content of the theorem is that on average, the initial points of
indeterminacy p1 and p2 are attracting. The problem is that this isn’t quite true: there are many
Newton in C2 59
inverse images near LÃ ! and LÃ !, and these are repelling. But they are only repelling linearly,
1 2
2 1
whereas elsewhere, more particularly near LÃ ! and LÃ !, the map is superattracting, at least
1 −1
−1 1
with an exponent 3/2. We chose our domains Ak carefully so that this would be the case; the point
of having a domain like A2 at all was to keep A3 away from LÃ !, which is again repelling. Then
1
0
it is not surprising that the superattracting behavior wins out over the linearly repelling behavior;
this fact is quite specifically illustrated by the K 1/4 in formula (3.38), which is really the key to
why the proof works at all.
We need to choose the numbers λk so that the pull-back of the measure µ itself is at most four
times itself. This is fairly easy for |k| ≥ 3, and for all k even, but choosing λ−3 , λ−1 , λ1 , λ3 is
quite difficult: because of the points of indeterminacy, there are a lot of inverse images near these
points.
We will find inequalities which express that the theorem is true in each Ak ; this is a different
problem in each Ak . We will start with the hardest case, A1 . Let ν be a probability measure on
R∞
C2 such that ν(Ak (C)) ≤ C µk (t) dt for all k ∈ Z.
Lemma 3.3.4. For any ǫ > 0, there exists K such that when C ≥ K, the measure N ∗ ν will satisfy
Z ∞
1 ∗
N ν(A1 (C)) ≤ µ1 (t) dt (3.30)
4 C
if
∞ ∞
√ √ √ X
p
1 X 2K(1 − ǫ2 )
(2 1 − ǫ − 2)λ1 > 2 λ2k+1 + p λ2k + . (3.31)
K(1 − ǫ) k=1 2
k=1
Proof. The region³ A´1 maps two-to-one to the union of all the Ak , k > 1, and in this region the
x-coordinate of N x
y is x/2 + o(|x|). In particular, if we choose K sufficiently large, we will have
¯³ ³ ´´ ¯
¯ |x|(1 − ǫ)
¯ N x
¯
y 1¯ ≥ , (3.32)
¯
¯ 2
where the index 1 on the left-hand side means the first coordinate. Since N (A1 (K)) is strictly
larger than ∪k>0 Ak (K), we must allow for the possibility that the pull-back measure may pick up
mass from outside UK ; this mass is at most 1/2, since ν is a probability measure, and a point has
√
at most two of its inverse images in A1 . We will represent this mass, rather wastefully, as a1 / C,
where a1 must be chosen sufficiently large that
p
a1 1 2K(1 + ǫ)
p ≥ , i.e., a1 = . (3.33)
2(K + ǫ) 2 2
60 Hubbard and Papadopol
Thus we find
∞ µ µ ¶¶
1 ∗ 1X C(1 − ǫ) a1
(N ν)(A1 (C)) ≤ ν Ak +√
4 2 2 C
k=1
Ã∞ Z ∞ ∞ Z ∞ !
1 X dt X dt a1
≤ λ2k+1 3/2
+ λ2k 2
+√
2 C(1−ǫ)/2 t C(1−ǫ)/2 t C
k=0 k=1
s ∞ ∞
(3.34)
2 X 1 X a1
= λ2k+1 + λ2k + √
C(1 − ǫ) C(1 − ǫ) C
k=0 k=1
s ∞ ∞
2 X 1 X a1
≤ λ2k+1 + √ λ2k + √
C(1 − ǫ) (1 − ǫ) CK k=1 C
k=0
Now let us deal with with A3 , which maps to A−1 and to a part of A0 ; which part depends on
K.
Lemma 3.3.5. For any ǫ > 0, there exists K such that the measure N ∗ ν will satisfy
1 ∗ ¡ ¢
(N ν)(A3 (C)) ≤ µ3 A3 (C) (3.37)
4
for C ≥ K if
λ0 λ−1
5/2
+ ≤ λ3 . (3.38)
K (1 − ǫ)3 1/4
K (1 − ǫ)3/4
Proof. The region A3 maps into A−1 ∪ A0 ; moreover, in the part of A3 that maps into A1 , the
second coordinate of the image is |x|3/2 + o(|x|3/2 ); see Figure 33. Thus for K sufficiently large we
have
¯³ ³ ´´ ¯
¯ N x 3/2
¯ ¯
y 2 ¯ > |x| (1 − ǫ) when |x| ≥ K. (3.39)
¯
¯
(The index 2 on the left-hand side indicates the second coordinate.) This map is 2 to 1 to its
image.
In the part of A3 that maps to A0 , which is a neighborhood of the polar hyperbola, the situation
is better yet. For one thing, this part of A3 maps to the part of A0 where |y| > |x|; moreover, the
second coordinate is |x|3 /8 + o(|x|3 ), so when K is sufficiently large we have
¯ |x|3 (1 − ǫ)
¯³ ³ ´´ ¯
¯ N x
¯
y 2¯ >
¯ when |x| ≥ K. (3.40)
¯ 8
This map is 1 to 1 to its image.
Newton in C2 61
1 1 2
3 2 3
1
C3 N(A3) 1
C2 N(A3)
q
p A0
C3/2
A-1
critical
value
parabola p’ 2
critical -1
q’
cubic
q”
p” 1
Polar -1
hyperbola A3
3
-4
K C
Figure 33. A schematic drawing of the mapping N in A3 . To understand its behavior in the real, we
must remember that one of the critical cubics passes through A3 , and that the image of A3 is bounded in
part by the critical value parabola y = x2 . The part of the neighborhood of the polar hyperbola where
|x| > C, shown in dark, maps into A0 , also shown in dark. The points p, q and their inverse images
p′ , p′′ , q ′ , q ′′ are drawn to help understand how the mapping can be 2-to-1 outside the neighborhood of the
polar hyperbola, and 1-to-1 inside it.
Of course, exactly the same results hold for λ−1 and λ−3 .
62 Hubbard and Papadopol
Proof. For part (a), observe that A0 maps 1 to 1 to A0 , and that when K is sufficiently large,
we have
1°° x °
³ ´° ° ³ ´°
x ° 1°° x °
³ ´°
° x °
°³ ´°
° y ° (1 + ǫ) ≥ °N y ° ≥ ° y ° (1 − ǫ) when ° y ° ≥ K. (3.47)
°
2 2
Again, we need to allow for the possibility that N ∗ ν picks up some mass from the mass of ν outside
UK . We will represent this mass as a0 /C, where a0 = K(1 + ǫ)/2. Then for any C < 2K(1 + ǫ),
i.e., for any C such that N (A0 (C)) may fail to be contained in UK , the term a0 /C is at least 1/4,
which is the maximum mass N ∗ ν/4 can pick up, since ν is a probability measure.
Thus
λ0 ∞
µ ³ ¶
1 ∗ ¡ ¢ 1 C(1 − ǫ) ´ a0 dt a0 λ0 a0
Z
N ν A0 (C) = ν A0 + ≤ + = + , (3.48)
4 4 2 C 4 C(1−ǫ)/2 t2 C 2C(1 − ǫ) C
which is indeed smaller than
∞
dt λ0 2a0 (1 − ǫ)
Z
λ0 2
= when λ0 > . (3.49)
C t C 1 + 2ǫ
The second part is a bit more delicate, and requires cutting A2 in two so that the part A′2 with
|y| ≤ 1/2 + o(1) maps to the part of A0 with |y| > |x|, and the part A′′2 with |y| ≥ 1/2 + o(1) maps
to the part of A0 where |x| > |y|. In both cases, the map is 1-to-1, and in both the image spills
out beyond UK . We will represent the mass that N ∗ ν picks up from the mass of ν outside UK by
a2 /C, where a2 is chosen so that
a2 1 a2 1
≥ and ≥ i.e., a2 = K(1 + ǫ). (3.50)
4K(1 + ǫ) 4 2K(1 + ǫ) 4
¯³ ³ ´´ ¯ |x|(1 − ǫ)
In A′2 , we have ¯ N xy 2¯ ≥ . Thus we find
¯ ¯
4
1 ∗ ¡ ′ ¢ 1 ³ ´ 1 µ ³ C(1 − ǫ) ´¶ a
′ 2
N ν A2 (C) = ν N (A2 (C)) ≤ ν A0 +
4 4 4 4 C
(3.51)
λ0 ∞ dt a2 λ0 a2
Z
= 2
+ = + .
4 C(1−ǫ)/4 t C C(1 − ǫ) C
Newton in C2 63
Finally, we come to the conditions on the λk , |k| > 3. These present no difficulty: Ak maps
well inside A2−k , there is no spillover to worry about.
Lemma 3.3.7. For any ǫ, there exists K such that the inequality
Z ∞
1 ∗ ¡ ¢
N ν Ak (C) ≤ µk dt (3.56)
4 C
is satisfied for C ≥ K if
λ2−k
λk ≥ if k is odd
(
√
K(1−ǫ)
(3.57)
λ2−k
λk ≥ K(1−ǫ)2 if k is even.
Proof. In all Ak with k > 3, the significant term of the Newton map is ¯ N x
¯³ ³ ´´ ¯
2 2
y 2 ¯ = |x| +o(|x| );
¯ ¯
Thus we find
1 ∗ ¡ ¢ 1 ³ ´ 1 ³ ³ ´´
N ν Ak (C) = ν N (Ak (C)) ≤ ν A2−k C 2 (1 − ǫ)2 . (3.59)
4 2 2
We now study separately the cases k even and k odd.
If k is even, we have
Z ∞
³ ³
2 2
´´ dt
ν A2−k C (1 − ǫ) ≤ λ2−k 2
C 2 (1−ǫ)2 t
(3.60)
λ2−k λ2−k
= 2 2
≤ .
C (1 − ǫ) KC(1 − ǫ)2
64 Hubbard and Papadopol
We are now in a position to prove Theorem 3.3.2: to find K, ǫ and λk , k ∈ Z such that all the
inequalities (3.31), (3.38), (3.44), (3.46), (3.57) are satisfied. Let us recopy them here:
∞ ∞
√ √ √ X
p
1 X 2K(1 − ǫ2 )
(2 1 − ǫ − 2)λ1 > 2 λ2k+1 + p λ2k + (3.31), again
K(1 − ǫ) 2
k=1 k=1
λ0 λ−1
λ3 ≥ + 1/4 (3.38), again
K 5/2 (1 − ǫ)3 K (1 − ǫ)3/4
K(1 − ǫ)2
λ0 > (3.44), again
1 + 2ǫ
λ0
λ2 > + K(1 + ǫ) (3.46), again
1−ǫ
( λ2−k
√
K(1−ǫ)
if k is odd, |k| > 3
λk ≥ (3.57), again
λ2−k
K(1−ǫ)2 if k is even, |k| > 3.
Let us consider first the associated homogeneous inequalities:
∞ ∞
√ √ √ X 1 X
(2 1 − ǫ − 2)λ1 > 2 λ2k+1 + p λ2k (3.64)
k=1
K(1 − ǫ) k=1
λ0 λ−1
λ3 ≥ 5/2
+ (3.65)
K (1 − ǫ)3 1/4
K (1 − ǫ)3/4
λ0 > 0 (3.66)
λ0
λ2 > (3.67)
1−ǫ
( λ2−k
√
K(1−ǫ)
if k is odd, |k| > 3
λk ≥ (3.68)
λ2−k
K(1−ǫ)2 if k is even, |k| > 3.
These can be satisfied as follows: first choose ǫ > 0 small
³ ´ (for instance 1/4), and K so large
that all the asymptotic inequalities for N are true for k x y k > K. Choose λ0 = 1, then λ2 so
that the fourth inequality is true, then λ3 so that the strict inequality
λ0
λ3 > 5/2
(3.69)
K (1 − ǫ)3
Newton in C2 65
is satisfied. Now choose λk , |k| > 3 so that the fifth inequalities are satisfied, and so that the series
in the first inequality is convergent. Now choose λ1 so that the first inequality is true. Next, choose
K larger yet so that the K 1/4 in the denominator in the second inequality allows it to be satisfied
too.
