J. London Math. Soc.

(2) 74 (2006) 397–414 e 2006 London Mathematical Society

C
doi:10.1112/S0024610706022988

A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE,

WITH APPLICATIONS TO COMPLEX DIMENSIONS

MICHEL L. LAPIDUS and ERIN P. J. PEARSE

Abstract
A formula for the interior ε-neighborhood of the classical von Koch snowflake curve is computed
in detail. This function of ε is shown to match quite closely with earlier predictions of what it
should be, but is also much more precise. The resulting ‘tube formula’ is expressed in terms of the
Fourier coefficients of a suitable nonlinear and periodic analog of the standard Cantor staircase
function and reflects the self-similarity of the Koch curve. As a consequence, the possible complex
dimensions of the Koch snowflake are computed explicitly.

1. Introduction
In [15], the authors lay the foundations for a theory of complex dimensions
with a rather thorough investigation of the theory of fractal strings (see also, e.g.,
[3, 4, 7–10, 13, 14, 16, 17]); that is, fractal subsets of R. Such an object may be
represented by a sequence of bounded open intervals of length lj :
∞


L := {lj }∞
j=1 , with lj < ∞. (1.1)
j=1

The authors are able to relate geometric and physical properties of such objects
through the use of zeta functions that contain geometric and spectral information
about the given string. This information includes the dimension and measurability
of the fractal under consideration, which we now recall.
For a nonempty bounded open set Ω ⊆ R, V (ε) is defined to be the inner
ε-neighborhood of Ω:
V (ε) := vol1 {x ∈ Ω : d(x, ∂Ω) < ε}, (1.2)
where vol1 denotes one-dimensional Lebesgue measure. Then the Minkowski dimen-
sion of the boundary ∂Ω (that is, of the fractal string L ) is
D = D∂Ω = inf{t
0 : V (ε) = O(ε1−t ) as ε → 0+ }. (1.3)
Finally, ∂Ω is Minkowski measurable if and only if the limit
M = M(D; ∂Ω) = lim V (ε)ε−(1−D) (1.4)
ε→0+

exists, and lies in (0, ∞). In this case, M is called the Minkowski content of ∂Ω.

Received 22 November 2004; revised 4 November 2005.

2000 Mathematics Subject Classification 26A30, 28A12, 28A75, 28A80 (primary), 11K55, 26A27,
28A78, 28D20, 42A16 (secondary).
The work of MLL was partially supported by the US National Science Foundation under grant
DMS-0070497. Support by the Centre Emile Borel of the Institut Henri Poincaré (IHP) in Paris
and by the Institut des Hautes Études Scientifiques (IHES) in Bures-sur-Yvette, France, during
completion of this article is also gratefully acknowledged by MLL.
398 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

If, more generally, Ω is an open subset of Rd , then analogous definitions hold

if 1 is replaced by d in (1.2)–(1.4). Hence, vold denotes the d-dimensional volume
(which is area for d = 2) in the counterpart of (1.2) and ε1−t , ε−(1−D) are replaced
by εd−t in (1.3), and ε−(d−D) in (1.4), respectively. In (1.2)–(1.4), d = 1 and the
positive numbers lj are the lengths of the connected components (open intervals)
of Ω, written in nonincreasing order. In much of the rest of the paper (where Ω
is the snowflake domain of R2 ), we have d = 2. See, for example, [7, 13–15, 18,
19, 22] and the relevant references therein for further information on the notions
of Minkowski–Bouligand dimension (also called ‘box dimension’) and Minkowski
content.
The complex dimensions of a fractal string L are defined to be the poles (of the
meromorphic continuation) of its geometric zeta function
∞


ζL (s) = ljs , (1.5)
j=1

in accordance with the result that

 ∞


DL = inf σ
0 : lj < ∞ ,
σ
(1.6)
j=1

that is, that the Minkowski dimension of a fractal string is the abscissa of conver-
gence of its geometric zeta function [8].
To rephrase, we define the complex dimensions of L to be the set
D = {ω ∈ C : ζL is defined and has a pole at ω}. (1.7)
One reason why these complex dimensions are important is the (explicit) tubu-
lar formula for fractal strings, a key result of [15]. Namely, that under suitable
conditions on the string L, we have the following tube formula:

(2ε)1−ω
V (ε) = cω + R(ε), (1.8)
ω(1 − ω)
ω∈D
where the sum is taken over the complex dimensions ω of L, and the error term
R(ε) is of lower order than the sum as ε → 0+ . (See [15, Theorem 6.1, p. 144].)
In the case when L is a self-similar string, that is, when ∂L is a self-similar subset of
R, one distinguishes two complementary cases (see [6, 9] and [15, § 2.3] for further
discussion of the lattice/nonlattice dichotomy) as follows.
(1) In the lattice case, that is, when the underlying scaling ratios are rationally
dependent, the error term is shown to vanish identically and the complex dimensions
lie periodically on vertical lines (including the line Re s = D).
(2) In the nonlattice case, the complex dimensions are quasiperiodically dis-
tributed and s = D is the only complex dimension with real part D. Estimates
for R(ε) are given in [15, Theorem 6.20, p. 154] and more precisely in [16, 17].
Also, L is Minkowski measurable if and only if it is nonlattice. See [15, Chapters 2
and 6] for details, including a discussion of quasiperiodicity.
These results pertain only to fractal subsets of R, although because this paper was
completed and using a different approach, some preliminary work on the extension
to the higher-dimensional case has been done in [11, 12, 20]. The foundational
special case of ‘fractal sprays’ [14] is discussed in [15, § 1.4].
It is the aim of the present paper to make some first steps in this direction.
We compute V (ε) for a well-known (and well-studied) example, the Koch snowflake,
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 399

