Fractals, Vol. 9, No.

4 (2001) 393–402
c World Scientific Publishing Company

π IN THE MANDELBROT SET

AARON KLEBANOFF
8519 Summershade Drive, Odenton, MD 21113, USA
[email protected]

Received March 12, 2001; Accepted April 3, 2001

Abstract
The Mandelbrot set is arguably one of the most beautiful sets in mathematics. In 1991, Dave
Boll discovered a surprising occurrence of the number π while exploring a seemingly unrelated
property of the Mandelbrot set.1 Boll’s finding is easy to describe and understand, and yet it is
not widely known — possibly because the result has not been rigorously shown. The purpose
of this paper is to present and prove Boll’s result.

1. INTRODUCTION
The Mandelbrot set M for
Qc (z) = z 2 + c (1)
(displayed in Fig. 1) is often defined to be the set of
c-values for which the orbit of zero, {Qnc (0)}|∞
n=1 , is
bounded. It is easy to show (see Appendix 1) that
if |z| ≥ |c| and |z| > 2 that the orbit {Qnc (z)}|∞n=1
diverges. From this, we know that M is contained
inside the disk |c| ≤ 2. So, we can refine our defini-
tion of M as follows:
M = {c ∈ C : |Qnc (0)| ≤ 2 for all n = 1, 2, 3, . . .} .
(2) Fig. 1 The Mandelbrot set (shown in black) for Qc (z) =
z 2 + c lies in the c–plane inside the disk c ≤ 2. The N th
The bands of shades or colors normally displayed shaded band outside M shows the c values for which the first
N iterations of the orbit, {Qn c (0)}|n=1 , escapes c ≤ 2. The
N
outside M represent the smallest number of itera- left side of the circle c = 2 is the outermost curve through
tions N such that {|Qnc (0)|}|N
n=1 exceeds 2. In what the left middle of the image.

393
394 A. Klebanoff

follows, we tell the story of an observation that Again we get the same type of relationship,
weds the number π to the banded region outside the this time the # of iterations ∗ sqrt(ε) gives pi.”
Mandelbrot set in a simple and yet remarkably
beautiful way. Boll’s discovery has since been popularized,
Our story began in 1991 when Dave Boll was mostly thanks to its publication in Chaos and
a computer science graduate student at Colorado Fractals: New Frontiers of Science,3 and also
State University. Boll had many curiosities, among thanks to Gerald Edgar4 at Ohio State University.
which was his fascination with computer-generated However, as pointed out by Edgar, although the
fractals. Boll was trying to convince himself that “π-result” for the sequence along the real axis was
only the single point c = (−0.75, 0) connects the known at least as early as 1980,5,6 only hueristic
cardioid and the disk to its left. He could not have arguments had been provided for the observation.
been prepared for what he was to discover. Boll We now provide a proof for the “π-result” into the
described his finding on his website2 which we cusp of M.
mostly follow here:

“I was trying to verify that the neck of the set

(which is at (−0.75, 0)) in the complex plane)
is infinitely thin (it is). Accordingly, I was
seeing how many iteration points of the form
(−0.75, ε) went through before escaping, with
ε being a small number. Here’s a table show-
ing the number of iterations for various values
of ε:
ε # of iterations
1.0 3
0.1 33
0.01 315
0.001 3143
0.0001 31417
Fig. 2 The two routes to π discovered by Boll. The
0.00001 314160 cusp route is c = (0.25 + ε, 0) and the vertical route is
0.000001 3141593 c = (−0.75, ε). Although not shown, c = (−0.75, −ε) will
0.0000001 31415928 work as well due to the symmetry of M.