We now have numbers that satisfy the homogeneous inequalities. Multiply all the λ’s by a
sufficiently large constant so that all the inhomogeneous inequalities are satisfied also. ¤
4
The Farey blow-up
In this chapter, we will pin down exactly what operation was performed when we constructed
the primitive space. More specifically, we define a new way of “blowing up a surface” at a point
(actually, an enriched point), which we call the Farey blow-up. The Farey blow-up has some very
pleasant naturality properties, and is probably the only way of constructing a natural domain for
a rational map with both points of indeterminacy and critical curves. We will proceed to compute
the cohomology of the smooth part of the Farey blow-up, and describe the real oriented blow-up
of the Farey blow-up along its exceptional divisor.
X0 = X, E0 = C1 ∪ C2 (4.1)
1 1 2 1 3 2 3 1 4 3 5 2 5 3 4
4 3 5 2 5 3 4 1 3 2 3 1 2 1 1
Figure 34. The Farey divisor is topologically a chain of beads: spheres that touch on a Cantor set.
Any two smooth curves that intersect at a point can be represented locally as x = 0, y = 0 in
appropriate local coordinates, so Example 4.1.1 is a model for all Farey blow-ups.
Example 4.1.1. Let us compute the Farey blow-up of C2 at the origin, with respect to the x and
the y axes; so let 0̃ be the Farey data consisting of the origin and the two axes.
66
Newton in C2 67
In this case, the appropriate Farey tree is the obvious one, shown in Figure 35:
0 1
³ ´ ³ ´
1 0 F0
1
³ ´
1 F1
1 2
³ ´ ³ ´
F2
³ ´ 2 ³ ´ ³ ´ 1 ³ ´
1 2 3 3 F3
3 3 2 1
Figure 35.
µ ¶ To get the entry between two entries, the corresponding
µ ¶ coordinates
µ are
¶ added; for example,
the entry 1 0 1
2 on line F2 is gotten by adding the entry 1 on line F0 and 1 on line F1 ; the entry
µ ¶ µ ¶ µ ¶
3 on line F is gotten by adding the “parents” 2 on line F and 1 on line F .
3 2 0
1 1 0
Let us denote by Fn the vectors in the nth line (starting at F0 ). Evidently the irreducible
components of Xn can be labeled by the elements of Fn , as shown in Figure 36.
0
0 1
1
1
0 1
1
1
0
0
0 1 2
1
3
1
1 1
1 2 3
1 2
1 1 2
2 1 1 1
2 3
3 0
0
1 1
Figure 36. The first four stages in constructing the Farey blow-up F(C2 , 0̃). The components (actually
rational lines) are represented as line segments, labeled by the pairs of integers appearing in the first four
stages of the Farey tree.
Since p and q are coprime, the pair pq is completely specified by the slope p/q of the ray
³ ´
through it. To lighten the notation, and also to make a distinction with a slighly different con-
struction in Chapter 5, we will label the components of the Farey blow-up Lp/q .
68 Hubbard and Papadopol
Since p and q are coprime, there exist integers k and l such that lq − kp = 1. Then the limit as
t → 0 of the parametrized curve
t 7→ (ml tp , mk tq ) (4.2)
2
³ ´ in F(C , 0̃) and these limits are all distinct for m 6= 0. The number m parametrizes the
exists
p component of F(C2 , 0̃).
q
This provides a fairly complete description of the smooth points of F(C2 , 0̃), but there are other
points. We next give an embedding of the positive irrationals ³ into´ the ³Farey´ blow-up. Given a
positive irrational α, it lies between two consecutive elements qp 1 and q p 2 of Fn , in the sense
1 2
that p1 /q1 < α < p2 /q2 . Set xn (α) to be the point
xn (α) = EÃ ! ∩ EÃ !. (4.3)
p1 p2
q1 q2
Since πn (xn (α)) = xn−1 (α), the sequence x0 (α), x1 (α), x2 (α), . . . defines an element x(α) of the
Farey blow-up F(C2 , 0̃). It should be clear that the map R+ − Q → F(C2 , 0̃), α 7→ x(α) is
injective, since between any two distinct irrationals there are rationals.
Such points x(α) are extremely singular. For instance, every neighborhood of such a point x(α)
has infinite-dimensional homology. ³ p´
Notice that in analogy to the curves t 7→ ttq , the curve
t
³ ´
t 7→ tα (4.4)
Proposition
in 4.1.2. (a) The pairs of integers appearing as successive labels on D are neighbors.
(b) The space obtained by performing Farey blow-ups at all the double points of D is precisely
F(X, x̃).
(c) Conversely, the part of a Farey blow-up appearing between components with neighboring
labels can be blown down, and the quotient point is a smooth point, where the images of the
two components intersect transversally.
Proof. Part (a) follows from the fact that the Farey sum of two neighbors is a neighbor to both
its parents.
Part (b) is more or less obvious: any double point y = L1 ∩ L2 of D appears at some point in
the construction of F(X, x̃), since all neighbors are adjacent in some Fn , and the further blow-ups
that take place there to construct F(X, x̃) are precisely the ones involved in constructing F(Y, ỹ),
where ỹ is y enriched by L1 and L2 .
Part (c) is exactly undoing part (b). ¤
Remark. You cannot blow down any segment of a Farey divisor between components whose labels
are not neighbors without creating singularities. △
Theorem
in 4.2.2. Let X, Y be complex surfaces, and f : X → Y be a finite analytic map; we
are particularly thinking of maps with critical curves. Suppose that z̃ is an enriched point of
Y , such that f −1 (z̃) is set-theoretically a union of enriched points of X. Then f extends to a
continuous mapping F(X, f −1 (z̃)) → F(z̃); this mapping is analytic on the smooth points of
F(X, f −1 (z̃)).
Proof. The statement is local on X, so we may assume that f −1 (z) is a single point. The
hypothesis that f −1 (z̃) is set-theoretically a union of enriched points of X is very strong, and we
70 Hubbard and Papadopol
will see that we may choose coordinates in the domain and in the range so that the Farey data
consists of the coordinate axes in both the domain and range.
We will show that near f −1 (z) the map can be written
³ k´
f: x x
³ ´
y →
7 yl (4.6)
We need to say things in terms of ideals rather than in terms of equations defining the set because,
for instance, the origin in R2 is defined by the equation x2 + y 2 = 0, but the one function x2 + y 2
generates a much smaller ideal than I(Z), which is generated by x and y.
We can then define ÛZ ⊂ U × S m−1 to be the closure of the set
(x, p) ∈ U × S m−1 | f (x) 6= 0 and f (x)/|f (x)| = p .
© ª
(4.7)
By contrast, the blow-up is the closure of the set
(x, l) ∈ U × Pm−1 | f (x) 6= 0 and f (x) ∈ l ;
© ª
(4.8)
the main difference is that we are now allowing for the orientability of the line l ∈ Pm−1 .
This actually defines the oriented blow-up only locally, but a standard result from algebraic
geometry [Har] says that the blow-up has a universal property, which allows us to glue the local
blow-ups together uniquely. This result goes over to the oriented context, as is shown in [HPV1],
allowing us to define a space X̂Z locally isomorphic to the model above.
where π0 is the functor that associates to a space its set of connected components, and the projective
limit is taken over all open neighborhoods V ⊂ X of y. The space E(X, Y ) comes with a natural
topology, which we will leave to the reader to define (see [Dou], Vol. II, p.197–198), and there is a
canonical map p : E(X, Y ) → X. We will denote by E(Y ) ⊂ E(X, Y ) the subset p−1 (Y ).
Definition
in 4.3.1. Let X be a real algebraic manifold, and Z ⊂ X an algebraic subset. The
real oriented blow-up BR∗ (X, Z) is the endpoint modification E(X̂Z , Ẑ), with the exceptional
divisor BR∗ (Z) ⊂ BR∗ (X, Z) being E(Ẑ) ⊂ E(X̂Z , Ẑ).
We want to understand the relation between the real oriented blow-up BR+ (X, D) and the real
oriented blow-up BR+ (F(X, x̃)), F(x̃) ∪ D).
To lighten notation, we will call this object B + F(X; x̃, D). It isn’t a priori clear what this is,
and we must take the projective limits in the proper order. Recall that F(X, x̃) is naturally the
pro-algebraic space
lim (Xn , πn ). (4.11)
←
Each Xn contains a divisor En , which with our definition contains both C1 and C2 . Let us define
Dn to be En ∪ D, from which C1 , or C2 , or both have been removed if they do not belong to D.
Here, removing a curve means taking the closure of the complement of the proper transform of
that curve.
The real oriented blow-ups BR+ (Xn , Dn ) form projective systems by Theorem 5.9 of [HPV]. We
then define
B + F(X; x̃, D) = lim BR+ (X, Dn ). (4.12)
←
To make a model for this object, we first define a way of taking the interval (0, ∞), and “opening
all the rational numbers to intervals.” To be more precise, consider the functions η1 , η2 : [0, ∞) → R
given by
X X
η1 (t) = e− sup(p,q) and η2 (t) = e− sup(p,q) . (4.13)
0<p/q<t,(p,q)=1 0<p/q≤t,(p,q)=1
Then the map t 7→ It = [η1 (t), η2 (t)] associates to every rational number a closed interval, and to
every irrational number a single point. These intervals and points are all disjoint. The limit
lim η1 (t) = lim η2 (t) = a (4.14)
t→∞ t→∞
is finite, so we may think of η1 and η2 as being defined on [0, ∞], and giving a homeomorphism of
“[0, ∞] with the positive rationals opened” to J = [0, a].
We will put an appropriate “real oriented blow-up structure” on J × (R/Z)2 , by defining a circle
action ∗ on the zones Ip/q × (R/Z)2 . If t ∈ Ip/q , we will let the circle R/Z act on {t} × (R/Z)2 by
¡ ¢ ¡ ¢
s ∗ t, (α, β) = t, (α + ps, β + qs) . (4.15)
Because the integers (p, q) are coprime, this action is free, and the orbits of the action are
(p, q)-torus knots. If pn /qn converges to an irrational number ω, then the orbits of the circle action
converge to leaves of the irrational foliation of slope ω on the torus η1 (ω) × (R/Z)2 .
For the remainder of this section, we will assume that C1 and C2 are contained in D.
Let us consider the real oriented blow-up p : BR+ (X, D) → X of X along D. This is a 4-
dimensional manifold with boundary; the part of the boundary above C1 − {x} and above C2 − {x}
carries a circle action, and the fiber above x is a torus T , which comes with an action of (R/Z)2 ,
so that the limit of the circle action on the part above C1 becomes the action of the first factor,
and the limit of the circle action on the part above C2 becomes the action of the second factor, as
shown on the left of Figure 37.
Choose a neighborhood U = U1 ∪ U2 of x in C1 ∪ C2 ; and identify p−1 (U ) with J × (R/Z)2 , so
that on the part of boundary above ∂U1 the action of the first factor of (R/Z)2 coincides with the
circle action there, and so that on the part of boundary above ∂U2 the action of the second factor
of (R/Z)2 coincides with the circle action there.
Newton in C2 73
Figure 37. Left: a torus, with circle actions on the inside and the outside. Right: the torus now thickened
into a Cantor set of concentric tori, with zones between them corresponding to the rational numbers. One
should imagine these zones also filled with circle orbits, forming torus knots of all different kinds.
Now let us call Z the set BR+ (X, C1 ∪C2 ), with the original circle actions on BR+ (C1 ∪C2 )−p−1 (U ).
Within p−1 (U ), put the circle actions on the parts corresponding to rational numbers, as described
in Equation (4.15).
Theorem
in 4.3.2. There is a homeomorphism
h : Z → B + F(X; x̃, D) (4.16)
that is the identity off a neighborhood of the modified T , and which preserves the circle and
torus actions.
¡ ¢
Proof. Consider the intervals JN , where all the components of J − ∪p/q∈F
/ N Ip/q have been
collapsed to points. There is a natural projective system
→ JN → JN −1 · · · → J0 = pt, (4.17)
and the projective limit is precisely the interval J.
There are corresponding quotients JN × (R/Z)2 → JN −1 × (R/Z)2 , and using the identification
above, we can define spaces ZN , together with maps forming a projective system
· · · → ZN → ZN −1 → · · · → Z0 (4.18)
with projective limit Z.
So the theorem will follow from giving compatible homeomorphisms
... → ZN → ZN −1 → ...
↓ hN ↓ hN −1 (4.19)
... → BR+ (XN , EN ) → BR+ (XN −1 , EN −1 ) → . . .