∂Ω

K Ω

Figure 1. The Koch curve K and Koch snowflake domain Ω.

with the hope that it may help in the development of a general higher-dimensional
theory of complex dimensions. This curve provides an example of a lattice self-
similar fractal and a nowhere differentiable plane curve. Further, the Koch snowflake
can be viewed as the boundary ∂Ω of a bounded and simply connected open set
Ω ⊆ R2 and that it is obtained by fitting together three congruent copies of the
Koch curve K, as shown in Figure 1. A general discussion of the Koch curve may
be found in [18, § II.6] or [1, Introduction and Chapter 9].
The Koch curve is a self-similar fractal with dimension D := log3 4 (Hausdorff
and Minkowski dimensions coincide for the Koch curve) and may be constructed
by means of its self-similar √structure (as in [5, p. 15] or [1, Chapter 9]) as follows:
let  = 12 + 2√
1
3
i, with i = −1, and define two maps on C by
f1 (z) := z and f2 (z) := (1 − )(z − 1) + 1.
Then the Koch curve is the self-similar set of R2 with respect to {f1 , f2 }; that is,
the unique nonempty compact set K ⊆ R2 satisfying K = f1 (K) ∪ f2 (K).
In this paper, we prove the following new result.

Theorem 1.1. The area of the inner ε-neighborhood of the Koch snowflake is
given by the following tube formula:
V (ε) = G1 (ε)ε2−D + G2 (ε)ε2 , (1.9)
where D = log3 4 is the Minkowski dimension of ∂Ω, p := 2π/log 3 is the oscillatory
period, and G1 and G2 are periodic functions (of multiplicative period 3) which are
discussed in full detail in Theorem 5.1. This formula may also be written as



V (ε) = ϕn ε2−D−inp + ψn ε2−inp , (1.10)
n∈Z n∈Z

for suitable constants ϕn , ψn which depend only on n. These constants are expressed
in terms of the Fourier coefficients gα of a multiplicative function which bears
structural similarities to the classical Cantor–Lebesgue function, and is described
in more detail in § 6.

While this formula is new, it should be noted that a previous approximation has
been obtained in [15, § 10.3]; see (2.10) in Remark 2.3 below. Our present formula,
however, is exact. By reading off the powers of ε appearing in (1.10), we immediately
obtain the following corollary.
400 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

K1 K3

K2 K4

Figure 2. The first four stages in the geometric construction of K.

Corollary 1.2. The possible complex dimensions of the Koch snowflake are
D∂Ω = {D + inp : n ∈ Z} ∪ {inp : n ∈ Z}. (1.11)

This is illustrated later in Figure 8. Also, for more precision regarding Corol-
lary 1.2, see Remark 5.3 below, as well as the discussion surrounding (5.5).

Remark 1.3. The significance of the tube formula (1.9) is that it gives a
detailed account of the oscillations that are intrinsic to the geometry of the Koch
snowflake curve. More precisely, the real part D yields the order of the amplitude
of these oscillations (as a function of ε) while the imaginary part np = 2πn/log 3
(n = 0, 1, 2, . . . ) gives their frequencies. This is in agreement with the ‘philosophy’ of
the mathematical theory of the complex dimensions of fractal strings as developed
in [15]. In addition, if one can show the existence of a complex dimension with real
part D and imaginary part inp, n = 0, then Theorem 1.1 immediately implies that
the Koch curve is not Minkowski measurable, as conjectured in [9, Conjectures 2
and 3, pp. 159, 163–164].

The rest of this paper is dedicated to the proof of Theorem 1.1 (stated more
precisely as Theorem 5.1). More specifically, in § 2 we approximate the area V (ε)
of the inner ε-neighborhood. In § 3 we take into account the ‘error’ resulting from
this approximation. We study the form of this error in § 3.1, and the amount of it
in § 3.2. In § 4 we combine § 2 and § 3 to deduce the tube formula (1.9) in the more
precise form given in § 5. We also include in § 5 some comments on the interpretation
of Theorem 5.1. Finally, in § 6 we sketch the graph and briefly discuss some of the
properties of the Cantor-like and multiplicatively periodic function h(ε), the Fourier
coefficients of which occur explicitly in the expansion of V (ε) stated in Theorem 5.1.

2. Estimating the area

Consider an approximation to the inner ε-neighborhood of the Koch curve, as
shown in Figure 2. Although we will eventually compute the neighborhood for the
entire snowflake, we work with one third of it throughout the sequel (as depicted
in the figure).
We will determine the area of the ε-neighborhood with functions that give the
area of each kind of piece (rectangle, wedge, fringe, as shown in the figure) in terms
of ε, and functions that count the number of each of these pieces, in terms of ε.
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 401

triangles
overlapping (from overlap)
rectangles fringe

wedges
ε
error rectangles

Figure 3. An approximation to the inner ε-neighborhood of the Koch curve, with

ε ∈ I2 . The refinement level here is based on the graph K2 , the second stage in the
geometric construction of the Koch curve (see Figure 2).

Figure 4. A smaller ε-neighborhood of the Koch curve, for ε ∈ I3 . This refinement

level is based on the graph K3 , the third stage in the construction of the Koch curve.