Does the product of ε and the number of iter-

ations strike you as a suspicious number? It’s
pi, to with +/ − ε. What the heck!
Let’s try it again, this time at the butt of the
set. The butt of the set occurs at (0.25, 0), and
here’s a table for points of the form (0.25+ε, 0)
ε # of iterations
1.0 2
0.1 8
0.01 30
0.001 97
0.0001 312
0.00001 991
0.000001 3140
Fig. 3 Along the route c = (0.25+ε, 0) into the cusp of M,
0.0000001 9933 the orbit {Qn c (0)} is real and is depicted by a web diagram.
0.00000001 31414 If ε = 0, the graph of y = Qc (x) and y = x touch at the fixed
0.000000001 99344 point x = 1/2. If ε > 0 but small, there is no fixed point,
0.0000000001 314157 but most iterations are close to x = 1/2.
 π in the Mandelbrot Set 395

2. MAIN THEOREM The results in (11) follow immediately after writing

(9) in its equivalent form
The π-Theorem: Choose ε > 0 and let N (ε) be the
number of iterations required for the orbit of zero, √ x(t) − 1/2
under the map Q1/4+ε (x) = x2 + 1/4 + ε, to exceed εt = arctan √ (12)
ε
2, i.e.
N (ε) = min Qn1/4+ε (0) > 2 . (3)
n and substituting the conditions (10) into (12). 
Then √
lim+ εN (ε) = π . (4) Remark 1. The terminal state 2 in (10) could be
ε→0 replaced with any number greater than 1/2. We
Figure 3 shows that the majority of steps in the use 2 because it is the radius of the smallest circle
orbit of 0 are spent close to 1/2 and that the step- centered at the origin containing the Mandelbrot
size goes to zero there as ε → 0+ . This gives us set M.
reason to believe that the solution of the difference What remains is to show that the solution of
equation the differential Eq. (7) really does approximate the
solution of the original difference Eq. (5). We begin
1
xk+1 = x2k + + ε, x0 = 0 (5) by solving
4
or equivalently 1
xk+1 = x2k + + ε, x0 = 1/2 . (13)
 2 4
1
xk+1 − xk = xk − + ε, x0 = 0 (6) The first four iterates give
2
may be well approximated by the differential 1
equation x1 = +ε
2
dx 1
= (x − 1/2)2 + ε , x(0) = 0 (7) x2 = + 2ε + ε2
dt 2
for ε small and x near 1/2. Sure enough, as claimed 1
x3 = + 3ε + 5ε2 + 4ε3 + ε4
by the following lemma, the differential equation for 2
x near x = 1/2 yields the “π-result” analogous to 1
Eq. (4). x4 = + 4ε + 14ε2 + 34ε3 + 50ε4 + · · ·
2

Lemma 1. If ε > 0, the time T (ε) that it takes and we recognize a pattern on the lowest order
for the state variable x(t) satisfying (7) to evolve to terms in order to conjecture that
x(T (ε)) = 2 satisfies !
√ 1 X
n−1
lim εT (ε) = π . (8) xn = + nε + 2
i ε2 + · · ·
ε→0+ 2 i=0
 
1 1 3 1 2 1
Proof. If we replace the initial condition in (7) = + nε + n − n + n ε2 + · · · .
2 3 2 6
with x(0) = 1/2, then
It is natural to Taylor expand the solution of the
1 √ √
x(t) = + ε tan εt . (9) differential equation [given in Eq. (9)]
2
To prove the lemma, we need to show that 1 √ √
x(n) = + ε tan( εn)
2
x(T− (ε)) = 0 and x(T+ (ε)) = 2 (10)
1 1 2
= + nε + n3 ε2 + n5 ε3 + · · · (14)
respectively imply 2 3 15
√ π √ π in order to make comparisons with the solution of
lim+ εT− (ε) = − and lim+ εT+ (ε) = .
ε→0 2 ε→0 2 the difference equation. Since we seek to choose the
√
(11) largest possible integer n, we let n = K/ ε (with
396 A. Klebanoff

K less than but as close to π/2 as possible so that Since the tangent series converges, we know that
n is an integer) to get given any ε̂ > 0, there is a sufficiently large value of
√ n for which the magnitude of the remainder term is
x(n) = x(K/ ε)
less than ε̂. Equivalently, there must be a positive
 