These will be constructed by induction. To start the induction, we choose a neighborhood U ′
of x in X, containing U . Then we can find a mapping
BR+ (X0 , E0 ) → BR+ (X0 , E0 ) (4.20)
that collapses p−1 (U1 ∪ U2 ) to the torus p−1 (x) = (R/Z)2 , which is otherwise injective, and the
identity outside of p−1 (U ′ ). This defines a homeomorphism BR+ (X0 , E0 ) → Z0 .
Now the construction of hN from hN −1 is precisely what is done in [HPV], Theorem 5.9. ¤
There is a straightforward modification of Proposition 4.3.2 in the case where D contains only
one of C1 or C2 , or neither.
The cases where D contains either C1 or C2 are identical. If it contains C2 , we start with
p : BR+ (X, C2 ) → X; the fiber p−1 (x) is now a circle, and the first thing is to thicken it to a solid
74 Hubbard and Papadopol
torus, and to identify this solid torus with D × R/Z, so that the R/Z-action on ∂D × R/Z coincides
with the given circle action. Use polar coordinates (ρ, θ) in D, with ρ ∈ J. Then in the regions
ρ ∈ Ip/q we can define a circle action as above, and the resulting space is our candidate for a model
of B + F(X; x̃, C2 ). The main difference is that the fiber above the point where C1 intersects E∞
is now a circle, not a torus.
If D contains neither C1 nor C2 , the topological model consists of removing a small 4-ball around
x, bounded by a 3-sphere. Then E1 is a rational curve, and BR+ (E1 ) ⊂ BR+ (X1 , E1 ) maps to E1 by
a map conjugate to the Hopf fibration. Represent E1 again in polar coordinates, where the radius,
which now lives in [0, ∞], should be thought of as belonging to J. Now repeat the construction
above; the points where E∞ intersects C1 and C2 both have circles above them, linking in BR+ (E1 ).
Theorem
in 4.4.1. (a) Let X, Y be complex surfaces, and f : X → Y be a finite analytic map.
Suppose that z̃ is an enriched point of Y , such that f −1 (z̃) is set-theoretically a union of enriched
points of X. Let D be a divisor in Y , which contains z only if it contains C1 or C2 or both.
Then f lifts to a continuous mapping
B + F X; f −1 (z̃), f −1 (D) → B + F(Y ; z̃, D).
¡ ¢
(4.21)
(b) This map is a covering map if D contains both C1 and C2 . If C1 is not in D, then the
proper transform of C1 is bounded by a circle in B + F(Y ; z̃, D); the inverse image of that circle
is critical, mapping to its image with the same local degree as the degree of f on C1 . The
analogous statement holds for C2 . These are the only places where the map fails to be a local
homeomorphism.
Proof. The statement is local in the domain, so we can specialize to the case of Example 4.1.1,
i.e., to the case where f = fà ! .
k
l
There is no problem at the smooth points of the Farey divisor: if a component is critical of
normal degree d, then circle orbits map to circle orbits with degree d. So let us consider a non-
algebraic point, corresponding to “slope” r, in the range. If r is rational, it corresponds to either
the upper or the lower end of a component; let us suppose the upper end if that happens. At each
level of the Farey tree, the real number r is then bracketed between two rational numbers:
′ p′N p′′N ′′
rN = ′ ≤ r < ′′ = rN . (4.22)
qN qN
The lower bracketing sequence will be eventually constant if r is rational.
A basis of neighborhoods of the non-algebraic point is then provided by the sets
′
¯ ¯rN ′
¯ ¯rN
¯x¯ 1 ¯x¯
¯ ¯ < , ¯ ¯ > ǫ,
¯y¯ ǫ ¯y¯ (4.23)
| arg x − λ1 | < ǫ, | arg y − λ2 | < ǫ, |x|2 + |y|2 < ǫ.
Newton in C2 75
³ k´
The inverse image of such a neighborhood by the mapping x x
³ ´
y →
7 y l is evidently the union
of kl disjoint sets, each of which contains such a set for a smaller ǫ. So the mapping on the real
oriented blow-up is continuous, of constant degree kl. ¤
Remark. It isn’t clear what the homology of the Farey blow-up itself is (as opposed to its
smooth points). This space has the sort of complication present in the Hawaiian earring, so we can
expect singular and Cech homology to be different. In fact, it seems more reasonable to study the
cohomology of F(X, z̃), but we will not consider this problem. Instead, we will develop a notion
of completed homology, which is a Hilbert space and hence isomorphic to its dual, and which will
serve as both homology and cohomology. △
Let X is a complex surface, and z̃ be an enriched point of X. We will denote by F ∗ (X, z̃) ⊂
F(X, z̃) the subset of smooth points. The space F(X, z̃) is a projective limit of spaces Xn with
divisors En . Let us denote Xn∗ the surface Xn from which the double points of En have been
removed. There are then natural inclusions
X0∗ ⊂ X1∗ ⊂ X2∗ ⊂ . . . , (4.24)
and the increasing union is F ∗ (X, z̃), which is thus represented as an inductive limit.
A first thing to realize (see Propositions 4.5.1 and 4.5.2) is that the homology is naturally a
subgroup of
∗
H2 (X) ⊕ ZIrr E∞ = H2 (X) ⊕ ZQ+ , (4.25)
where the second factor is the product (not the sum) of copies of Z, i.e., it consists of giving
arbitrary integer weights to the components of E∞ , which are naturally labeled by the positive
rationals Q∗+ . Of course, it isn’t the entire group, for instance because it is countable, and the
product group is uncountable. But neither is it H2 (X) ⊕ Z(Irr E∞ ) , where all but finitely many
coefficients vanish.
For the next proposition, it is essential to realize that removing a point (or finitely many points)
from a complex surface does not change its second homology group; in particular the inclusion
Xn∗ → Xn induces ∗
³ ´an isomorphism H2 (Xn ) = H2 (Xn ). This has the consequence that the compo-
nents Lp/q for pq ∈ Fn represent homology elements [Lp/q ] of H2 (Xn∗ ).
Proposition
in 4.5.1. For each n, the homomorphism
³ ´ X
H2 (X) ⊕ ZFn → H2 (Xn∗ ), c, (αp/q )p/q∈Fn 7→ i∗ c + αp/q [Lp/q ] (4.26)
p/q∈Fn
is an isomorphism.
Every element of this inductive limit is represented by a sequence vm ∈ ZFm , vm+1 ∈ ZFm+1 , . . .
starting with some m.
Proposition
in 4.5.2. The weight assigned by vk to elements of Fl with l < k is independent of
k.
Proposition 4.5.2 provides an embedding of H2 (F ∗ (X, z̃)) → H2 (X) ⊕ ZIrr E∞ , but we will
describe the image much more precisely. Let Q be the first quadrant in R2 .
Definition
in 4.5.3. Let S be the commutative group of functions f : Q → R that
• are continuous;
• are piecewise-linear, linear except on finitely many rays through the origin with rational
slope;
• take on integer values on Q ∩ Z2 ; and
• vanish on the axes.
This group is a free commutative group on generators ep/q , p/q ∈ Q∗+ , one for each positive
rational number, defined as follows. Each rational number first appears in the Farey tree as the
Farey sum
p p1 + p2
= (4.28)
q q1 + q2
of two numbers that appeared earlier (its parents). Let ep/q ∈ S be the function that is linear
except on the rays of slope through pq1 , pq , and pq2 , and which assigns 1 to pq and 0 to
³ ´ ³ ´ ³ ´ ³ ´
1 2
p1 , p 2 , 0 , 1 .
³ ´ ³ ´ ³ ´ ³ ´
q q 1 0 (4.29)
1 2
ep/q , pq ∈ Fn − F0
³ ´
(4.30)
belong to it. We need to know that they span, or equivalently, that they are linearly independent.
Indeed, if a linear combination is 0, the the coefficients corresponding to each element of Fn − Fn−1
must vanish, since otherwise the linear combination would not be linear at that element of Fn −
Fn−1 . But then, for the same reason, the coefficient of each element of Fn−1 − Fn−2 must vanish,
etc.
Thus the ep/q , pq ∈ Fn are linearly independent in Sn , so that every element of Sn can be
³ ´
written
X
ap/q ep/q , (4.31)
à !
p ∈Fn −F0
q
Newton in C2 77
with rational coefficients ap/q , and we need to show that these coefficients are all integers. Suppose
not, and let m ≤ n be the smallest index for which there exists pq ∈ Fm with ap/q not an integer.
³ ´
′
This´ is a contradiction, since all the basis functions corresponding to indices m > ³ m´ vanish at
p , as do as the others of index m, and since all the earlier ones give integers at p , .
³
q q
Clearly together the ep/q form a basis of lim Sn = S. ¤
→
There is an evident isomorphism Sn = ZFn , simply giving the values at the elements of Fn ,
and this isomorphism induces a canonical projection S → Sn given by restricting a function f
to the points of Z2 that are in Fn . There is also a natural inclusion Sn → S, which consists
of extending linearly to each sector bounded by rays through points of Fn . This map induces
mappings in : Sn → Sn+1 .
Let X be a complex surface, and z̃ be an enriched point of³X.´ We define a homomorphism
Φ : S → F ∗ (X, z̃) as follows. Choose p/q ∈ Q∗+ , and suppose pq ∈ Fn − Fn−1 . There is then
an exceptional divisor Lp/q ⊂ Xn labeled pq . Let φp/q : P1 → Lp/q be an isomorphism, and let
³ ´
φ′p/q : P1 → Xn∗ be a small deformation whose image avoids the double points.
Remark. You can only deform differentiably: analytically, exceptional divisors are rigid. The
deformation can take place in an arbitrarily small neighborhood of Lp/q , but the deformed 2-sphere
must intersect Lp/q since this curve has self-intersection −1. △
Then φ′p/q : P1 → F ∗ (X, z̃) maps the fundamental class in H2 (P1 ) to a well-defined element
Φ(ep/q ) ∈ H2 (F ∗ (X, z̃) (this is the definition of Φ). For Theorem 4.5.5, we need one more general-
ity: the inclusion X − {z} → X induces an isomorphism H2 (X − {z}) → H2 (X); using its inverse
we see that there is a canonical map H2 (X) → H2 (F ∗ (X, z̃)).
Theorem
in 4.5.5. The mapping Φ and the canonical map H2 (X) → H2 (F ∗ (X, z̃)) induce an
isomorphism
H2 (X) ⊕ S → H2 F ∗ (X, z̃) ,
¡ ¢
(4.32)
which we will still call Φ.
Proof. The space F ∗ (X, z̃) is the inductive limit of the Xn∗ , and since homology commutes with
inductive limits, we see that what we need to prove is that for every n, the map
Sn ⊕ H2 (X) → H2 (Xn∗ ) (4.33)
is an isomorphism, and that the diagrams
H2 (X) ⊕ Sn → H2 (Xn∗ )
↓ ↓ (4.34)
∗
H2 (X) ⊕ Sn+1 → H2 (Xn+1 )
all commute.
This is proved by induction, starting with n = 0, where it is the isomorphism H2 (X) = H2 (X −
{z}), and going from step to step using Proposition 4.6 of [HPV]. ¤
Give S a quadratic form, by making distinct ep/q orthogonal, and setting e2p/q = −1. The spaces
H2 (X) and F ∗ (X, z̃) both carry quadratic forms defined by intersection, and these are all related
by the following statement.
78 Hubbard and Papadopol
Theorem
in 4.5.6. If we take H2 (X) ⊕ S to be an orthogonal direct sum, then the isomorphism
Φ : H2 (X) ⊕ S → H2 F ∗ (X, z̃) .
¡ ¢
(4.35)
is an isomorphism of spaces with inner products.
Proof. Begin by representing all elements of H2 (X) by cycles that avoid a neighborhood of z; that
is really the meaning of the isomorphism H2 (X) = H2 (X − {z}). Then our first exceptional divisor
L1/1 is orthogonal to all of H2 (X), and has self-intersection −1, being an exceptional divisor. Now
the statement should be clear from the way Φ was defined. Every homology class, when it is
created, is disjoint from all the previous ones, and has self-intersection −1. Of course, the same
thing is true of φ′p/q (P1 ), thus none of these classes intersect any other class, and they all have
self-intersection −1. ¤
act on the homology. We will work this out only for the mapping fà !, which we will denote
2
1
simply by f for the rest of this chapter. This turns out quite nicely, but a general proof along the
lines of the one given here would be very cumbersome.