As seen by comparing Figure 2 with Figure 4, the number of such pieces increases
exponentially.
We take the base of the Koch curve to have length 1, and carry out our approx-
imation for different values of ε, as ε → 0+ . In particular, let
√ √
In := (3−(n+1) / 3, 3−n / 3]. (2.1)
√
Whenever ε = 3−n / 3, the approximation shifts to the next level of refinement.
For example, Figure 2 shows ε ∈ I2 , and Figure 4 shows ε ∈ I3 . Consequently, for
ε ∈ I0 , it suffices to consider an ε-neighborhood of the prefractal curve K0 , and for
ε ∈ I1 , it suffices to consider an ε-neighborhood of the prefractal curve K1 , and so
on. Figure 2 shows these prefractal approximations.
In general, we will discuss a neighborhood of Kn , and we define the function
 
1
n = n(ε) := log3 √ = [x] (2.2)
ε 3
to tell us for what n we have ε ∈ In . Here, the square brackets indicate the floor
function (integer part) and
√
x := −log3 (ε 3), (2.3)
402 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

notation which will frequently prove convenient in the sequel. Further, we let
{x} := x − [x] ∈ [0, 1) (2.4)
denote the fractional part of x.
Observe that for ε ∈ In , n is fixed even as ε is changing. To see how this is useful,
consider that for all ε in this interval, the number of rectangles (including those
which overlap in the corners) is readily seen to be the fixed number
rn := 4n . (2.5)
Also, each of these rectangles has area ε3−n , where n is fixed as ε traverses In .
Continuing in this constructive manner, we prove the following lemma.

Lemma 2.1. For ε ∈ In , there are:

(i) rn = 4n rectangles, each with area ε3−n ;
(ii) wn = 23 (4n − 1) wedges, each with area πε2√ /6;
(iii) tn = 23 (4n + 2) triangles, each with area ε√
2
3/2; and
(iv) 4n components of fringe, each with area ( 3/20)9−n.

Proof. We have already established part (i).

For part (ii), we exploit the self-similarity of the Koch curve K to obtain the
recurrence relation wn = 4wn−1 + 2, which we solve to find the number of wedges


n−1
wn := 2 · 4j = 23 (4n − 1). (2.6)
j=0

The area of each wedge is clearly πε2 /6, as the angle is always fixed at π/3.
(iii) To prevent double-counting, we will need to keep track of the number of
rectangles that overlap in the acute angles so that we may subtract the appropriate
number of triangles


n−1
tn := 4n − 4j = 23 (4n + 2). (2.7)
j=1
√ 2
Each of these triangles has area ε 3/2.
(iv) To measure√the area of the fringe, note that the area under the entire Koch
curve is given by 3/20, so the fringe of Kn will be this number scaled by (3−n )2 .
There are 4n components, one atop each rectangle (see Figure 2).

Lemma 2.1 gives a preliminary area formula for the ε-neighborhood. Here, ‘pre-
liminary’ indicates the absence of the ‘error estimate’ developed in § 3.

Lemma 2.2. The ε-neighborhood of the Koch curve has the approximate area
 √ √    
 2−D −{x} 3 3 {x} 3 {x} 1 π √ ε2 π √
V (ε) = ε 4 9 + 3 + − 3 − + 2 3 . (2.8)
40 2 6 3 3 3

This formula is approximate in the sense that it measures a region slightly larger
than the actual ε-neighborhood. This discrepancy is accounted for and analyzed in
detail in § 3.
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 403
Proof of Lemma 2.2. Using (2.2) and (2.3), we obtain
 x √  x
1 4 3 1−D 4 3
4 = ε−D ,
x
9 −x 2
= 3ε , = ε , = ε2−D . (2.9)
2 3 2 9 2

Now using n = [x] = x − {x}, we compute the contributions of the rectangles,

wedges, triangles and fringe, respectively, as
 n √
 4 2−D 3
Vr (ε) = ε =ε · 4−{x} 3{x} ,
3 2
πε2 n π π
Vw (ε) = (4 − 1) = ε2−D · 4−{x} − ε2 ,
9√ 18√ 9 √
2
ε 3 3 2 3
Vt (ε) = n
(4 + 2) = ε 2−D
·4 −{x}
+ ε2 ,
3
 n  √  √6 3
4 3 3 3 −{x} {x}
Vf (ε) = = ε2−D ·4 9 .
9 20 40

Putting all this together, V = Vr + Vw − Vt + Vf gives the result.

Remark 2.3. It is pleasing to find that this is in agreement with earlier pre-
dictions of how V (ε) should look. In particular, [15, p. 209] gives the estimate
√  
3 −{x} 3 {x}
V (ε) ≈ ε2−D 4 9 + 6 · 3{x} − 1 , (2.10)
4 5

which differs only from our formula for V (ε) in (2.8) by some constants and the
final term of order ε2 .

In (3.9) we will require the Fourier series of the periodic function ε−(2−D) V (ε)
as given by (2.8), so we recall the formula

a−1
e2πinx
a−{x} = . (2.11)
a log a + 2πin
n∈Z

This formula is valid for a > 0, a = 1 and has been used repeatedly in [15].
Note that it follows from Dirichlet’s theorem and thus holds in the sense of Fourier
series. In particular, the series in (2.11) converges pointwise; this is also true in
(2.13) and (2.14) below.
We will make frequent use of the following identity in the sequel:
√
e2πinx = (ε 3)−inp = (−1)n ε−inp , for n ∈ Z, (2.12)

where p√ = 2π/log 3 is the oscillatory period as in Theorem 1.1. With x =

− log3 (ε 3), we can rewrite the Fourier expansion of a−{x} given in (2.11) as

a − 1
(−1)n ε−inp
a−{x} = . (2.13)
a log 3 log3 a + inp
n∈Z
404 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

A3 A2 A3 A1 A3 A2 A3

Figure 5. An error block for ε ∈ In . The central third of the block contains
one large isosceles triangle, two wedges and the trianglet A1 .