1 √ 1 2 exponent a so that the error term is less than ε̂ = εa
= + ε K + K3 + K5 + · · · (15) so that (17) gives
2 3 15
which should be compared with the solution of the 3 √
difference equation √ = tan K − O(εa ) + O( ε) , a>0
2 ε
1 K
xn = +√ ε which leads to
2 ε
 3  2  !  
√ 3 √
+
1 K
√ −
1 K
√ +
1 K
√ ε2 + · · · lim n ε = lim tan−1 √ + O(εa ) − O( ε)
3 ε 2 ε 6 ε ε→0+ ε→0+ 2 ε
  π
1 √ 1 2 = .
= + ε K + K3 + K5 + · · · + O(ε) . 2
2 3 15
(16) So, our goal is to show that the sum in (19) is
really the partial sum for the tangent series. We’ll
√ also need to show a similar result for n negative.
Remark 2. We have chosen n so√that n ε ≈ π/2 We begin with a lemma that describes the form of
for ε small and so that limε→0+ n ε = π/2. the solution xn of the difference equation.
Remark 3. The expression (K + 13 K 3 + 152
K5 + Lemma 2. The solution xn of the difference
· · · ) from x(n) in (15) is the tangent series. How- Eq. (13) is
ever, a similar expression from xn in (16) is a partial
sum of 2n−1 terms that we aim to show are the first n−1
2X
2n−1 nonzero terms of the tangent series. xn = pj (n)εj , n = 1, 2, 3, . . . (19)
j=0
Taylor’s theorem says that
K 3 2K 5 where p0 (n) = 1/2 and pj (n) is a 2j − 1 degree
+ · · · + a2n−1 K 2 −1
n
tan K = K + + polynomial in n (for j = 1, 2, 3, . . .) that satisfies
3 15
the recurrence relation
n
tan(2 ) κ 2n
+ K X
j−1
(2n )! pj (n + 1) = pj (n) + pi (n)pj−i (n)
for 0 ≤ κ ≤ K where tan(j) x denotes the jth i=1

derivative of tan x, and aj denotes the jth nonzero p1 (n) = n , j = 2, 3, 4, . . . . (20)

term in the tangent series. So, once we can show
that (16) contains a partial sum for tan K, Eq. (16)
becomes Proof. Since x1 = 1/2 + ε, we have p0 (1) = 1/2
!
1 √ and p1 (1) = 1 so (19) holds for n = 1. Now assume
n−1
tan(2 +1) κ
n
xn = + ε tan K − K2 + O(ε) . (19) is true for n = m. Then
2 (2n−1 + 1)!
Setting xn = xK/√ε = 2 in (16) implies xm+1 = (p0 (m) + p1 (m)ε + p2 (m)ε2 + p3 (m)ε3

2 − 1/2 tan(2 ) κ 2n √
n m−1 1
√ = tan K − K + O( ε) (17) + · · · + p2m−1 (m)ε2 )2 + +ε (21)
ε n
(2 )! 4
 
1
which should be contrasted with setting the
√ solution = p20 (m) + + [2p0 (m)p1 (m) + 1]ε
of the differential equation x(n) = x(K/ ε) = 2 in 4
 
(15) to get Xj
2 − 1/2 + · · · +  pi (m)pj−i (m) εj + · · · (22)
√ = tan K . (18)
ε i=0
 π in the Mandelbrot Set 397