Let us denote fQ : R2 → R2 the map
fQ : pq = 2p
³ ´ ³ ´
q . (4.37)
Theorem
in 4.6.1. The map induced by f on the homology is
f˜∗ ep/q = 2ep/q ◦ fQ
−1
¡ ¢
. (4.38)
Proof. The idea is to work at a level of blow-up where ep/q has just appeared. Let the Farey
parents of pq be pq1 and pq2 , and assume that q1 is even and q2 is odd. (The case q2 even,
³ ´ ³ ´ ³ ´
1 2
q1 odd is identical, and q 1 and q 2 can’t
³ both´ be even; we will deal with the case where both are odd
p 2p
³ ´
1
later.) Observe that q /2 and q 2 are then neighbors, and we can consider the sequence of
1 2
two blow-ups in the domain and range represented in Figure 38.
Note that the map µ f extends ¶ at the top and the bottom level, but not at the middle level, where
p1 + p2
the divisor labeled is collapsed to the point marked with a circle, except for one point
q1 + q2
of indeterminacy, marked with a star.
The class ep/q is represented by the line Lp/q after one blow-up, and by Lp/q + L(p1 +2p2 )/(q1 +q2 )
after the second blow-up. This maps to
2L(2p1 +2p2 )/(q1 +q2 ) + L(p1 +2p2 )/(q1 /2+q2 ) . (4.39)
Newton in C2 79
Domain Range
p1
p1
q1/2
q1
p2
q2 2p2
q2
p1 p1
q1 q1/2
p1+p2 p1+2p2 2p2
p2 q1/2+q2
q1+q2 q2
q2
p1 p1+2p2
q1 p1
p1+p2 q1+2q2
q1/2
q1+q2
p1+2p2 p1+2p2
q1+2q2 q1/2+q2
p2 2p2
q2 q2
Figure 38. The map f can be locally extended before doing any blow-ups (i.e., at the top level), and
after doing two blow-ups (i.e., at the bottom level), but it cannot be extended at the middle level (after
just one blow-up).
On the other hand, the class e(p1 +2p2 )/(q1 /2+q2 ) is represented by L(p1 +2p2 )/(q1 /2+q2 ) after one
blow-up in the range, and by
L(2p1 +2p2 )/(q1 +q2 ) + L(p1 +2p2 )/(q1 /2+q2 ) (4.40)
after two blow-ups.
Thus the image of ep/q is represented by
e(p1 +2p2 )/(q1 /2+q2 ) + L(2p1 +2p2 )/(q1 +q2 ) = e(p1 +2p2 )/(q1 /2+q2 ) + e(2p1 +2p2 )/(q1 +q2 ) , (4.41)
since L(2p1 +2p2 )/(q1 +q2 ) is the last exceptional divisor.
This is a formula for the action of f˜Ã ! on homology (at least in the unramified case), but it
2
1
isn’t quite clear that it is what the statement promised. Before dealing with that, let us find the
analogous formula if both q1 and q2 are odd.
This is covered by Figure 38 also: both parents of pq1 + 2p2 have odd second coordinates;
³ ´
1 + 2q2
we will leave it to the reader to see that all lines with such labels appear in this setting. Here
L(p1 +2p2 )/(q1 +2q2 ) is an exceptional divisor created by the most recent blow-up, so it represents
(after two blow-ups) the class e(p1 +2p2 )/(q1 +2q2 ) , and it maps to
L(p1 +2p2 )/(q1 /2+q2 ) = e(p1 +2p2 )/(q1 /2+q2 ) − e(2p1 +2p2 )/(q1 +q2 ) . (4.42)
Now we need to understand these elements of the commutative group S as functions (see Def-
inition 4.5.3). We will restrict our functions to the line q = 1 to draw them. The functions ep/q
80 Hubbard and Papadopol
³ ′´
restrict to tent functions, of course, but the miracle that makes things work is that if pq ′ is a
child of pq , then ep/q ± ep′ /q′ is also a tent function; the slopes are exactly what is needed for
³ ´
q1+q2
e 2p1+2p2 e p1+2p2 e 2p1+2p2 e p1+2p2
q1/2+q2 q1/2+q2
q1+q2 q1+q2
Figure 39. Left: the sum e(p1 +2p2 )/(q1 /2+q2 ) + e(2p1 +2p2 )/(q1 +q2 ) is the tent function
−1
2e(p1 +p2 )/(q1 +q2 ) ◦ fQ .
Right: the difference e(p1 +2p2 )/(q1 /2+q2 ) − e(2p1 +2p2 )/(q1 +q2 ) is the tent function
−1
2(e(p1 +2p2 )/(q1 /2+q2 ) ) ◦ fQ .
This is already a fairly familiar space, but quite surprisingly, it is also naturally an old friend
from analysis, the Sobolev space H01 (I) of continuous functions on g on [0, 1], whose distributional
derivative is in L2 , and such that g(0) = g(1) = 0. We use a script H to avoid confusion with
homology; the subscript 0 indicates that the function vanishes at the endpoints.
t
³ ´
1 − t , and give it the measure |dt|, so it has length 1. The restriction to I maps S to a space
of functions on I. For future reference, we will specify what the image is.
Newton in C2 81
Proof. This should be clear from Lemma 4.5.4: if you extend a function g as above to be
homogeneous of ³degree
´ 1, you will find a continuous, piecewise linear function on the quadrant q,
p
and its value at q will be (p + q)g(p/q). ¤
Theorem
in 4.7.2. The map f 7→ f |I induces an isometry of Hilbert spaces
Ψ : Ŝ → H01 (I). (4.46)
Proof. Notice that the functions f |I are tent functions. First let us see that Ψ(ep/q ) and Ψ(ep′ /q′ )
are orthogonal if p/q 6= p′ /q ′ . By their definition, either the supports of Ψ(ep/q ) and Ψ(ep′ /q′ ) are
disjoint, or the support of one, say Ψ(ep/q ), is contained in an interval where the other is linear,
hence its derivative is constant. Thus what we need to show is that
d
Z
t
³ ´
I dt
ep/q 1 − t |dt| = 0. (4.47)
t
³ ´
(The function ep/q 1 − t is represented in Figure 40.)
1
p1+q1+p2+q2
p1 p2
p1+q1 p2+q2
p1+p2
p1+q1+p2+q2
µ ¶
Figure 40. The function ep/q t
1−t .
This shows that Ψ is an isometric inclusion. Now we need to show that the image of S is dense,
but that is fairly clear from Lemma 4.5.4: the derivatives of elements of S give all step-functions
on I with rational discontinuities and integral 0. These are certainly dense in the hyperplane in
L2 (I) made up of functions with integral 0; for instance, they include the Haar basis. ¤
There is quite a bit of freedom in this description, corresponding to the freedom in Proposition
4.1.2; we will need only the following special case.
Proposition
in 4.7.3. (a) If you draw the vertical ray Rc defined by p = c, q > 0 in Q, and give
it the measure |dq|/c, then the restriction of elements of S to Rc is an isometric inclusion of
S into H1 (Rc ), and the³image 1
´ of S ⊗Z C is dense in the subspace of H (Rc ) of functions that
vanish at the endpoint 0c of Rc .
(b) Consider the mapping g : Rc → Rc given by g(q) = q/2. Then
(fQ )∗ f |Rc = g∗ f |Rc . (4.50)
Proof. We need to show that the restrictions of the ep/q are an orthonormal family; this is a
computation just like the one to justify Equation (4.49), and we omit it. The density follows, for
instance, from the fact that S contains the Haar functions. ¤
5
The compactification when a = b =
In this chapter, we will construct a compact space X∞ on which the Newton map N of Chapter
3 is well defined. We will further analyze the topology of this space, showing that its behavior at
infinity is very reminiscent of the KAM theorem [Ar].
Here we are denoting components of D1 with vector indices, not fractions, because the primitive
points of QP are not quite specified by their slopes.
We will meet these points again, so we give them names: A1 , B1 , B2 , A2 .
• The polar hyperbola 4xy = 1 and the component LÃ ! will enrich A1 , giving Ã1 ;
1
−1
83
84 Hubbard and Papadopol
• The parabola x = −2y 2 and the component LÃ ! will enrich B1 , giving B̃1 ;
2
1
• The parabola y = −2x2 and the component LÃ ! will enrich B2 , giving B̃2 ;
1
2
• The polar hyperbola 4xy = 1 and the component LÃ ! will enrich A2 , giving Ã2 .
−1
1
We set Z̃1 to be the union of these four enriched points and the three ordinary points q1 , q2 , q3 ,
and define X2 = F(X1 , Z̃1 ). It is not quite so easy to define N2 : X2 → X1 .
Proposition
in 5.1.1. The Newton map N extends to a mapping N2 : X2 → X1 .
Proof. There is no difficulty on E2 , the blow-ups of q1 , q2 , q3 ; this was done in Lemma 1.2.4.
The difficulty is to show that N extends to the Farey divisors corresponding to the enriched points
Ã1 , Ã2 , B̃1 , B̃2 . In all cases, this follows from Theorem 4.2.2, but one must make some preliminary
constructions to see how.
Begin by performing just three blow-ups on P2 : one at p1 , to create LÃ ! ; a second at the
1
0
intersection LÃ ! with the line at infinity (here called LÃ ! ), to create the divisor LÃ !; and
1 1 2
0 1 1
finally, a third at the point of coordinate m = −2 on LÃ ! , to create a line we will temporarily
2
1
call simply L. Call the surface constructed Y . We are now in a position to apply Theorem 4.2.2.
Indeed, the Newton map extends to a neighborhood of the union LÃ ! ∪ LÃ ! ∪ L, mapping
1 2
0 1
LÃ ! to itself by the rational function given by Equation (3.25). The line L is critical, mapping
2
1
to LÃ !, and finally LÃ ! maps to the line at infinity.
1 1
0 0
In particular,
• the point x̃1 where L meets LÃ !, enriched by L and LÃ !, maps to the point ỹ1 of coordinate
2 2
1 1
0 on LÃ !, enriched by LÃ ! and LÃ !;
1 1 2
0 0 1
• the point x̃2 where L meets the proper transform parabola of equation x + 2y 2 = 0, enriched
by L and the proper transform of the parabola, maps to the point ỹ2 of coordinate ∞ on LÃ ! ,
1
0
enriched by LÃ ! and the x-axis.
1
0
Next, we need to extend the Newton mapping to the Farey divisor at Ã1 ; this is easier. Perform
in the domain P2 two blow-ups at p1 , to construct first LÃ ! and then LÃ ! . In the range
1 1
0 −1
P2 , perform two blow-ups at p2 , to construct LÃ ! and LÃ ! . Then the Newton map is a local
0 1
1 2
isomorphism from the enriched point Ã1 to LÃ ! ∩ LÃ ! , enriched by the lines themselves. Again,
1 1
2 1
Theorem 4.2.2 now provides the extension; note that the Farey divisor in the range is a sub-Farey
divisor of the divisor associated to p̃2 , so the range is X1 .
The constructions for B̃2 and Ã2 are completely parallel. ¤
-1 0
1 1
1
2
1
N
1 0
p
y-axis 11 q
1 2
3
0
Py
1
Figure 41. This represents how the Farey blow-up of B̃2 maps to X1 .
Then the point with these coordinates tends to the point of coordinate m of the pq -component
³ ´
of the Farey blow-up as s → 0. Set s = 1/t; changing back to the (x, y) coordinates, we see that
the curve
tp+q t2p+q
x= k , y = (5.3)
m (−2tq + ml ) m2k (−2tq + ml )
Remark. These formulas represent x and y as Puiseux expansions in t with two terms. This is
typical: as we blow up more and more times, paths leading to the deeper blow-ups need to be
sneakier and sneakier, and require more and more terms. No finite Puiseux expansion gives paths
that lead to the points of the projective limit; these require series, and the curves leading to them
are necessarily transcendental. This sort of thing was explored earlier (for Hénon mappings) in
[Bou], [BS].
The image of this curve is
2t4p+3q t4p+2q
+ m4k (−2t
à !
m3k (−2tq + ml )2 m4k (−2tq +ml )3 q +ml )2
. (5.4)
4t3p+2q − m3k (−2tq + ml )2 2t5p+3q
+ t2p+2q
m5k (−2tq +ml )3 m2k (−2tq +ml )2
The characteristic of the curve tending to a point of indeterminacy is a cancellation of principal
terms, specifically in the expression
2t4p+3q t4p+2q t4p+2q 2tq
µ ¶
+ 4k = 4k +1
m4k (−2tq + ml )3 m (−2tq + ml )2 m (−2tq + ml )2 −2tq + ml
(5.5)
ml t4p+2q
= 4k .
m (−2tq + ml )3
Other than this, we only have to find the principal terms of everything that remains, which
gives
à ml−4k t4p−q ! µ l−k
m3k − − m 8 tp−q
¶
8
t5p
= . (5.6)
t3p − 4m 5k
− 4m12k t2p
³ We´ see that the situation is different for p − q odd and for p − q even. If p − q is odd, the
p -component of the Farey blow-up maps to LÃ ! by the map m 7→ −1/(24p+2q m2 ), and
q p−q
2p
the component is not critical.³ ´
If p − q is even, then the pq -component of the Farey blow-up maps to LÃ ! by an
(p − q)/2
p
isomorphism, and the component is critical.