Recall that {x} = x − [x] denotes the fractional part of x. With D = log3 4, we
use (2.13) to express (2.8) as a pointwise convergent Fourier series in ε:

1
−35/2 33/2
V (ε) = +
3 log 3 25 (D − 2 + inp) 23 (D − 1 + inp)
n∈Z
  
π − 33/2 n 2−D−inp 1 π √
+ 3 (−1) ε − + 2 3 ε2 . (2.14)
2 (D + inp) 3 3

3. Computing the error

Now we must account for all of the little ‘trianglets’, the small regions shaped
like a crest of water on an ocean wave. These regions were included in our original
calculation, but now must be subtracted. This error appeared in each of the rectan-
gles counted earlier, and so we refer to all of the error from one of these rectangles
as an ‘error block’. Figure 5 shows how this error is incurred and how it inherits a
Cantoresque structure from the Koch curve. Actually, we will see in § 3.2 that this
‘error’ will have several terms, some of which are of the same order as the leading
term in V (ε), which is proportional to ε2−D by (2.14). Hence, caution should be
exercised when carrying out such computations and one should not be too quick to
set aside terms that appear negligible.

3.1. Finding the area of an ‘error block’

In calculating the error, we begin by finding the area of one of these error blocks.
Later, we will count how many of these error blocks there are, as a function of ε.
Note that n is fixed throughout § 3.1, but ε varies within In . We define the function

w = w(ε) := 3−n = 3−[x] . (3.1)

This function w(ε) gives the width of one of the rectangles, as a function of ε (see
Figure 6). Note that w(ε) is constant as ε traverses In , as is n = n(ε). From Figure 6,
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 405
w(ε)=3−n

θ
H ε

b=_16 w(ε)

Figure 6. Finding the height of the central triangle.

one can work out that the area of both wedges adjacent to Ak is
 
w
ε2 sin−1 ,
2 · 3k ε
and that the area of the triangle above Ak is
 2
w w
ε 1 − .
2 · 3k 2 · 3k ε
Then for k = 1, 2, . . . , the area of the trianglet Ak is given by

   2
w(ε) w(ε) w(ε) w(ε)
Ak (ε) = ε k − ε2 sin−1 − ε 1 − , (3.2)
3 2 · 3k ε 2 · 3k 2 · 3k ε
k−1
and appears with multiplicity
√ 1 −{x}2 , as in Figure 5. We use (3.1) and (2.3) to write
−x {x}
w(ε) = 3 3 = ε 3( 3 ) , and define
w
3xk := k = 3{x}−k+1/2 . (3.3)
3 ε
Hence, the entire contribution of one error block may be written as
  x  x 2 

∞ x
−1 3 3 3k
B(ε) := 2 k−1
3k − sin
x k
− k
1− ε2 . (3.4)
2 2 2
k=1
Recall the power series expansions
∞

(2m)! u2m+1
∞

(2m)! u2m+2
sin−1 u = and 1 − u2 = 1 − ,
m=0
22m (m!)2 (2m + 1) m=0
22m+1 m!(m + 1)!
which are valid for |u| < 1. We use these formulae with u = w/(2 · 3k ε), so
convergence is guaranteed by
√ √
w 3{x} 3 3
0 =  < 1,
2·3 εk 2·3 k 2
and the fact that the series in (3.4) starts with k = 1. Then (3.4) becomes

∞  x
∞
∞ 
3 (2m)! (3xk )2m+3 (2m)! (3xk )2m+1
B(ε) = 2k−1 k + 4m+4 m!(m + 1)!
− 4m+1 (m!)2 (2m + 1)
ε2
2 m=0
2 m=0
2
k=1

∞ 
∞ x 2m+1
∞ 
(2m − 2)! (3k ) (2m)! (3xk )2m+1
= 2k−1 4m (m − 1)!m!
− 4m+1 (m!)2 (2m + 1)
ε2
m=1
2 m=1
2
k=1
406 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

partial

complete

Figure 7. Error block formation. The ends are counted as partial because three
of these pieces will be added together to make the entire snowflake.

∞


∞  
2k−1 (2m − 2)! (3{x}+1/2 )2m+1 (2m − 1)2m 2
= 1 − ε
m=1 k=1
(32m+1 )k 24m (m − 1)!m! 2m(2m + 1)

∞  
1 1
=
m=1
32m+1 (32m+1 − 2)/32m+1
√  −{x}
(2m − 2)! ( 3)2m+1 1
× 4m−1 ε2
2 (m − 1)!m!(2m + 1) 32m+1

∞  −{x}
(2m − 2)! 1
= 4m−1 (m − 1)!m!(2m + 1)(32m+1 − 2) 32m+1
ε2 . (3.5)
m=1
2
The interchange of sums is validated by checking absolute convergence of the final
series via the ratio test, and then applying Fubini’s theorem to retrace our steps.

3.2. Counting the error blocks

Some blocks are present in their entirety as ε traverses an interval In , while
others are in the process of forming: two in each of the peaks and one at each end
(see Figure 7). Using the same notation as in Lemma 2.1, we count the complete
and partial error blocks with
cn = rn − tn = 13 (4n − 4) and pn = tn = 23 (4n + 2).
By means of (2.2)–(2.4), we convert cn and pn into functions of the continuous
variable ε, where ε > 0:
   
1 ε−D −{x} 2 ε−D −{x}
c(ε) = 4 −4 and p(ε) = 4 +2 . (3.6)
3 2 3 2
With B(ε) given by (3.5), the total error is thus
E(ε) = B(ε)[c(ε) + p(ε)h(ε)]. (3.7)

Remark 3.1. The function h(ε) in (3.7) is some periodic function that oscillates
multiplicatively in a region bounded between 0 and 1, indicating what portion of the
partial error block has formed; see Figures 5 and 7. We do not know h(ε) explicitly,
but we do know by the self-similarity of K that it has multiplicative period 3;
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 407
that is, h(ε) = h(ε/3). Using (2.12), the Fourier expansion



h(ε) = gα (−1)α ε−iαp = gα e2πiαx = g(x) (3.8)
α∈Z α∈Z

shows that we may also √ consider h(ε) as an additively periodic function of the
variable x = − log3 (ε 3), with additive period 1. We refer the interested reader to
§ 6 below for a further discussion of h(ε), including a sketch of its graph, justification
of the convergence of (3.8) and a brief discussion of some of its properties.