for j = 2, . . . , 2m . We have now shown that (19) Equating coefficients of n2j−2 gives
is true for all n = 1, 2, 3 . . . . We must now prove
1 X
the properties of the pj ’s hold. Since our induc- j−1
tion hypothesis includes p0 (m) = 1/2, (21) implies aj = ai aj−i
2j − 1 i=1
that p0 (m + 1) = (1/2)2 + 1/4 = 1/2 proving
that p0 (n) = 1/2 for all n = 1, 2, 3 . . . . Similarly,
we assumed that p1 (m) = m, so (21) implies that and a1 = 1 since b1 (n) = 1 · n. 
p1 (m + 1) = 2 · 12 · m + 1 proving that p1 (n) = n
for all n = 1, 2, 3 . . . . Since p0 (n) = 1/2 for all n, Next, we show that the aj ’s are coefficients for
the coefficient of εj in xn+1 = x2n + 14 + ε gives the the tangent series.
recurrence relation
Lemma 4. The nonzero coefficients of the tangent
X
j−1
series are also generated by (25).
pj (m + 1) = pj (m) + pi (m)pj−i (m) (23)
i=1
Proof. We can write
directly. Finally, (23) is equivalent to
∞
X
X X
m j−1
tan x = ck xk , −π/2 < x < π/2
pj (m + 1) = pi (k)pj−i (k) (24) k=1
k=1 i=1

since pj (1), the coefficient of εj in x1 , is zero. Since since tan x is analytic on −π/2 < x < π/2. Futher-
pi (m) and pj−i (m) are 2i − 1 and 2(j − i) − 1 degree more, since tan x is odd, ck = 0 for even k. We
polynomials, respectively, pi (m)pj−i (m) is a 2j − 2 can differentiate the convergent power series and
P we have
degree polynomial and so is j−1 i=1 pi (k)pj−i (k). It
follows then that the right-hand-side of Eq. (24) is d
a 2j − 1 degree polynomial.  tan x = sec2 x = 1 + tan2 x
dx
We saw in Eq. (16) that the lower order so that
terms in the polynomial coefficients pj (n) contri- ! !
bute to O(ε) terms which ultimately have no effect ∞
X ∞
X ∞
X
k−1 k k
on the result. What we need is the coefficient aj of kck x =1+ ck x ck x . (26)
the highest degree term in pj (n). k=1 k=1 k=1

Equating coefficients of xn for n = 2j = 2, 4, 6, . . .

Lemma 3. Define aj to be the coefficient of the
in (26) gives
highest degree term in pj (n). Then aj satisfies the
recurrence relation X
(2j + 1)c2j+1 = cm cl . (27)
1 X
j−1
m+l=2j
aj = ai aj−i , a1 = 1 . (25) m, l odd
2j − 1 i=1
If we then substitute ai = c2i−1 , (27) becomes
Proof. By definition, pj (n) = aj n2j−1 +cj n2j−2 +
X
j
· · · where cj is the constant coefficient of n2j−2 . (2j + 1)aj+1 = ai aj+1−i
From the recurrence relation (23) for pj (n), we have i=1

(aj n2j−1 + cj n2j−2 + · · · )

which completes the proof after replacing j with
X
j−1 j − 1. 
+ (ai n2i−1 + ci n2i−2 + · · · )
i=1 Having just shown that the leading coefficients
in pj (n) match the jth nonzero term in the tan-
× (aj−i n2(j−i)−1 + cj−i n2(j−i)−2 + · · · )
gent series, we have completed the proof that it
= aj (n + 1)2j−1 + cj (n + 1)2j−2 + · · · is valid to replace the difference equation wth the
differential equation for 1/2 ≤ xn , x(n) ≤ 2, n ≥
= aj n2j−1 + (2j − 1)aj n2j−2 + cj n2j−2 + · · · . 0. We next tackle the case of 0 ≤ xn , x(n) ≤
398 A. Klebanoff

