A spectral study of the Minkowski Curve

Nizare Riane, Claire David

October 12, 2018

Sorbonne Université
CNRS, Laboratoire Jacques-Louis Lions, 4, place Jussieu 75005, Paris, France
arXiv:1711.10350v3 [math.AP] 4 Feb 2018

Abstract
In the following, we give an explicit construction of a Laplacian on the Minkowski curve, with
energy forms that bear the geometric characteristic of the structure. The spectrum of the Laplacian
is obtained by means of spectral decimation.

Keywords: Laplacian - Minkowski Curve - Einstein relation - Spectral decimation.

AMS Classification: 37F20- 28A80-05C63.

1 Introduction
The so-called Minkowski curve, a fractal meandering one, whose origin seems to go back to Hermann
Minkowski, but is nevertheless difficult to trace precisely, appears as an interesting fractal object, and
a good candidate for whom aims at catching an overview of spectral properties of fractal curves.

The curve is obtained through an iterative process, starting from a straight line which is replaced
by eight segments, and, then, repeating this operation. One may note that the length of the curve
grows faster than the one of the Koch one. What are the consequences in the case of a diffusion process
on this curve ?

The topic is of interest. One may also note that, in electromagnetism, fractal antenna, specifically,
on the model of the aforementioned curve, which miniaturized design turn out to perfectly fit wideband
or broadband transmission, are increasingly used.

It thus seemed interesting to us to build a specific Laplacian on those curves. To this purpose, the
analytical approach initiated by J. Kigami [Kig89], [Kig93], taken up, developed and popularized by
R. S. Strichartz [Str99], [Str06], appeared as the best suited one. The Laplacian is obtained through
a weak formulation, by means of Dirichlet forms, built by induction on a sequence of graphs that
converges towards the considered domain. It is these Dirichlet forms that enable one to obtain energy
forms on this domain.

1
 Laplacians on fractal curves are not that simple to implement. One must of course bear in mind
that a fractal curve is topologically equivalent to a line segment. Thus, how can one make a distinction
between the spectral properties of a curve, and a line segment ? Dirichlet forms solely depend on
the topology of the domain, and not of its geometry. The solution is to consider energy forms more
sophisticated than classical ones, by means of normalization constants that could, not only bear the
topology, but, also, the geometric characteristics. One may refer to the works of U. Mosco [Mos02] for
the Sierpiński curve, and U. Freiberg [FL04], where the authors build an energy form on non-self sim-
ilar closed fractal curves, by integrating the Lagrangian on this curve. It is not the case of all existing
works: in [UD14], the authors just use topological normalization constants for the energy forms in stake.

In the sequel, we give an explicit construction of a Laplacian on the Minkowski curve, with energy
forms that bear the geometric characteristic of the structure. The spectrum of the Laplacian is obtained
by means of spectral decimation, in the spirit of the works of M. Fukushima and T. Shima [FS92].
On doing so, we choose three different methods. This enable us to initiate a detailed study of the
spectrum.

Figure 1 – Minkowski sausage.

Figure 2 – The initial line segment.

2
 0
1

Figure 3 – The graph obtained at the first iteration.

2 Framework of the study

In the sequel, we place ourselves in the Euclidean plane of dimension 2, referred to a direct orthonormal
frame. The usual Cartesian coordinates are (x, y).

Notation. Let us denote by P0 and P1 the points:

P0 = (0, 0) , P1 = (1, 0)

Notation. Let us denote by θ ∈ ]0, 2 π[, k > 0, T1 , and T2 real numbers. Let us consider the rotation
matrix:  
cos θ − sin θ
RO,θ =
sin θ cos θ
We introduce the iterated function system of the family of maps from R2 to R2 :

{f1 , ..., f8 }

where, for any integer i belonging to {1, ..., 8}, and any X ∈ R2 :
 
T1
fi (X) = k RO,θ X +
T2

Remark 2.1. If 0 < k < 1, the family {f1 , ..., f8 } is a family of contractions from R2 to R2 , the ratio of
which is k.

Property 2.1. According to [Hut81], there exists a unique subset MC ⊂ R2 such that:
8
[
MC = fi (MC)
i=1

which will be called the Minkowski Curve.

One may note that this curve is not strictly self-similar in the sense defined by B. Mandelbrot [Man77], [Hut81]:

∃ (i, j) ∈ {f1 , ..., f8 }2 , i 6= j : fi (MC) ∩ fj (MC) 6= ∅

Definition 2.1. We will denote by V0 the ordered set, of the points:

{P0 , P1 }

3
The set of points V0 , where, for any i of {0, 1}, the point P0 is linked to the point P1 , constitutes an
oriented graph, that we will denote by MC0 . V0 is called the set of vertices of the graph MC0 .

For any strictly positive integer m, we set:

8
[
Vm = fi (Vm−1 )
i=1

The set of points Vm , where the points of an m-cell are linked in the same way as MC0 , is an oriented
graph, which we will denote by MCm . Vm is called the set of vertices of the graph MCm . We will
denote, in the following, by Nm the number of vertices of the graph MCm .