Further extensions
From here on, the extensions are easy. The inverse image Z̃2 = N2−1 (Z̃1 ) is Farey data on X2 , and
so forth. There is one irritant that must be mentioned: we started with the three points q1 , q2 , q3
in the finite plane; these are part of Z1 . Then the inverse images of these points are part of Z2 ,
the inverse images of these are parts of Z3 , and so forth.
It may happen that at some stage, one of these points z ∈ Zk−1 may lie on a critical value
parabola; the values of the parameters where this occurs form a countable union of sets of codi-
mension 1, and may well be dense in the parameter space.
If that happens, we cannot simply blow up the inverse images Nk−1 (z) in the critical cubic. As
we saw in Example 4.2.1, the Newton map does not then extend to the exceptional divisor; instead,
it blows it down to a point, except for one point of indeterminacy (corresponding to the kernel of
the derivative DNk (Nk−1 (z)), which must be blown up again. We will deal with this by giving the
points Nk−1 (z) an enhanced structure that allows us to define this double blow-up; in this case the
enhanced structure is a tangent direction at the point.
Newton in C2 87
Once some point is enhanced, all its inverse images will be enhanced also. This is no problem
so long as the inverse images are not critical, but if one of these enhanced points again lies on a
critical value parabola, its inverse image will need to be further enhanced, by whatever structure it
takes to define the finite sequence of blow-ups required for the Newton map to extend. Presumably,
this sort of enhancement occurs at most twice, but we don’t know this for sure.
By induction, define Z̃i = Ni−1 (Z̃i−1 ), (where the inverse image of an enhanced point is an
enhanced point, perhaps further enhanced as above) and Xi+1 = F(Xi , Z̃i ). This is possible
because of Proposition 5.1.2.
Proposition
in 5.1.2. (a) For each i ≥ 2, the set Z̃i is Farey data, contained in the most recent
Farey divisor.
(b) For every i ≥ 1, the map N extends to a map Ni+1 : Xi+1 → Xi .
Proof. We assume by induction that Ni−1 maps F(Z̃i−2 ) to Z̃i−2 , and is a finite map on smooth
points. Then Z̃i−1 is Farey data, hence Xi exists, and Theorem 4.2.2 guarantees that (Ni−1 :
Xi−1 − Zi−1 ) → (Xi−2 − Zi−2 ) extends to Ni : Xi → Xi−1 , and is a finite mapping on the smooth
points. ¤
We now consider the projective limit X∞ = lim (Xi , πi ), which comes with the mappings
←
πi : X∞ → Xi , and a map N∞ : X∞ → X∞ . The space X∞ contains a “divisor at infinity”
D∞ = (π0 )−1 (l∞ ).
The space X∞ is central to our way of understanding the Newton map, and we will spend the
rest of this chapter trying to analyze it.
Proposition
in 5.2.1. The map Π is not injective.
Proof. Our computation to show this will depend on our computational approach to the extension
at B̃2 (see Figure 41), and will use the notation set up there. Consider the component L1/4 in the
Farey blow-up of the point B̃2 , according to the parametrization above. This component maps to
LÃ ! , and this map is 2-to-1. Choose some point y1 ∈ LÃ ! , and let its two inverse images in
−3 −3
2 2
88 Hubbard and Papadopol
L be y0′ and y0′′ . These points can be thought of as belonging to X∞ , since the component LÃ !
−3
2
maps to the component LÃ ! in zone 2, which itself maps to LÃ ! in zone 1, where the Newton
4 4
1 3
map is the identity. So no blow-ups will ever be performed on the points y0′ , y0′′ .
Thus the two distinct points of X∞
(B2 , y0′ , y0′ , . . . ) and (B2 , y0′′ , y0′′ , . . . ) (5.8)
both map to the sequence in Ξ∞
B2 , y1 , N (y1 ), N ◦2 y1 , N ◦3 y1 = N ◦2 y1 , . . . .
¡ ¢
(5.9)
This shows that the spaces X∞ and Ξ∞ are different. ¤
It turns out that Π is a homeomorphism on the most interesting set of points. The most
recurrent sequences in Ξ∞ are the subset Υ∞ of sequences that contain an infinity of B1 ’s or B2 ’s.
Let us call Y∞ = Π−1 Υ∞ .
Proposition
in 5.2.2. The map Π : Y∞ → Υ∞ is a homeomorphism.
Proof. The proof is an unpleasant computation. First, observe that the components of D1 that
eventually map to LÃ ! are
1
2
LÃ !, LÃ !, LÃ !, LÃ !, LÃ !, (5.10)
1 −4 1 −22 1
−1 1 −10 1 −46
where the indices ni that are not 1 are obtained recursively by the rule −ni+1 = −(2ni + 2).
In particular, these indices are all even, and these components are all critical, except for LÃ !.
1
−1
We will show that any component of a Farey exceptional divisor of Xn that maps to one of these
by N1 ◦ N2 ◦ · · · ◦ Nn is a critical curve. This will prove the proposition.
Each Farey blow-up is “parametrized” by the positive reals in a natural way, for instance by the
points of the line p = 1 in the associated quadrant Q.
These enriched points where we are making the blow-ups are either critical or not; if they are
critical, then we will call the associated Farey blow-up critical. Only components of critical Farey
blow-ups can be critical.
To begin
³ ´with, we saw above that the critical components of the Farey blow-up of B1 and B2
are the pq -components with q − p even, and that these map to the components
p (p − q)/2
³ ´ ³ ´
(p − q)/2 and p (5.11)
1
³ ´
respectively. So the inverse image of LÃ ! with q < 2 is the component 1 − 2q (in the blow-up
1
q
of B1 ), and exactly the same holds at the blow-up at B2 . All these components are critical.
Now it is easy to proceed by inverse
³ images. At every step, the inverse image of one of these
1
´
components is a component labeled q for some q; indeed, only such a component can map
Newton in C2 89
BR+ (Xn , Dn ) = B + F(Xn−1 ; Z̃n−1 , Dn−1 ) and BR+ (Dn ) = B + F(Z̃n−1 , Dn−1 ) (5.12)
Theorem
in 5.3.1. (a) The subset BR+ (D∞ ) ⊂ BR+ (X∞ , D∞ ) is homeomorphic to a 3-sphere.
(b) The map N∞ : X∞ → X∞ lifts to a map BR+ (D∞ ) − (S1 ∪ S2 ) → BR+ (D∞ ), where S1 and
S2 are the two circles corresponding to the points where the polar hyperbola meets LÃ ! and
1
−1
LÃ !. This map is a ramified covering map, ramified along the four circles where the critical
−1
1
cubics intersect LÃ !, LÃ ! , LÃ ! , LÃ ! .
1 −1 1 2
−1 1 2 1
This description is clean, but not very geometric. Figure 42 is an attempt to draw the first
stage of the real oriented blow-up. Make the real oriented blow-up of the primitive space XP = X1
along the divisor D1 at infinity. This is a 3-sphere with a Cantor set of concentric tori in it. The
zones (or gaps) between tori correspond to the components of D∞ , and hence are labeled by the
primitive points, i.e., ³
by pairs of coprime integers
³ ´ with at least one positive. These gaps are foliated
by closed curves; the q -gap is foliated by pq torus knots. Two of the tori are degenerate; they
p
´
are round (ordinary) circles in the 3-sphere, and correspond to the adherence of the x and y axes.
90 Hubbard and Papadopol
1
-1 1
0 2
1
1
1
1
2
0
1
-1
1
Figure 42. A sketch of the 3-sphere at infinity if we take a real oriented blow-up of X1 along D1 . We have
not managed to actually draw a Cantor set of tori; one should imagine each of those drawn as thickened
to give a Cantor-set section.
We next need to understand the Farey blow-up of the point of coordinate −2 in LÃ !. That
1
2
blow-up thickens a circle orbit in the zone corresponding to 21 , and puts another Cantor set
³ ´
of tori inside the thickened circle orbit, now a solid torus. Figure 43 attempts to draw the first
thickened circle, now torus (which looks flatter than we would like).
µ ¶
Figure 43. Left: one circle orbit in the zone 12 has been thickened to form a solid torus; one should
imagine it filled with a Cantor set of concentric tori, with zones in between corresponding to the components
of the Farey blow-up of the point of coordinate −2 in Lµ1¶ . Right: the solid torus by itself, shown winding
2
µ ¶
twice in one direction and once in the other, as its label 1
2 specifies.
Newton in C2 91
The real oriented blow-up contains three other new tori, in the zones corresponding to LÃ !,
2
1
LÃ !, and LÃ !; it is beyond our software to draw these. At the next level, we will need to
1 1
−1 −1
thicken more circles (12 of them), of which two will be inside the zones within the most recent
Farey blow-up. We ³won’t
´ try to draw this in any detail, but Figure 44 attempts to sketch what
1
we find in the zone 2 .
1
1
S1
....
.
2
1
....
.
S1
³ ´
2
Figure 44. In a cross-section of zone , we see the region between two circles. Crossing from the inner
1
one to the outer one, in two opposite places, we see the Julia set of the rational map (3.18). Originally,
this Julia set is just a line segment (joining 0 to ∞), but the critical point (of coordinate −2) has been
blown up to a solid torus, as have all its inverse images, which are dense in the interval and blown up,
make a string of beads joining the interior to the exterior torus. In three dimensions, it is more like a
ribbon. Each of these beads is itself a Cantor set of circles, with gaps, of which some themselves contain
Cantor sets of circles, and so on ad infinitum. In addition, the circle S1 is in this zone, and intersects our
cross-section in two points, indicated by stars. This circle is one of the four circles of ramification of the
√
map N∞ . It is attracted to the circle corresponding to the fixed point at coordinate 2 − 2. This circle
also intersects the cross-section in two points, indicated as black dots.
What are the fibers of the projection BR+ (D∞ ) → D∞ ? Some are easy to understand, namely
those that correspond to a point z of some Dn at which no further Farey blow-ups have been
performed. Then the fiber is either:
• A circle, if z ∈ Xn∗ .
• A torus, if z is not a smooth point of D∞ . This torus carries a foliation: an irrational foliation
if z is not in the closure of any component of Dn , and a rational foliation by closed curves if z
92 Hubbard and Papadopol
is in the closure of some component of Dn . This applies in particular to the points z of Dn that
correspond to points of some earlier Dm where Farey blow-ups were performed.
The more complicated fibers occur above the points of Y∞ ; they are solenoids. These solenoids
are
³ accumulations of more and more elaborate iterated torus knots, and come with a sequence
p1 , p2 , . . . that describes what sort of solenoid it is, and how it is embedded in S 3 . Objects
´ ³ ´
q1 q2
of this sort are implicit in the standard picture of the KAM theorem, as the limits of solutions
that live on islands off islands off islands . . . ; however, such solutions of Hamiltonian systems do
not appear to have been analyzed in the literature.
There is much more to say about the dynamics of this real oriented blow-up, more particularly
about the topology of the “Cantor set of solenoids,” and the invariant measures which it carries,
which will be the topic of a future paper.
H2 (X1∗ ) = Z ⊕ S1 ⊕ S2 , (5.14)
Proposition
in 5.4.1. The natural inclusion H2 (X1∗ ) → ZIrr D1 is an isomorphism onto SP .
Proof. This is very similar to Theorem 4.5.5, or alternately follows from it with a bit of work. ¤
Newton in C2 93
5 4 3
1 2 3 4 5
µ ¶ µ ¶ µ ¶
Figure 45. Left: the “Farey plane” of Figure 31, with the line segments joining 0 to 1 to 1
1 1 0
drawn in, and with seven distinguished points on the segments. Right: the graph of the function that
encodes where each component of D1 is mapped; this function will be denoted NJ .