1/2, n ≤ 0. Since we are restricted to x ≥ 0, the which is equivalent to (28) via yk = zk + 1/4 so that
map f (x) = x2 + 1/4 + ε has the inverse (29) becomes
q
f −1 (x) = x − 1/4 − ε . 1 X∞
zn = q0 (n) − + qj (n)εj . (32)
So, x−n = yn for n ≥ 0 if 4 j=1
q
yn+1 = yn − 1/4 − ε , y0 = 1/2 . (28) We first note that q0 (n) = 1/2 and q1 (n) =
−n follow trivially by mathematical induction
The first few iterations are (expanded in ε) upon substituting Eq. (32) into (31) and equating
q
y1 = 1/4 − ε coefficients of ε0 and ε1 to get
r
1 1
= 1/2 − ε − ε − 2ε − · · ·
2 3
q0 (n + 1) = q0 (n) − , q0 (1) =
4 2
q
y2 = 1/4 − 2ε − · · · and

= 1/2 − 2ε − 5ε2 − 22ε3 − · · · q1 (n + 1) = q1 (n) − 1 , q0 (1) = −1

q
y3 = 1/4 − 3ε − · · · respectively.
Next, we seek a recurrence relation for qj (n).
= 1/2 − 3ε − 14ε2 − 106ε3 − · · · In order to do this, we must know what the first
j + 1 termspare in the Taylor series expansion
q
for f (ε) = z(ε) − ε − 1/4. We denote the jth
y4 = 1/4 − 4ε − · · ·
derivative of f or z by f (j) or z (j) , respectively.
= 1/2 − 4ε − 30ε2 − 346ε3 − · · · .
Claim. For each j = 2, 3, 4, . . ., the jth derivative
Conveniently, we see patterns similar to those found of f (ε) is
in the case where x ≥ 1/2. We conjecture and prove  
the following. !
1 X j (i) (j−i) 
j−1
−1
f (j) (ε) = (z − ε)−3/2  z z
4 2 i=1 i
Lemma 5. The backwards solution of the differ-
ence equation xn+1 = x2n + 1/4 + ε, x0 = 1/2, k = 1
0, −1, −2, . . . (with 0 ≤ xk < 1/2)
p or equivalently + (z − ε)−1/2 z (j) + · · · (33)
2
the forward solution of yn+1 = yn − 1/4 − ε, y0 =
1/2 by substituting x−n = yn is where the terms denoted by “· · · ” are of the form
∞
X Y
yn = qj (n)εj (29) α [z (dk ) ]pk (34)
j=0 k

where q0 (n) = 1/2 and qj (n) is a 2j − 1 degree poly- where

nomial in n (for j = 1, 2, 3, . . .) that satisfies the X X
recurrence relation dk pk = j and pk ≥ 3 or (35)
k k
X
j−1
qj (n + 1) = qj (n) − qi (n)qj−i (n) + O(n2j−3 ) X
i=1
dk pk < j (36)
k
q1 (n) = −n , j = 2, 3, 4, . . . (30)
and α = γ(z − ε)−r/2 for some constant γ and odd
where O(n2j−3 ) denotes a polynomial of degree at
natural number r.
most 2j − 3. The careful reader has already noticed
the similarities with Lemma 2. P
Remark. If we define 0i=1 x(i) = 0, then the
claim holds for j = 1 as well.
Proof. We will find it more convenient to work
with the difference equation
√ Proof of Claim. We first note that the terms of
zn+1 = zn − ε − 1/4 , z0 = 1/4 (31) f (j)(ε) displayed in (33) do not satisfy condition
 π in the Mandelbrot Set 399

(35) or (36). The ith component of the first j − 1 Before going further, we note that the first term
terms in (33) can be written can be written as
!  
−1 j Y2 !
(z − ε)−3/2 z (i) z (j−i) = α [z (dk ) ]pk 3 1 X
j−1
j (i) (j−i) 
8 i (z − ε)−5/2 (z (1) − 1)  z z
k=1 8 2 i=1
i
−1 j


with α = −
8 i (z ε)−3/2 ,
d1 = i, d2 = j − i, and
P2 P X
j−1 Y
3 X
j−1 Y
2
p1 = p2 = 1. So, k=1 pk = 2 < 3 and 2k=1 dk = αi [z (dk ) ]pk − αi [z (dk ) ]pk
pk = j so that neither (35) nor (36) hold. The last i=1 k=1 i=1 k=1
term displayed in (33) term can be written