We now return to the computation; substituting (3.5) and (3.6) into (3.7) gives
 −D   
ε −{x} 1 4
E(ε) = B(ε) 4 + h(ε) + (h(ε) − 1)
3 2 3
∞  −{x}
1
(2m − 2)!(h(ε) + 1/2) 4
= ε2−D
3 m=1 24m−1 (m − 1)!m!(2m + 1)(32m+1 − 2) 32m+1
∞  −{x}
1
(2m − 2)!(h(ε) − 1) 1
+ ε2
3 m=1 24m−3 (m − 1)!m!(2m + 1)(32m+1 − 2) 32m+1
∞
1

(2m − 2)!(4 − 32m+1 )(−1)n (h(ε) + 1/2)ε2−D−inp
= 4m+1
3 log 3 m=1 2 (m − 1)!m!(2m + 1)(32m+1 − 2)(D − 2m − 1 + inp)
n∈Z
∞
1

(2m − 2)!(1 − 32m+1 )(−1)n (h(ε) − 1)ε2−inp
+
3 log 3 m=1 24m−3 (m − 1)!m!(2m + 1)(32m+1 − 2)(−2m − 1 + inp)
n∈Z

1

= (h(ε) + 1/2)(−bn )(−1)n ε2−D−inp
3 log 3
n∈Z

1

+ (h(ε) − 1)(−τn )(−1)n ε2−inp , (3.9)
3 log 3
n∈Z

where we have written the constants bn and τn in the shorthand notation as follows:
∞

(2m − 2)!(32m+1 − 4)
bn := , (3.10)
m=1
24m+1 (m − 1)!m!(2m + 1)(32m+1 − 2)(D − 2m − 1 + inp)
∞

(2m − 2)!(32m+1 − 1)
τn := . (3.11)
m=1
24m−3 (m − 1)!m!(2m + 1)(32m+1 − 2)(−2m − 1 + inp)

In the third equality of (3.9), we have applied (2.13) to a = 4/32m+1 and to

a = 1/32m+1 , respectively. By the ratio test, the complex numbers bn and τn given
by (3.10) and (3.11) are well defined. This fact, combined with Fubini’s theorem,
enables us to justify the interchange of sums in the last equality of (3.9).

4. Computing the area

Now that we have estimate (2.14) for the area of the neighborhood of the Koch
curve V (ε), and formula (3.9) for the ‘error’ E(ε), we can find the exact area of the
408 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

inner neighborhood of the full Koch snowflake as follows:

V (ε) = 3(V (ε) − E(ε))

1
−35/2 33/2
= 5
+ 3
log 3 2 (D − 2 + inp) 2 (D − 1 + inp)
n∈Z

π − 33/2
+ 3 (−1)n ε2−D−inp
2 (D + inp)
1

+ (h(ε) + 1/2)(−1)nbn ε2−D−inp
log 3
n∈Z
 
π √ 1

− + 2 3 ε2 + (h(ε) − 1)(−1)n τn ε2−inp
3 log 3
n∈Z

 5/2
1 −3 33/2 π − 33/2
= 5
+ 3 + 3
log 3 2 (D − 2 + inp) 2 (D − 1 + inp) 2 (D + inp)
n∈Z

bn
+ + h(ε)bn (−1)n ε2−D−inp
2
   
1
π √
+ −τn − log 3 + 2 3 δ0 + h(ε)τn (−1)n ε2−inp ,
n
(4.1)
log 3 3
n∈Z

where δ0n is the Kronecker delta. Therefore, we have

V (ε) = G1 (ε)ε2−D + G2 (ε)ε2 , (4.2)
where the periodic functions G1 and G2 are given by
1

G1 (ε) := (an + bn h(ε))(−1)n ε−inp , (4.3)
log 3
n∈Z
1

G2 (ε) := (σn + τn h(ε))(−1)n ε−inp . (4.4)
log 3
n∈Z

Here we have used (3.10) and (3.11), and introduced the notation an and σn (stated
explicitly in (5.3)). We wish to rearrange these series so as to collect all factors of ε.
First, we can split the sum in (4.3) as
1
h(ε)

G1 (ε) = an (−1)n ε−inp + bn (−1)n ε−inp (4.5)
log 3 log 3
n∈Z n∈Z
 n −inp
 n −inp
because a(ε) := n∈Z an (−1) ε and b(ε) := n∈Z bn (−1) ε are each
convergent: a(ε) converges for the same reason as (2.14), and one can show that

∞  −{x}
(4 log 3)(2m − 2)! 4
b(ε) = (4.6)
m=0
24m+1 3m−1/2 (m − 1)!m!(2m + 1)(32m+1 − 2) 32m+1
converges to a well-defined distribution induced by a locally integrable function;
one proves directly that |bn |  c/|n| by writing (3.10) as
∞