Proposition
in 5.4.2. The elements eà ! of SP form a basis.
p
q
Proof. This is almost identical to Lemma 4.5.4; again we omit the proof. ¤
Intersection on X1∗ induces a pseudo-inner product on H2 (X1∗ ), and our basis is well adapted to
this product.
Proposition
in 5.4.3. For the natural inner product, we have
heà ! , eà ! i = 1, (5.15)
1 1
1 1
p 6= 1 ,
³ ´ ³ ´
heà ! , eà ! i = −1 for all primitive vectors q 1 (5.16)
p p
q q
p1 p2 .
³ ´ ³ ´
heà ! , eà !i = 0 when q1 6= q2 (5.17)
p1 p2
q1 q2
94 Hubbard and Papadopol
Proof. This was implied by Proposition 5.4.1. Indeed, the homology class eà ! is represented
p
q
by the divisor LÃ ! at the level of blow-up at which it first appears; it is then deformed so as to
p
q
avoid all the points where future blow-ups will take place. This apples to all such classes except
LÃ ! , which is the original line at infinity, similarly deformed (for instance into the diagonal, or
1
1
any line that avoids p1 and p2 ).
Thus clearly all these classes have self-intersection −1, just as they have at the moment of cre-
ation (except LÃ ! , which wasn’t created, and has self-intersection 1), and they are all orthogonal
1
1
to each other. ¤
Proposition
in 5.4.4. The restriction mapping SP → H1 (J) is an isometric embedding, and the
image of SP ⊗Z C is dense in H1 (J).
Proof. It is enough to see that the images of the eà ! are orthogonal and have the correct pseudo-
p
q
length. The function eà ! restricts to the constant 1 and hence has the right pseudo-length. For
1
1
the other basis vectors, the proof is identical to that of Theorem 4.7.2. For the density, consider
for instance that the derivatives of our functions include the Haar functions. ¤
The next step is to compute the map (NP )∗ : H2 (X1∗ ) → H2 (X1∗ ); Recall that NP is defined in
Proposition 3.2.1. The computation is very similar to the one in the proof of Theorem 4.6.1, and
we won’t go through all the details, especially as most of the work
³ ´ was done in Section 3.2. Just
as there, we must deal separately with the classes eà ! when pq is in zones 1-5, and also look
p
q
carefully at the special cases.
Let us begin with the hardest case, zone 1. We will compute the image of eà ! when the parents
p
q
are pq1 and pq2 with q1 even. If we perform two blow-ups in the domain and range, we get
³ ´ ³ ´
1 2
Figure 46, very analogous to Figure 38, where the labels describing what maps to what, and what
is critical, were all computed in Section 3.2.
Newton in C2 95
p1+q1/2
p1
p1
q1
2p2+q2
2p2
p2
p1+q1/2 q2
p1
p1
q1
2p2+q2+p1+q1/2
2p2+q2 p2
2p2+p1
2p2 q2
p1+p2
p1+q1/2 q1+q2
p1
2p2+q2+2p1+q1 p1
p1+p2 q1
2p2+2p1 q1+q2
Figure 46. The Newton mapping, acting on a partial blow-up, on the way to constructing X1 .
LÃ ! + LÃ !. (5.19)
p p1 +2p2
q q1 +2q2
Finally, this combination of divisor components, at this level of blow-up, represents the class
eà ! + eà !. (5.21)
p1 +2p2 +q2 +q1 /2 2p1 +2p2 +q1 +q2
2p2 +p1 2p2 +2p1
The same computation shows that eà ! (for which both parents have second coordinate
p1 +2p2
q1 +2q2
odd) maps to
eà ! − eà !. (5.22)
p1 +2p2 +q2 +q1 /2 2p1 +2p2 +q1 +q2
2p2 +p1 2p2 +2p1
This computes the action on homology in zone 1; zone 5 is identical, after exchanging the first
and second coordinates.
In zone 2, the Newton map simply takes eà ! to eà ! , and in zone 3, it acts as the identity.
p p
q p−q
96 Hubbard and Papadopol
This leaves the exceptional cases, which we simply list; the computations are all the same sort
of thing: blow up just until you can see what is going on, and no more:
eà ! 7→ eà !
1 1
1 1
eà ! 7→ eà ! + eà ! eà ! 7→ eà ! + eà !
2 2 1 1 1 0 (5.23)
1 1 0 2 2 1
eà ! 7→ eà ! + eà ! eà ! 7→ eà ! + eà ! .
1 1 0 −1 2 1
−1 2 1 1 1 0
In some sense, this is a complete computation of the action of N on the homology of the primitive
space. However, it is extremely unwieldy, and as it turns out, there is a remarkably simple way of
writing all of it.
To this end, we notice that there is a map NJ : J → J that encodes where each component of
D1 is mapped; we saw the graph of this function in Figure 45, on the right.
Using this map, we can “define” a mapping
X
(NJ )∗ : H1 (J) → H1 (J) by (NJ )∗ (f )(x) = f (y). (5.24)
y∈NJ−1 (x)
Theorem
in 5.4.5. The mapping (NP )∗ : H2 (XP∗ ) → H2 (XP∗ ) extends uniquely to the mapping
[
(N 1 1
P )∗ : H (J) → H (J) (5.25)
given by
µ ³ ¶
[ 1´ ³1´
(NP )∗ (f ) = (NJ )∗ (f ) + 2 f − χ[−1/2,−∞) + f χ[1/2,∞) , (5.26)
2 2
where χI is the characteristic function of I.
Proof. We will prove this by showing that it is true for each eà !. The most interesting case is
p
q
p is in zone 1 or 5, but in those cases the computation is identical to that in Theorem
³ ´
where q
4.6.1, and more specifically illustrated in Figure 39, and we omit it. For pq in zones 2, 3, or 4,
³ ´
the formula is easy to check, but we need to look carefully at the exceptional cases.
is the constant 3 on [−1/2, 1/2], and the constant function 1 elsewhere. The extra terms add on
the constant 2 outside of [−1/2], [1/2], giving the constant function 3.
The case of eà !. In this case,
2
1
is the function 2x on [0, 1/2], extended by 0. The extra term adds the constant 1 on [1/2, ∞), and
these two together are equal to eà ! + eà ! .
2 1
1 0
is the constant 1 on (−∞, 1/2], and the constant function 0 elsewhere. The extra terms add on
the constant 1 on [1/2, ∞], giving the constant function 1, i.e., eà ! . ¤
1
0
This proves the result, but the proof is very unsatisfactory. The statement of Theorem 5.4.5 is
so clean that it seems impossible that the interpretation of homology in terms of Sobolev spaces is
just a fluke. There must be a reason why this works other than several miraculous computations,
but we were unable to find a conceptual proof.
∗
5.5. The homology H2 (X∞ ).
∗
We will now compute the homology of X∞ , as the inductive limit lim H2 (Xn∗ ) under the obvious
→
inclusions. Exactly as in Proposition 4.5.2, this homology is naturally a subgroup of ZIrr(D∞ ∪E∞ ) ,
which we will identify.
Each Xi+1 is obtained from Xi by performing some finite set of Farey blow-ups on Xi , at points
z ∈ Di . For each such Farey blow-up, take a new copy Qz of the first quadrant in R2 . This copy
comes with the commutative group Sz of continuous, piecewise-linear functions on Qz that are
linear except on finitely many rays with rational slopes, and³ take
´ integral values at integral points.
p
This group comes with a basis ep/q,z labeled by the points q with coprime integer components,
as in Equation (4.29). When discussing these functions, which are defined on a single quadrant,
we return to the notation of Chapter 4, and encode them by the slope p/q of the ray through pq .
³ ´
+
We will need to add one element to this group:
³ ´define Sz to be the same set of functions, except
that we no longer insist that they vanish at 10 . There is a basis vector e1/0 corresponding to
this element, which is simply the function p, if we call (p, q) the coordinates on Qz .
Now we define a two-dimensional tree-like “complex,” starting with Q1 = QP , and making Qn
by attaching a copy of Qz to Qn−1 for each point z ∈ Dn−1 at which a Farey blow-up is performed.
98 Hubbard and Papadopol
Qn−1 corresponding to the component to which z belongs, identifying the point 10 to that point
³ ´
To describe the homology of X∞ we must also deal with the blow-ups of points in the finite
plane, which we will do by simply defining In = Irr En , and I∞ = Irr E∞ = ∪n In .
Theorem
in 5.5.1. The map
S∞ ⊕ Z(I∞ ) → ZIrr D∞ ∪Irr E∞ (5.31)
that on the first summand restricts functions to the integral points with coprime coordi-
nates, and on the second is the obvious inclusion Z(Irr E∞ ) → ZIrr E∞ , is an isomorphism onto
∗
H2 (X∞ ) ⊂ ZIrr D∞ ∪Irr E∞ .
for each of these spaces, the homology is generated by the fundamental classes of all the proper
transforms of the exceptional divisors, plus the proper transform of the line at infinity.
At the time an exceptional divisor is created, we can also deform it so that it avoids any points
∗
where future blow-ups will take place. It then represents an element of H2 (X∞ ), but that class is
not the proper transform of the divisor in question after further blow-ups have been performed;
the precise statement is given in Proposition 4.8 of [HPV], repeated here as Proposition 5.5.2.
Proposition
in 5.5.2. The composition
¡ ¢
H2 (X) → H2 X − {z} → H2 (X̃z ) (5.35)
maps the fundamental class [C] of a curve C in X with m smooth branches through z to
[C ′ ] + m[E] ∈ H2 (X̃z ), where C ′ is the proper transform of C in X̃z , and E is the exceptional
divisor.
Suppose now that a Farey blow-up is to be performed at a point z ∈ Dn , and that z is a point
of some component D. In the system of finite blow-ups, let us put ourselves at a level at which D
has just been created. When we make our first blow-up at z, creating the component L1/1,z of the
eventual Farey blow-up, the class [D] is represented by [D′ ] + [L1/1,z ]. When we next blow up the
point of intersection of these two, creating L2/1,z , the image of [D] will be
[D′ ] + [L1/1,z ] + 2[L2/1,z ]. (5.36)
By induction, it should be clear that when the whole Farey blow-up has been performed, the
class [D] will become [D′ ] + e1/0,z . This shows that the diagrams (5.34) all commute, and proves
Theorem 5.5.1. ¤
The space Q∞ is a bit unwieldy, and we want to replace it by something more tangible, so that
S∞ will become a space of functions on a graph. We will draw this graph in Q∞ as follows, shown
in Figure 47. Begin by setting J1 = J ⊂ Q1 = QP . Next, any time we attach a quadrant Qz to
Qn−1 , the ray where the attaching takes place will intersect Jn−1 in a unique point. Draw through
that point the parallel to the free boundary ray of Qz . Note that the function e1/0,z restricts to a
constant on this ray. The union of the rays obtained this way is our graph J∞ .
{
{
J1
J2
J3
Figure 47. A sketch of J1 , J2 and J3 , together with the way the map NJ transforms segments.
100 Hubbard and Papadopol
Give each ray I of J the unique multiple of Lebesgue measure so that the restrictions of the
functions ep/q,z to the ray are orthonormal in H1 (I). This is possible by Theorem 4.7.3. This
measure can be written |dt|, for a unique affine parametrization of the ray that is 0 at the point of
attachment.
We will still denote by |dt| the non-atomic measure on J that restricts to |dt| on each ray. This
puts a pseudo-inner product on S∞ , by setting
Z
hf, gi = − f ′ (t)g ′ (t) |dt| + f (0)g(0) (5.37)
J
In this formula, the point 0 is the point in the middle of J1 , i.e., the point 11 of QP . Note that
³ ´
this integral is finite since elements of S∞ are constant except on finitely many rays, and there
their derivatives have compact support.
∗
Theorem
in 5.5.3. The isomorphism H2 (X∞ ) → S∞ of Theorem 5.5.1 transforms the intersec-
∗
tion inner product of H2 (X∞ ) into the inner product
Z
hf, gi = − f ′ (t))g ′ (t)|dt| + f (0)g(0) (5.38)
J
on S∞ .
Proof. This follows from Theorem 4.7.3, together with Theorem 4.5.6. ¤
∗ ∗
Define Ĥ2 (X∞ ) to be the completion of H2 (X∞ ) with respect to the inner product. Then
Theorem 5.5.3 has the following corollary.
∗
Corollary
in 5.5.4. The isomorphism H2 (X∞ ) → S∞ of Theorem 5.5.1 extends to an isometry
∗ 1
Ĥ2 (X∞ ) → H (J).