1 where αi = 16 3
(z − ε)−5/2 ji , d1 = i, d2 = j − i,
(z − ε)−1/2 z (j) = α[z (d1 ) ]p1 P3
2 d3 = 1, pkP= 1 for k = 1, 2, 3 so thatP k=1 dk pk =
2 2
P j + 1 and k=1 pk = 3 (first part) k=1 dk pk = j <
withP d1 = j and p1 = 1 so that 1k=1 pk = 1 < 3 j + 1 (second part). Also, the fourth term fits the
and 1k=1 dk pk = j. form
We next show that f (2) (ε) satisfies the claim. 1
Differentiating f gives (z − ε)−3/2 z (j) = α[z (d1 ) ]p1
4
1 P1
f (1) (ε) = (z − ε)−1/2 (z (1) − 1) so that k=1 dk pk = j < j + 1. All that is left is
2
and 1
f (j+1) (ε) = − (z − ε)−3/2
−1 4
f (2) (ε) = (z − ε)−3/2 (z (1) − 1)2  ! 
4 1X j
j−1

1 × [z (i+1) z (j−i) +z (i) z (j+1−i) ]

+ (z − ε)−1/2 z (2) 2 i=1 i
2
−1 1 1
= (z − ε)−3/2 [z (1) ]2 + (z − ε)−1/2 z (2) − (z − ε)−3/2 z (1) z (j)
4 2 4
1 1 1
+ (z − ε)−3/2 z (1) − (z − ε)−3/2 . (37) + (z − ε)−1/2 z (j+1) + · · · .
2 4 2
Taking j = 2 in the first displayed part of (33) gives We need, then, to show that
the first term in (37) and taking j = 2 in the second
displayed part of (33) gives the second term in (37). !
1 X j + 1 (i) (j+1−i)
j
The last two terms in (37) are of the form (34) and z z
both clearly satisfy condition (36). 2 i=1 i
We now proceed to assume the claim is true up !
1X j
j−1
to j and prove that it follows then for j + 1. We = [z (i+1) z (j−i) + z (i) z (j+1−i) ]
differentiate (33) to get 2 i=1 i
3
f (j+1) (ε) = (z − ε)−5/2 (z (1) − 1) + z (1) z (j) .
8
 ! 
1 X j (i) (j−i)  1
j−1 But
× z z − (z − ε)−3/2 !
2 i=1 i 4
1X j
j−1

  [z (i+1) z (j−i) + z (i) z (j+1−i) ] + z (1) z (j)

! 2 i=1 i
1X j
j−1
× [z (i+1) z (j−i) +z (i) z (j+1−i) ] ( !
2 i=1 i 1 X
j
j
= z (i) z (j+1−i)
1 1 2 i=2
i−1
− (z − ε)−3/2 z (1) z (j) + (z − ε)−3/2 z (j) ! )
4 4 X
j−1
j (i) (j+1−i)
1 + z z + z (1) z (j)
+ (z − ε)−1/2 z (j+1) + · · · . i=1
i
2
400 A. Klebanoff
( ! j−1 " ! !#
1 j (1) (j) X j j we have
= z z + +
2 1 i−1 i   −3
i=2
(j) −1 1 2
! ) f (0) =
j 4 4
× z (i) z (j+1−i) + z (j) z (1) + z (1) z (j)  ! 
j−1
1X j
j−1
! × i!qi (n)(j − i)!qj−i (n)
1 X j+1
j−1 2 i=1 i
= (j + 1)z (1) z (j) +
2 i=2 i   −1
1 1 2
! + j!bj (n) + g(j) (0)
1 X
j
j + 1 (i) (j+1−i) 2 4
× z (i) z (j+1−i) = z z .
2 i=1
i where