∞


βm
bn = , with βm < ∞. (4.7)
m=1
D − 2m − 1 + inp m=1

Then the rearrangement leading to (4.6) is justified via the ‘descent method’ with
q = 2, as described in Remark 4.1 below. Note that the right-hand side of (4.6)
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 409
converges by the ratio test and is thus defined pointwise on R ∼ Z. Therefore, one
also sees that b(ε) and √ h(ε) are periodic functions. Considered as functions of the
variable x = log3 (1/ε 3), both have period 1 with b continuous for 0  x < 1 and
h continuous for 0 < x  1. Further, each is monotonic on its period interval, and
possesses a bounded jump discontinuity only at the endpoint. This may be seen for
b(ε) from (4.6) and for h(ε) from § 6.
Although we have used distributional arguments to obtain (4.6), the right-hand
side of (4.6) is locally integrable and has a representation as a piecewise continuous
function. Thus, the distributional equality in (4.6) actually holds pointwise in the
sense of Fourier series and all of our results are still valid pointwise. Recall that
the Dirichlet–Jordan theorem [24, Theorem II.8.1] states that if f is periodic and
(locally) of bounded variation, its Fourier series converges pointwise to (f (x−) +
f (x+))/2. As described above, b(ε) and h(ε) are each of bounded variation and
therefore by [24, Theorem II.4.12] we have bn = O(1/n) and gn = O(1/n) as n →
±∞. Finally, [24, Theorem IX.4.11] may be applied to yield the pointwise equality



b(ε)h(ε) = bα gn−α (−1)n ε−inp . (4.8)
n∈Z α∈Z

This theorem applies because (4.6) shows that b(ε) is bounded away from 0.
Now that all the ε are combined, we substitute (4.8) back into (4.5) and rewrite
G1 (ε) = (a(ε) + b(ε)h(ε))/log 3 as in (5.2a):
 
1


G1 (ε) = an + bα gn−α (−1)n ε−inp . (4.9)
log 3
n∈Z α∈Z

Manipulating G2 (ε) similarly, we are able to rewrite (4.4) in its final form (5.2b),
and thereby complete the proof of Theorem 5.1.

Remark 4.1. The convergence of the Fourier series associated to a periodic

distribution is proved via the descent method by integrating both sides q times
(for sufficiently large q) so that one has pointwise convergence. After enough inte-
grations, the distribution will be a smooth function, the series involved will converge
absolutely and the Weierstrass theorem can be applied pointwise. At this point,
rearrangements or interchanges of series are justified pointwise, and we obtain
a pointwise formula for the qth antiderivative of the desired function. Then one
takes the distributional derivative q times to obtain the desired formula. See [15,
Remark 4.14]. How large the positive integer q needs to be depends on the order
of polynomial growth of the Fourier coefficients. Recall that the Fourier series of a
periodic distribution converges distributionally if and only if the Fourier coefficients
are of slow growth, that is, do not grow faster than polynomially. Moreover, from the
point of view of distributions, there is no distinction to be made between convergent
trigonometric series and Fourier series. See [21, § VII, I].

5. Main results
We can now state our main result in the following more precise form of
Theorem 1.1.
410 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

Theorem 5.1. The area of the inner ε-neighborhood of the Koch snowflake is
given pointwise by the following tube formula:
V (ε) = G1 (ε)ε2−D + G2 (ε)ε2 , (5.1)
where G1 and G2 are periodic functions of multiplicative period 3, given by
 
1


G1 (ε) := an + bα gn−α (−1)n ε−inp , (5.2a)
log 3
n∈Z α∈Z
 
1


G2 (ε) := σn + τα gn−α (−1)n ε−inp , (5.2b)
log 3
n∈Z α∈Z

where an , bn , σn and τn are the complex numbers given by

35/2 33/2 π − 33/2 1
an = − + + + bn ,
25 (D − 2 + inp) 23 (D − 1 + inp) 23 (D + inp) 2

∞
(2m)! (32m+1 − 4)
bn = ,
42m+1 (m!)2 (4m2 − 1)(32m+1 − 2)(D − 2m − 1 + inp)
m=1
 
π √
σn = −log 3 + 2 3 δ0n − τn , (5.3)
3

∞
(2m)! (32m+1 − 1)
τn = ,
m=1
42m−1 (m!)2 (4m2 − 1)(32m+1 − 2)(−2m − 1 + inp)
where δ00 = 1 and δ0n = 0 for n = 0 is the Kronecker delta.

In Theorem 5.1, D = log3 4 is the Minkowski dimension of the Koch snowflake

∂Ω and p = 2π/log 3 is its oscillatory period, following the terminology of [15].
The numbers gα appearing in (5.2) are the Fourier coefficients of the periodic
function h(ε), a suitable nonlinear analog of the Cantor–Lebesgue function, defined
in Remark 3.1 and further discussed in § 6 below.
The reader may easily check that formula (5.1) can also be written



V (ε) = ϕn ε2−D−inp + ψn ε2−inp , (5.4)
n∈Z n∈Z

for suitable constants ϕn , ψn which depend only on n. Now by analogy with the tube
formula (1.8) from [15], we interpret the exponents of ε in (5.4) as the ‘complex co-
dimensions’ of ∂Ω. Hence, we can simply read off the possible complex dimensions
and, as depicted in Figure 8, we obtain a set of possible complex dimensions
D∂Ω = {D + inp : n ∈ Z} ∪ {inp : n ∈ Z}. (5.5)
One caveat should be mentioned: this is assuming that none of the coefficients
ϕn or ψn vanish in (5.4). Indeed, in that case we would only be able to say that
the complex dimensions are a subset of the right-hand side of (5.5). We expect
that, following the ‘approximate tube formula’ obtained in [15], the set of complex
dimensions should contain all numbers of the form D + inp.

Remark 5.2. We expect our methods to work for all lattice self-similar fractals
(that is, those for which the underlying scaling ratios are rationally dependent)
as well as for other examples considered in [15] including the Cantor–Lebesgue
curve, a self-affine fractal. For example, we have already obtained the counterpart
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 411

2p

−4 −3 −2 −1 0 D 2 3 4

Figure 8. The possible complex dimensions of K and ∂Ω. The Minkowski

dimension is D = log3 4 and the oscillatory period is p = 2π/log 3.

of Theorem 5.1 for the square (rather than triangular) snowflake curve. By applying
density arguments (as in [15, Chapter 2]), our methods may also yield information
about the complex dimensions of nonlattice fractals.