µ ¶ µ ¶
p 2p + q
7→ in zone 1
q 2p
µ ¶ µ ¶
p p
7 → in zone 2
p−q
q
µp¶ µ ¶
p
7→ in zone 3 (5.39)
q q
µ ¶ µ ¶
q−p
p
→
7 in zone 4
q q
µ ¶ µ ¶
p 2p
7→ in zone 5.
q 2p + q
We also computed in Equation (5.6) what it was on the quadrant corresponding to the blow-up
at B2 ; this and a symmetric computation at B1 give
p 7→ 2q p 7→ p − q at B .
³ ´ ³ ´ ³ ´ ³ ´
q q−p at B1 , q 2p 2 (5.40)
Any other Farey blow-up at some point z ∈ Xk will map to a previous Farey blow-up at Nk (z),
and there are natural coordinates (p, q) in the domain quadrant and in the range quadrant. The
map NQ is then written
µ ¶ µ ¶
p p
7→ if z is not critical
q q
µ ¶ µ ¶ (5.42)
p 7→ 2p
if z is critical
q q
To see that this is compatible with Theorem 4.6.1, we need to know that if the blow-up is at an
enriched point z̃ that is critical, then the critical curve through³z ´
is the component of Dk , not the
1
enriching curve in the finite plane. Thus the attaching point 0 corresponds to the end of the
³ 1´
parametrized curve tt0 , which tends as t → 0 to a point of the critical curve.
Theorem
in 5.6.1. The mapping H2 : X2∗ (X∞ ) → H2 (X∞ ∗
) , viewed as a mapping S∞ → S∞ , is
given by the formula
¡ ¢ ¡ ¢
N∗ (f ) = (NQ )∗ f |Q′∞ + 2 NQ )∗ (f |Q′′∞ . (5.43)
Proof. We will check that the formula is true for each ep/q,z ; there are some cases where it is
easy, and several where it is quite complicated.
First, let us see why the 2 in the formula is necessary. Figure 48 illustrates what is going on.
102 Hubbard and Papadopol
1 2 3
1 2 3
1 2 3
1 2 3
1 2 3
Figure 48. Left: a Farey quadrant Qz attached to an earlier quadrant. Right: their images.
On the left we see a quadrant Qz , attached to the quadrant Q of some earlier Farey blow-up.
Note that since the component to which it is attached is critical, this earlier Farey blow-up itself
was at a critical point. The marked point on the left, primitive on the ray to which the new
quadrant is attached, is mapped to a non-primitive point in the quadrant Q′ = (NQ )∗ (Q), in fact
twice a primitive point, as marked on the right.
On Q, we have a function (on the left) which is 1 at the primitive point (its values off the ray
are irrelevant). It pushes forward, by (NQ )∗ to the function which is 1 at the marked point on the
right. But the image on the homology is twice that: it is the function which is 1 at the primitive
point on the ray in the range. Thus to be continuous the function e1/0,z must map to twice the
function e1/0 of the image quadrant.
This argument proves a lot of Theorem 5.6.1, in fact, together with Theorem 4.6.1 and Theorem
5.4.5, it proves it for all ep/q,z with support only where NQ is a local homeomorphism.
It is then clear that we only need to verify the formula for the exceptional cases:
eà !, eà !, eà !, eà !, eà !, eà !, eà !. (5.44)
1 1 2 1 1 0 −1
−1 0 1 1 2 1 1
This is quite tedious, and extremely prone to error. We will do it in detail only for eà ! and
1
−1
eà !, and leave the others to the reader.
1
0
The case of eà !. Figure 49 illustrates what is happening in this case. We know from Equation
1
−1
(5.23) that
³ ´
N ∗ eà ! = eà ! + eà ! (5.45)
1 1 0
−1 2 1
2 2 2 2 0
1
1 1 1 0
0 0 0
0 0 0
0 1 1 1
0 1 2 2
0 1 2 3
1 2 2
1 2 2
1 2 2
Figure 49. On the left, the primitive plane and the quadrant corresponding to the blow-up at A1 . On
the right, the primitive plane, viewed as the range.
However, the function eà ! also has support in the quadrant corresponding to the blow-up of
1
−1
A1 ; in fact it is the function e1/0 in that quadrant, as shown in the lower half of the picture. That
quadrant also maps into the shaded region on the right, and the two together add to the correct
value.
The case of eà !. Figure 50 illustrates what is happening in this case, which is quite a bit more
1
0
complicated than the former one.
0 1 2 2 2 2
0 1 2 1 1 1
2
1 2 3
0 0 0 1 2
0 1 2 3
1 2
0 1 2 3
1 2
0 1 2 3
1 2 2 1 2 2
Q
Q B1
A1 1 2 2 1 2 2
1 2 2 1 2 2
Figure 50. On the left, the primitive plane and the quadrant corresponding to the blow-ups at A1 and
B1 . On the right, the primitive plane, viewed as the range.
There are now six separate regions in the domain to think about. The part in zone 1 pushes
forward to the restriction of eà ! in zones 4 and 5. The part in zone 2 with q < 0 pushes forward
1
1
to half the restriction of eà ! to the part of zone 3 where q > p. The part in zone 2 where q > 0
1
1
and the part in zone 3 both push forward to the region in zone 3 where q < p, and there each
contributes half to the restriction of eà ! to that region.
1
1
Since both A1 and B1 are in the support of eà !, we need to think about what both of their
1
0
quadrants contribute. The part in QA1 contributes the missing half in zone 3, q > p, and the part
in QB1 contributes the restriction of eà ! to zones 1 and 2. ¤
1
1
This has evidently been cooked up to map S∞ to S∞ . Before we can describe the action of the
Newton map in terms of H1 (J∞ ), we need one more bit of notation: Call
′
J∞ = J∞ ∩ Q′∞ ′′
and J∞ = J∞ ∩ Q′′∞ . (5.52)
J J
NQ(J)
J J NQ(J)
∗ ∗
Theorem
in 5.6.3. The mapping N∗ : H2 (X∞ ) → H2 (X∞ ) is written, using the isomorphism
∗
H2 (X∞ ) with S∞ of Theorem 5.5.1, as
N∗ (f ) = (NJ )× (f |J ′ ) + 2(NJ )× (f |J ′′ ). (5.53)
∗ ∗
Corollary
in 5.6.4. The mapping N̂∗ : Ĥ2 (X∞ ) → Ĥ2 (X∞ ) is given on H1 (J∞ ) by the formula
N∗ (f ) = (NJ )× (f |J ′ ) + 2(NJ )× (f |J ′′ ). (5.54)
This allows us to understand N̂∗ as a bounded operator in a Hilbert space. The spectral analysis
of this operator should be interesting: it should contain all the cohomological information about
invariant (1, 1)-currents. We intend to write another paper about this.
6
The case where a and b are arbitrary
In this chapter, we will see that essentially everything we have done in the case a = b = 0 in
Chapters 3 and 5 is true for all a and b. The main difference is that whereas the x and y axes map
to each other when a = b = 0, they no longer play any special dynamical role when a and b are
arbitrary. The first step is to discover the appropriate curves to take the place of the x and y axes
in the enriched points p̃1 and p̃2 . It turns out that there are also curves of order two in this case,
although the curves are transcendental, and we don’t know much about their global behavior.
Theorem
in 6.1.1. There exist R > 0 and unique analytic mappings
µ ¶ µ b ¡ 1 ¢¶
t + o(t) 2t2 + o t2
γ1 (t) = a
¡ 1
¢ and γ2 (t) = (6.1)
2t2 + o t2 t + o(t)
defined for |t| > R, such that
N (γ1 (t)) = γ2 (t2 ) and N (γ2 (t)) = γ1 (t2 ). (6.2)
Proof. The following lemma will be the principal tool. Let UR = {|z| ≥ R} and let ER be the
Banach space of bounded continuous functions on UR , analytic on the interior. We will denote by
kf kR the sup norm on this Banach space. We will denote by Γ(f ) and Γ′ (f ) the curves of equation
y = f (x) and x = f (y).
Lemma 6.1.2. There exist R > 0 and constants 0 < K1 < K2 (depending on a and b) such that
for any function f ∈ ER with kf k < K2 , there exists a unique f˜ ∈ ER with kf˜kR < K1 such that
N −1 (Γ(f )) ∩ |y| ≥ R, |x| < K2 = Γ′ (f˜).
© ª
(6.3)
Proof of Lemma 6.1.2. Looking at the dominant terms of the formula for N , we see that for
all M > 0, there exist R and K such that if we set
½³ ´ ¾
x M
AM,R = y | |y| ≥ R, x ≤ |x|2 , (6.4)
106
Newton in C2 107
Figure 52. This shows how the regions in Equation (6.4) are situated.
It follows that any horizontal disc y = c, |x| ≤ M/|c|2 maps injectively to an almost vertical disc,
with x-coordinate approximately c2 , and with y-coordinate approximately filling the disc of radius
2M around a. Choose M = |a| + 1, so that this disc will contain the disc of radius M around 0
for R sufficiently large.
Then such a disc will intersect Γ(f ) in a single point if kf kR < M , and we see the N −1 (Γ(f ))
contains an analytic curve in the interior of AM,R that intersects every analytic disc in a single
point. Such a curve is Γ′ (f˜) for a well-defined f˜ that depends analytically on f . To complete the
proof of the lemma, we must show that there exists M ′ < M such that kf˜kR . This is clear from
the definition of AM,R : the maximum of |x| over AM,R is M/R2 , and this will be less that M if
R > 1. ¤
Proof of Theorem 6.1.1 from Lemma 6.1.2. Consider the map (f, g) 7→ (g̃, f˜) from the
product of the ball of radius C2 with itself in ER × ER , to itself. Since the image is in fact in the
product of balls of radius C1 , this is an analytic map of a ball in a Banach space strictly inside
itself. As such, it is strongly contracting and has a unique fixed point (F, G) ∈ ER × ER . Now
N ◦2 : Γ(F ) → Γ(F ) is a map with a super attractive fixed point: Γ(F ) is parametrized by x, and
the map is of the form x 7→ x4 + . . . . Thus there is a Böttcher coordinate t and in fact a unique
one if we impose that t ∼ x.
The same argument applies to Γ′ (G), and the two Böttcher coordinates solve our problem. ¤
Theorem
in 6.2.1. The Newton map, defined on P2 with the five points of indeterminacy removed,
extends to a mapping N1 : X1 → X1 , with seven points of indeterminacy.
108 Hubbard and Papadopol
Proof. When a = b = 0, we proved this using explicit computations. We could have proved it
using the naturality statement 4.2.2, and we will do so here, as any explicit computation would
require knowing the equations of the curves C1 and C2 .
We must make a number of preliminary blow-ups first, to be in a position to apply Theorem
4.2.2.
Blow-up p1 to create the exceptional divisor LÃ ! ; then again at the intersection with the
1
0
proper transform of C1 , to create LÃ ! ; then one more time, to create LÃ ! . In the range,
1 1
−1 −2
we blow up once at p2 , to create LÃ ! ; and once at the intersection of LÃ ! and LÃ ! . Then
0 0 1
1 1 1
the Newton map is well defined on a neighborhood of LÃ ! , mapping it with normal degree 2
1
−2
to LÃ !. The situation is drawn in Figure 53: our one-term asymptotic expansion (6.1) of the
0
1
equation of C1 guarantees that after these blow-ups, the proper transform of C1 does intersect
LÃ ! , and not at the point where LÃ ! intersects it.
1 1
−2 −1
0
1 1
2
C2
1
1
1
1 0
-1
1
-2
C1
We can now use Theorem 4.2.2 to extend the Newton mapping to Farey blow-ups at the points
of intersection of LÃ ! with C1 and with LÃ ! , and going to the Farey blow-ups at the double
1 1
−2 −1
points on LÃ !.
0
1
Near p1 , we still need to extend the Newton map to the Farey blow-ups of the intersections
LÃ ! ∩LÃ ! and LÃ ! ∩LÃ ! . At the first intersection, we can use Theorem 4.2.2 again, or claim
1 1 1 1
−1 0 0 1
that the computations of Chapter 3 go through without modification, as the terms contributed by a
and b are never the dominant term in that region. The second strategy also works at LÃ ! ∩ LÃ ! ;
1 1
0 1
Newton in C2 109
if we wish to apply the first, we will need to do another preliminary blow-up, since one of the
critical curves goes to this point, and is not part of the enriching structure.