We must now show that the derivative of terms X Y

g(j) (0) = αi [z (dk ) ]pk
originally of the form (34) retain the form (34)
i k ε=0
under differentiation. Let
X  −r/2 Y
1
Y
m
(dk ) pk
= γi dk !qdpkk (n) .
g(ε) = α [z ] i
4 k
k=1
It follows then that for j = 2, 3, . . . ,
with properties for (34) satisfied along with condi-
tions (35) and (36). Then f (j)(0)
qj (n + 1) =
j!
d dα (1) Ym
g(ε) = (z − 1) [z (dk ) ]pk X
j−1
dε dε k=1 =− qi (n)qj−i (n) + qj (n)
X
m Y i=1
+α [z (dk ) ]pk pi [z (di ) ]pi −1 z (di +1) X 2r γi Y
i=1 i6=k + dk !qdpkk (n) (38)
i
j! k
| {z }
dα Ym Y
dα m+1
=− [z (dk ) ]pk + [z (dk ) ]pk remainder
dε k=1 dε k=1
| {z } | {z } where q1 (n) = −n. Since sums and products of
g1 g2
polynomials are polynomials, qj (n) is a polynomial.
X
m Y We now show that qj (n) is a 2j − 1 degree poly-
+α pi [z (dk ) ]pk [z (di ) ]pi −1 z (di +1) . nomial for j = 1, 2, 3, . . . . Since it is true for i = 1,
|
i=1 i6=k
{z }
we now assume true for i = 1, . . . , j − 1. Using the
g3 notation O(d) to denote a polynomial of degree d,
we have
Clearly, piece g1 satisfies whichever condition (35)
or (36) that g did which immediately implies qj (n + 1) − qj (n)
Pm Pm+1
Pk=1
d p < j + 1.
Pm
Piece g has k=1 k pk =
P
d
k k 2
= O(2i − 1) · O(2(j + 1) − 1)
k=1 dk pk = j, thenP k=1 pk ≥ 3
m m
k=1 dk pk + 1. If !
m+1
(since (35) must hold) and we have Pm k=1 k k
d p = X
j + 1.P If, on the other hand, k=1 dk p k < j, +O (2dk − 1)pk
m+1 k
then
P
d
k=1 k kp < j + 1. The ith term
Pm
in g3 has
!
d p
k6=i k k + d (p
i i − 1) + di + 1 = k=1 k k + 1.
d p X X
So, as for g2 , g3 still satisfies either condition (35) = O(2j − 2) + O 2 dk pk − pk .
or (36) with j replaced by j + 1. This completes the k k
proof of the claim. P
Recalling, now, that Conditions (35)P
and (36) imply
P
that either k
dk pk =Pj and P k p k ≥ 3 or k dk p k < j. Either
1 way, 2 k dk pk − k pk ≤ 2j − 3. So, the remainder
z= + q1 (n)ε + q2 (n)ε2 + q3 (n)ε3 + · · · term in (38) is at most a 2j − 3 degree polynomial
4
 π in the Mandelbrot Set 401

and the difference qj (n+1)−qj (n) is a 2j −2 degree of M). Another open problem is to determine the
polynomial: function of ε that multiplies N (ε). So far, we have

qj (n + 1) − qj (n) = β1 n2j−2 + β2 n2j−3 + · · · aεb N (ε) → π (39)

or equivalently, where we have seen a = 1, 2 and b = 1, 1/2. In gen-

eral, should we expect (39) to hold for some rational
X
n
values a and b? If so, what does the pinch location
qj (n + 1) = (β1 k2j−2 + β2 k2j−3 + · · · ) in M tell us about a and b?
k=1

which is a 2j −1 degree polynomial. This completes

the proof of Lemma 5. 