Remark 5.3. It should be pointed out that in the present paper, we do not
provide a direct definition of the complex dimensions of the Koch snowflake curve
∂Ω (or of other fractals in R2 ). Instead, we reason by analogy with formula (1.8)
above to deduce from our tube formula (5.1) the possible complex dimensions of
∂Ω (or of K). As is seen in the proof of Theorem 5.1, the tube formula for the
Koch curve K is of the same form as for that of the snowflake curve ∂Ω. It follows
that K and ∂Ω have the same possible complex dimensions (see Corollary 1.2).
In subsequent work, we plan to define the geometric zeta function ζ∂Ω = ζ∂Ω (s) in
this context to actually deduce the complex dimensions of ∂Ω directly as the poles
of the meromorphic continuation of ζ∂Ω .
In fact, this project is well underway. A self-similar tiling defined in terms of an
iterated function system on Rd is constructed in [20]. In [12], we use this tiling
and aspects of geometric measure theory discussed in [11] to define a zeta function
whose poles give the complex dimensions directly. We then use the zeta function
and complex dimensions to obtain an explicit distributional inner tube formula for
self-similar systems in Rd , analogous to [15, Chapter 6]. We stress that although
Theorem 5.1 was helpful in developing [11, 12, 20], it is not a consequence or
corollary of this more recent work.

Remark 5.4. In the long-term, by analogy with Weyl’s tube formula [23]
for smooth Riemannian submanifolds (see [2]), we would like to interpret the
coefficients an , bn , σn , τn of the tube formula (5.1) in terms of an appropriate sub-
stitute of the ‘Weyl curvatures’ in this context. See the corresponding discussion in
[15, § 6.1.1 and § 10.5] for fractal strings; see also [17].
412 MICHEL L. LAPIDUS AND ERIN P. J. PEARSE

This is a very difficult open problem and is still far from being resolved, even in
the one-dimensional case of fractal strings. See [15, § 6.1.1] and [11, 12].

Remark 5.5 (Reality principle). As is the case for the complex dimensions of
self-similar strings (see [15, Chapter 2] and [16, 17]), the possible complex dimen-
sions of ∂Ω come in complex conjugate pairs, with attached complex conjugate
coefficients. Indeed, since gα = g−α (see § 6), a simple inspection of the formulas in
(5.3) shows that for every n ∈ N,
an = a−n , bn = b−n , σn = σ−n and τn = τ−n . (5.6)
It follows that a0 , b0 , σ0 and τ0 are reals and that G1 and G2 in (5.2) are real-valued,
in agreement with the fact that V (ε) represents an area.

6. The Cantor-like function h(ε)

We close this paper by further discussing the nonlinear Cantor-like function



h(ε) = gα (−1)α ε−iαp = gα e2πiαx = g(x), (6.1)
α∈Z α∈Z

introduced in (3.8). Note that since h is real valued (in fact, 0  h < µ < 1),
we have g−α = gα for all α ∈ Z. Further, recall from § 3.2 that in view of the
self-similarity of K, h(ε) is multiplicatively periodic with period 3, that is, h(ε) =
h(ε/3). Alternatively, (6.1) shows that it can be thought of as an additively periodic
function of x with period 1, that is, g(x) = g(x + 1). By the geometric definition
of h(ε), we see that it is continuous and even monotonic when restricted to one of
its period intervals In := (3−n−3/2 , 3−n−1/2 ]. Since h is of bounded variation, its
Fourier series converges pointwise by [24, Theorem II.8.1] and its Fourier coefficients
satisfy
gα = O(1/α), as α → ±∞ (6.2)
by [24, Theorem II.4.12], as discussed in § 4.
Further, since h(ε) is defined as a ratio of areas (see Remark 3.1), we have h(ε) ∈
[0, µ) for all ε > 0. Note that h(ε)  µ < 1 and so h(ε) does not attain the value 1;
the error blocks being formed are never complete. The partial error blocks only
form across the first 23 of √ the line segment beneath them. They reach this point
precisely when ε = 3−n / 3 for some n
1. Back in (3.4), we found the error of a
single error block to be given by
∞


B(ε) = 2k−1 Ak (ε),
k=1

where Ak (ε) is given by (3.2). Thus, the supremum of h(ε) will be the ratio
 
B(εk ) − A1 (εk )
+ A1 (εk ) B(εk ),
2
which will be the same constant for each εk = 3−k−1/2 , k = 1, 2, . . . (see Figure 9).
In other words, the number we need is
 ∞
A1 (εk ) + 12 ∞k=2 2
k−1
Ak (εk ) A1 (εk ) + k=2 2k−2 Ak (εk )
µ := ∞ k−1 = ∞ k−1 ∈ (0, 1).
k=1 2 Ak (εk ) k=1 2 Ak (εk )
 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE 413

C(ε)

B(ε)

Figure 9. The ratio µ is vol2 (C(ε))/vol2 (B(ε)).

µ µ
~
h(ε)

h(ε)

1 1 1 1 1 1
9 3 3 3 √3 9 3 3 3 √3

Figure 10. A comparison between the graph of the Cantor-like function h and the
graph of its approximation h̃.