A completely symmetric construction will extend the mapping near p2 . ¤
Theorem
in 6.3.1. For all a, b, the Russakovskii-Shiffman measure for Na,b does not charge the
points p1 and p2 .
Proof. We will not repeat the proof in detail, as it is almost identical with the proof in Chapter
3. We will simply comment on that proof, showing where minor modifications need to be made.
The first step is to see that the “Markov” decomposition can still be defined, i.e., that Lemma
3.3.1 is still true, and there is no difficulty there. Theorem 2.3.1 was made without restrictions,
and it asserts that all points of the lines LÃ ! and LÃ ! have unstable manifolds. Unions of these
2 3
3 2
manifolds and their inverse images were used to construct the boundaries Mk of the regions Ak .
One has to look more carefully to make sure that Lemmas 3.3.4, 3.3.5, 3.3.6, and 3.3.7 are still
true. For the first three, there is no problem, since in the corresponding regions Ak , −3 ≤ k ≤ 3,
the asymptotics of the particular map used there coincide with the asymptotics in the general
case. As for Lemma 3.3.7, the only way the asymptotics of the mapping N were used was that in
those regions (Ak , |k| > 3) either the x2 in the second coordinate (near p1 ) or the y 2 in the first
coordinate (near p2 ) was the dominant term of that coordinate function. That is certainly still
true; in fact it is the main thing used in the proof of Theorem 6.1.1.
Thus the entire argument goes through without modifications, except to adjust the regions Ak
to the dynamics at hand. ¤
Theorem
in 6.5.1. In X∞ , the basins of the roots are connected and holomorphically convex. If
they contain no enhanced points, they are Stein.
Proof. Let a be a root, Ua be its basin, and Ga : Ua → R ∪ {−∞} be the potential function of
the root, as defined in [HP]:
1 ◦n
Ga (x) = lim log kN∞ (x) − ak. (6.7)
n→∞ 2n
Clearly the basins of the roots are entirely made up of smooth points of X∞ , since all such
points are mapped by an iterate of N∞ to a small neighborhood of the roots that is an open subset
of C2 . The key point is that Ga is proper pluri-subharmonic function, i.e., that for all k > 0, the
set where Ga (x) ≤ −k is compact (this is not true in C2 ).
First, the function Ga is continuous on Ua , except at the root itself, because the root has no
inverse images other than itself. Moreover, there exists a compact neighborhood V of a and a
constant K > 0 such that Ga (x) > −K if x ∈ / V.
Choose k > 0, and a sequence xn satisfying Ga (xn ) ≤ −k. We can certainly extract a convergent
subsequence in X∞ , converging to some point y. If y is in the basin of a, we are done by the
continuity of Ga . Otherwise, y must be in the boundary of the basin. But then, for any m > 0 we
◦m
can find n > 0 such that N∞ (xn ) ∈
/ V , since
◦m ◦m
lim N∞ (xn ) = N∞ (y) (6.8)
m→∞
is in the boundary of the basin. That implies that Ga (xn ) > −K/2m , and contradicts Ga (xn ) ≤ −k
if 2m k > K.
This proves that the basins are holomorphically convex; to see that they are Stein, we must
show that the global functions separate points. There is no problem for pairs of points that do
not belong to the same component of some En : the projection to C2 will then separate the points.
Neither is there a problem if both points belong to a single exceptional divisor: just so long as that
divisor is mapped to one of the double tangents by some iterate of N , then π∞ ◦ N k will separate
the points. But if some exceptional divisor is collapsed to a point in the interior of a basin, which
may happen as far as we know, then the global functions will of course not separate points on that
exceptional divisor. ¤
Remark. In P2 , sequences in a basin can approach points of indeterminacy, without the potential
there tending to 0. △
This gives some information about the complex analytic structure of the basins in C2 .
Theorem
in 6.5.2. The basins of the roots in C2 are Stein manifolds.
Proof. Let us first consider the case where there are no enhanced points in the basin Ua of a
root a. Then the basin in C2 is the complement of ∪n En in Ua , and the complement of a closed
hypersurface in a Stein manifold is Stein.
Newton in C2 111
Actually, this argument also works if there are enhanced points. We must first blow down all the
projective lines that are blown down by some iterate of N . We then obtain a Stein space, which
has singularities since even in the simplest case the lines being blown down have self-intersection
−2; but this space is holomorphically convex, and the global functions do separate points.
Again in this space, the basin in C2 is the complement of ∪n En , which is a closed hypersur-
face. ¤
Bibliography
[Ar] V.I. Arnold, Small divisor problems in classical and celestial mechanics, Usp. Mat. Nauk
18, no. 6 (114), 91–192 (1963).
[B] E. Bedford, Iteration of polynomial automorphisms of C2 , Proc. Int. Cong. Math. Kyoto,
(1990), 847-858.
[BS] E. Bedford and J. Smillie, Fatou-Bieberbach domains arising from polynomial automor-
phisms, Indiana Math. Jour., 40 (1991), 789-792.
[BS1] E. Bedford and J. Smillie, Polynomial diffeomorphisms of C2 I: Currents, equilibrium mea-
sure and hyperbolicity, Invent. Math., 87 (1990), 69-99.
[BS2] E. Bedford and J. Smillie, Polynomial diffeomorphisms of C2 II: Stable manifolds and re-
currence, J Amer. Math. Soc., 4 (1991), 657-679.
[BS3] E. Bedford and J. Smillie, Polynomial diffeomorphisms of C2 III: Ergodicity, exponents and
entropy of the equilibrium measure, Math. Ann., 294 (1992), 395-420.
[BS5] E. Bedford and J. Smillie, Polynomial automorphisms of C2 V: Critical points and Lyapunov
exponents, Journal of Geometric Analysis 8,3 (1998).
[BS6] E. Bedford and J. Smillie, Polynomial diffeomorphisms of C2 VI: Connectivity of J, Ann.
Math., 148(1998), 695-735.
[BS7] E. Bedford and J. Smillie, Polynomial diffeomorphisms of C2 VII: Hyperbolicity and external
rays. Annales Scient. E.N.S. Paris 32(1999), 455-497.
[BS8] E. Bedford and J. Smillie, External rays in the dynamics of polynomial diffeomorphisms of
C2 , Contemporary math. 222(1999), 41-79.
[BCSS] L. Blum, F. Cucker, M. Shub, S. Smale, Complexity and real computation, Springer, New
York (1998).
[Bou] T. Bousch, Quelques problèmes de dynamique holomorphe, Thesis, University of Paris at
Orsay, 1992
[Br] J.Y.Briend, Exposants de Liapounof et points périodiques d’endomorphismes holomorphes
de CP k , Thesis, Toulouse (1997).
[De] P. Deligne, Intersections sur les surfaces régulières, SGA 7, Exposé 10, IHES (1969).
[Dil] J. Diller, Dynamics of birational maps of P2 , Ind. J. Math. 45.3 (1996), 721-772.
112 Hubbard and Papadopol
[Dou] R. and A. Douady, Algèbre et Théorie Galoisiennes, I and II, Cedic/Fernand Nathan, Paris
1977, 1979.
[FS1] J. Fornaess and N. Sibony, Complex Dynamics in Higher Dimension I , Asterisque 222
(1994), 201-231.
[FS2] J. Fornaess and N. Sibony, Complex Dynamics in Higher Dimension II , Ann. Math. Studies
137 (1995), 134-182.
[FS3] J. Fornaess and N. Sibony, Complex Dynamics in Higher Dimension, in Complex Potential
Theory (1995), Kluwer Academic Publishers, 131-186.
[Fo] J. Fornaess, Dynamics in several complex variables, CBMS 87
[Fr1] S. Friedland, Entropy of polynomial and rational maps, Ann. Math, 133 (1991), 359-368.
[Fr2] S. Friedland, Entropy of algebraic maps, J. Fourier Anal. and Appl.(1995), 215-228
[GH] P. Griffiths and J. Harris, Principles of Algebraic Geometry, Wiley, N.Y., 1978.
[Har] R. Hartshorne, Algebraic Geometry, Springer-Verlag.
[Hat1] A. Hatcher, Topology, Cambridge University Press (2000).
[HMS] M. Hirsch, J. Marsden, M. Shup, From topology to computation: Proceedings of the Smale-
fest, Springer (1993).
[Hirz] F. Hirzebruch, Hilbert Modular Surfaces, l’Ens. Math. 19 (1973), 183-281.
[HNK] F. Hirzebruch, W.D. Neumann and S.S. Koh, Differentiable manifolds and quadratic forms,
Marcel Dekker, N.Y. 1971.
[HBH] J. H. Hubbard and B.B. Hubbard Vector Calculus, Linear Algebra, and Differential Forms:
A Unified Approach, Prentice Hall, Upper Saddle River, N.J., 1999.
[Hu] J.H. Hubbard, The Hénon mapping in the complex domain, in Chaotic Dynamics and Frac-
tals, Barnsley and Demko, eds., Academic Press, New York (1986), 101-111.
[HO1] J.H. Hubbard and R.W. Oberste-Vorth, Hénon mappings in the complex domain I: The
global topology of dynamical space, Pub. Math. IHES 79 (1994), 5-46.
[HO2] J.H. Hubbard and R.W. Oberste-Vorth, Hénon mappings in the complex domain II: Projec-
tive and inductive limits of polynomials, in Real and Complex Dynamical Systems, Branner
and Hjorth, eds., Kluwer Academic Publishers (1995), 89-132.
[HP] J.H. Hubbard and P. Papadopol, Superattractive fixed points in Cn , Ind. J. Math. 43.1
(1994), 321-365.
[HSS] J.H. Hubbard, D. Schleicher, and S. Sutherland, How to really find roots of polynomials by
Newton’s method, preprint, submitted.
[Jo] M. Jonsson, Dynamics of polynomial skew products on C2 , Math. Ann. 314 (1999), no. 3,
403–447.
[KH] A. Katok, B. Hasselblatt, Introduction to the modern theory of dynamical systems, Cam-
bridge U. Press (1995).
[MSS] R. Mane, P. Sad, D. Sullivan, On the dynamics of rational maps, Ann. Scient. E.N.S. Paris,
16 (1983), 193-217.
[Ma] Y. Manin, Cubic forms, North Holland (1974).
[Mi1] J. Milnor, Singular points of complex hypersurfaces, Ann. Math. Stud. 61, Princeton
University Press (1968).
[Mi2] J. Milnor, Dynamics in one complex variable, Springer (1999).
[Pe] Y. Pesin, Characteristic Liapounov exponents and smooth ergodic theory, Russ. Math. Sur-
veys, 32 (1977), 55-114.
[PS] C. Pugh and M. Shub, Ergodic attractors, TAMS 312 (1989), 1-54.
[RS] Russakovskii A., and Shiffman, B., Value distribution for sequences of rational mappings
and complex dynamics, Ind., Jour. Math. 46 (1997), 897–932.
[Se] O. Sester, Dynamique des polynômes fibrés, Thesis, University of Paris at Orsay, 1996.
[S1] I.R. Shafarevic, Algebraic surfaces, Proc. Stek. Inst. Math., 75 (1965), AMS Providence
(1967).
[S2] I.R. Shafarevic, Basic Algebraic Geometry, I&II , Springer Verlag, NY 1994.
[Sh1] M. Shub, Global stability of dynamical systems, Springer (1987).
[ShSm] M. Shub and S. Smale, Computational complexity: On the geometry of polynomials and a
theory of cost I., Ann. Sci. Ecole Norm. Sup. 18 (1985), 107–142.
[Si] N. Sibony, Dynamique des applications rationelles de Pk , preprint 1998.
[Sm1] S. Smale, A convergent process of price adjustment and global Newton methods, J. Math.
Economy 3 (1976), 107–120.
[Sm2] S. Smale, The fundamental theorem of algebra and complexity theory, Bulletin of the Amer.
Math. Soc. 4 (1981), 1–36.
[Sm3] S. Smale, Algorithms for solving equations, Proceedings of the International Congress of
Mathematicians (1986), pp. 172–195, American Mathematical Society.
[Sm4] S. Smale, Newton’s method estimates from data at one point, in R. Ewing, K. Gross, and C.
Martin (eds.), The Merging of Disciplines: New Directions in Pure, Applied, and Compu-
tational Mathematics (1986). Springer-Verlag.
[Spa] E. Spanier, Algebraic topology, McGraw Hill, N.Y., 1966.
114 Hubbard and Papadopol