Finally, we use the fact that qj (n) is a 2j − 1

degree polynomial along with the fact that the re-
mainder terms in (38) are at most a 2j − 3 degree
P
polynomial to rewrite qj (n) − j−1 i=1 qi (n)qj−i (n) +
· · · = qj (n + 1) in (38) as

(aj n2j−1 + cj n2j−2 + · · · )

X
j−1
Fig. 4 The parabolic route c = (−1.25 − ε2 , ε) into
− (ai n2i−1 + ci n2i−2 + · · · ) the point (−1.25, 0) avoids crossing the decorations of
i=1 M and produces another π-result. The symmetric route
c = (−1.25 − ε2 , −ε) gives the same result.
× (aj−i n2(j−i)−1 + cj−i n2(j−i)−2 + · · · ) + |{z}
···
O(2j−3)
ACKNOWLEDGMENTS
= aj (n + 1)2j−1 + cj (n + 1)2j−2 + · · ·
This work was supported by the National Science
= aj n2j−1 + (2j − 1)aj n2j−2 + cj n2j−2 + · · · .
Foundation under grant DMS-9704639. I also thank
my friends and colleagues at the United States
Equating coefficients of n2j−2 gives
Naval Academy who helped and motivated me to
complete the proof.
−1 X
j−1
aj = ai aj−i
2j − 1 i=1
REFERENCES
and we already know that a1 = −1 since q1 (n) =
1. D. Boll, Pi and the Mandelbrot Set (Again), USENET
−1 · n. The proof of the “π-result” into the cusp of
article
M now follows from the lemmas!
[email protected].
2. D. Boll, Pi and the Mandelbrot Set,
http://www.frii.com/˜dboll/mandel.html.
3. CONCLUSION 3. H.-O. Peitgen, H. Jürgens and D. Saupe, Chaos and
Rather than attempt to complete the proof of Boll’s Fractals: New Frontiers of Science (Springer-Verlag,
New York, 1992).
vertical route shown in Fig. 2, we do something
4. G. Edgar, Pi and the Mandelbrot Set,
much easier. We conjecture that there are infinitely
http://www.math.ohio-state.edu/˜edgar/piand.html.
many such routes at each of the infinitely many 5. Y. Pomeau and P. Manneville, “Intermittent Tran-
pinches of M. In fact, in 1997, Jay Hill found that sition to Turbulence in Dissipative Dynamical
the parabolic route c = (−1.25−ε2 , ε) into the pinch Systems,” Commun. Math. Phys. 74, 189–197 (1980).
at (−1.25, 0) yields a “π-result”.4 We provide a 6. J. Guckenheimer and P. Holmes, “Nonlinear Oscilla-
table of our own experiment for this route in tions, Dynamical Systems, and Bifurcations of Vec-
Fig. 4. We know of no other reported routes to date tor Fields,” in Applied Mathematical Sciences, Series
(except the obvious routes given by the symmetry Vol. 42. (Springer-Verlag, New York, 1983).
402 A. Klebanoff

Qc (z) to be our next z, the result follows. 

Proof. Let |z0 | > 2 and |z0 | ≥ |c|. Then

def
|z1 | = |Qc (z0 )| = |z02 + c|
≥ |z0 |2 − |c| (triangle inequality)
≥ |z0 |2 − |z0 | (|z0 | ≥ |c|)
= |z0 |(|z0 | − 1)
= λ0 |z0 |

where λ0 = |z0 | − 1 > 1. This implies that |z1 | >

|z0 | > 2, so continuing we have
Fig. 5 If |z| > 2 and |z| ≥ |c|, then |Qc (z)| > |z|.
def
|z2 | = |Qc (z1 )| ≥ λ1 |z1 | ≥ λ1 λ0 |z0 | > λ20 |z0 |
APPENDIX
where we have defined λn = |zn | − 1 so that clearly
Theorem (Escape Criterion): If |z| > 2 and λn > λm for n > m. Mathematical induction on n
|z| ≥ |c|, then Qc (z) → ∞ as n → ∞. gives |zn | > λn0 |z0 |. Taking n → ∞ completes the
proof. 
Picture Proof. Figure 5 gives the main idea of
the proof: if |z| > 2, then |Qc (z)| > |z| > 2. Taking