Note that although this definition of µ initially appears to depend on k, the ratio
in question is between two areas that have exactly the same proportion at each εk ;
this is a direct consequence of the self-similarity of the Koch curve. In other words,
if C(ε) is the area indicated in Figure 9, then the relations
   
ε ε
3C = C(ε) and 3B = B(ε)
3 3
show that µ is well defined.
We now approximate h(ε) by a function which shares its essential properties:
(i) h(εk ) = limϑ→0− h(εk + ϑ) = 0;
(ii) limϑ→0+ h(εk + ϑ) = µ; √
where for k = 1, 2, . . . , εk = 3−k / 3 again. That is, h(ε) goes
√ from 0 to µ as ε goes
from 3−k to 3−(k+1) . Using again the notation x = − log3 (ε 3) and {x} = x − [x],
we see that the function
h̃(ε) = µ · {−[x] − x} (6.3)
shares both of these properties but is much smoother. Indeed, h̃(ε) only has points
of nondifferentiability at each εk and is otherwise a smooth logarithmic curve; see
Figure 10. The true h(ε), by contrast, is a much more complex object that deserves
further study in later work.

Acknowledgements. The authors are grateful to Machiel van Frankenhuysen for

his comments on a preliminary version of this paper. In particular, for indicating
the current, more elegant and concise presentation of the main result (5.1). We also
wish to thank Victor Shapiro for a helpful discussion concerning Fourier series,
especially with regard to the discussion surrounding (4.8).
414 A TUBE FORMULA FOR THE KOCH SNOWFLAKE CURVE

References
1. K. J. Falconer, Fractal geometry – mathematical foundations and applications (Wiley,
Chichester, 1990).
2. A. Gray, Tubes, 2nd edn, Progress in Mathematics 221 (Birkhäuser, MA, Boston, 2004).
3. B. M. Hambly and M. L. Lapidus, ‘Random fractal strings: their zeta functions, complex
dimensions and spectral asymptotics’, Trans. Amer. Math. Soc. 358 (2006) 285–314.
4. C. Q. He and M. L. Lapidus, ‘Generalized Minkowski content, spectrum of fractal drums,
fractal strings and the Riemann zeta-function’, Mem. Amer. Math. Soc. 127 (1997) no. 608,
1–97.
5. J. Kigami, Analysis on fractals (Cambridge University Press, Cambridge, 2001).
6. S. P. Lalley, ‘Packing and covering functions of some self-similar fractals’, Indiana Univ.
Math. J. 37 (1988) 699–709.
7. M. L. Lapidus, ‘Fractal drum, inverse spectral problems for elliptic operators and a partial
resolution of the Weyl–Berry conjecture’, Trans. Amer. Math. Soc. 325 (1991) 465–529.
8. M. L. Lapidus, ‘Spectral and fractal geometry: from the Weyl–Berry conjecture for the
vibrations of fractal drums to the Riemann zeta-function’, Proceedings of the 4th
UAB International Conference on Differential Equations and Mathematical Physics
(ed. C. Bennewitz; Academic Press, New York, 1992) 151–182.
9. M. L. Lapidus, ‘Vibrations of fractal drums, the Riemann hypothesis, waves in fractal
media, and the Weyl–Berry conjecture’, Proceedings of the 12th International Conference
on Ordinary and Partial Differential Equations, vol. IV, Pitman Research Notes in
Mathematics Series 289 (ed. B. D. Sleeman and R. J. Jarvis; Longman Scientific and
Technical, London, 1993) 126–209.
10. M. L. Lapidus and H. Maier, ‘The Riemann hypothesis and inverse spectral problems for
fractal strings’, J. London Math. Soc. (2) 52 (1995) 15–34.
11. M. L. Lapidus and E. P. J. Pearse, ‘Curvature measures and tube formulas of compact
sets’, in preparation.
12. M. L. Lapidus and E. P. J. Pearse, ‘Tube formulas and complex dimensions of self-similar
tilings’, Preprint, 2006, arXiv:math.DS/0605527.
13. M. L. Lapidus and C. Pomerance, ‘The Riemann zeta-function and the one-dimensional
Weyl–Berry conjecture for fractal drums’, Proc. London Math. Soc. (3) 66 (1993) 41–69.
14. M. L. Lapidus and C. Pomerance, ‘Counterexamples to the modified Weyl–Berry conjecture
on fractal drums’, Math. Proc. Cambridge Philos. Soc. 119 (1996) 167–178.
15. M. L. Lapidus and M. van Frankenhuysen, Fractal geometry and number theory: complex
dimensions of fractal strings and zeros of zeta functions (Birkhäuser, Boston, MA, 2000).
16. M. L. Lapidus and M. van Frankenhuysen, ‘Complex dimensions of self-similar fractal
strings and Diophantine approximation’, J. Exp. Math. 12 (2003) 41–69.
17. M. L. Lapidus and M. van Frankenhuysen, ‘Fractality, self-similarity and complex dimen-
sions’, Fractal geometry and applications: a jubilee of Benoı̂t Mandelbrot, Proceedings of
Symposia in Pure Mathematics 72, Part 1 (American Mathematical Society, Providence,
RI, 2004) 349–373.
18. B. B. Mandelbrot, The fractal geometry of nature (W.H. Freeman, New York, 1983)
19. P. Mattila, Geometry of sets and measures in Euclidean spaces. Fractals and rectifiability
(Cambridge University Press, Cambridge, MA, 1995).
20. E. P. J. Pearse, ‘Canonical self-similar tilings by IFS’, Preprint, 2005,
arXiv:math.MG/0606111.
21. L. Schwartz, Théorie des distributions (Hermann, Paris, 1966).
22. C. Tricot, Curves and fractal dimensions (Springer, New York, 1995).
23. H. Weyl, ‘On the volume of tubes’, Amer. J. Math. 61 (1939) 461–472.
24. A. Zygmund, Trigonometric series (Cambridge University Press, Cambridge, 1959)

Michel L. Lapidus and Erin P. J. Pearse

Department of Mathematics
University of California
Riverside, CA 92521-0135
USA
[email protected]
[email protected]