Notation. We introduce the following similarities, from R2 to R2 , such that, for any X ∈ R2 :

        
1 0 1 1 1 1
f1 (X) = RO,0 X + , f2 (X) = RO, π2 X+ , f3 (X) = RO,0 X +
4 0 4 0 4 1

       
1 2 1 2 1 2
f4 (X) = RO, 3 π X+ , f5 (X) = RO, 3 π X+ , f6 (X) = RO,0 X +
4 2 1 4 2 0 4 −1
    
1 3 1 3
f7 (X) = RO, π2 X+ , f8 (X) = RO,0 X +
4 −1 4 0

Property 2.2. The set of vertices (Vm )m∈N is dense in MC.

Proposition 2.3. Given a natural integer m, we will denote by Nm the number of vertices of the graph
MCm . One has:
N0 = 2
and, for any strictly positive integer m:

Nm = 8 × Nm−1 − 7

Proof. The proposition results from the fact that, for any strictly positive integer m, each graph MCm
is the union of eight copies of MCm−1 : every copy shares one vertex with its predecessor.

Definition 2.2. Consecutive vertices of MC

Let us set: P2 = P0 . Two points X and Y of MC will be called consecutive vertices of MC if there
exists a natural integer m, and an integer j of {0, 1}, such that:

X = (fi1 ◦ . . . ◦ fim ) (Pj ) and Y = (fi1 ◦ . . . ◦ fim ) (Pj+1 ) (i1 , . . . , im ) ∈ {1, . . . , 8}m

4
Definition 2.3. Edge relation, on the graph MC

Given a natural integer m, two points X and Y of MCm will be called adjacent if and only if X
and Y are two consecutive vertices of MCm . We will write:

X∼Y
m

Definition 2.4. For any positive integer m, the Nm consecutive vertices of the graph MCm are, also,
the vertices of two triangles Tm,lef t , Tm,right , and 2 × 8m−1 squares Qm,j , 1 6 j 6 2 × 8m−1 . For any
integer j such that 1 6 j 6 2 × 8m−1 , one obtains each square by linking the point number j to the
point number j + 1 if j = i mod 9, 1 6 i 6 3, and the point number j to the point number j − 3
if j = 4 mod 9, and the triangles by linking the point number j to the point number j + 1 and j + 2 if
j = 0 mod 9. One has to consider those polygons as semi-closed ones, since, for any of those 4−gons,
the starting vertex, i.e. the point number j, is not connected, on the graph MCm , to the extreme one,
i.e. the point number j − 3. In the same way, the extreme vertices of the two triangles Tm,lef t , Tm,right ,
are not connected, on the graph MCm .
The reunion of those triangles and squares generate a Borel set of R2 .

1,left
1,2

1,1
1,right

Figure 4 – The triangles T1,lef t , T1,right , and the squares Q1,1 , Q1,2 .

Definition 2.5. Quasi-quadrilateral domain delimited by the graph MCm , m ∈ N

For any natural integer m, we will call quasi-quadrilateral domain delimited by MCm , and denote
by D (MCm ), the reunion of the triangles Tm,j , and the squares Qm,j .

Definition 2.6. Quasi-quadrilateral domain delimited by the Minkowski Curve MC

We will call quasi-quadrilateral domain delimited by MC, and denote by D (MC), the limit:
D (MC) = lim D (MCm )
n→+∞

which has to be understood in the following way: given a continuous function u on the curve MC, and
a measure with full support µ on R2 , then:
 
m−1
Z 1 2×8
X 1 X X 1 X 
u dµ = lim u (X) µ (Tm,lef t ) + u (X) µ (Qm,j ) + u (X) µ (Tm,right )
D(MC) m→+∞  3 4 3 
X vertex of Tm,lef t j=1 X vertex of Qm,j X vertex of Tm,right

5
Remark 2.2. One may wonder why consider such a quasi-quadrilateral domain, whereas we presently
deal with a curve ? If it seemed, to us, the most natural way to perform integration on our frac-
tal curve, there are, also, additionnal reasons. Part of an explanation can be found in the paper
of U. Mosco [Mos02], where the author suggests, in the case of the Sierpiński curve, to "fill each small
simplex (...) not only with its edges, but with the whole portion of the limit curve which it encom-
passes." It is also natural, in so far as it enables one to take into account the self-similar patterns that
appear thanks to the curve, patterns, the measure of which plays the part of a pound. This joins the
seminal work of J. Harrison et al. [HN91], [HN92], [Har93].

Definition 2.7. Word, on the Minkowski curve MC

Let m be a strictly positive integer. We will call number-letter any integer Wi of {1, . . . , 8}, and
word of length |W| = m, on the graph MC, any set of number-letters of the form:

W = (W1 , . . . , Wm )
We will write:

fW = fW1 ◦ . . . ◦ fWm

Proposition 2.4. Adresses, on the the Minkowski Curve

i. Every Pi , i ∈ {0, 1}, has exactly one neighbor.

ii. Given a strictly positive integer m, every X ∈ Vm \V0 has exactly two neighbors and two addresses

X = f(W1 ,...,Wm ) (P0 ) = f(W ′ 1 ,...,W ′ m ) (P1 )

where W = (W1 , . . . , Wm ) and W ′ = (W ′ 1 , . . . , W ′ m ) denote two appropriate words of length m,

on the graph MCm (see [Str06]).

Proposition 2.5. Let us set:

[
V⋆ = Vm
m∈N

The set V⋆ is dense in MC.

6
3 Energy forms, on the Minkowski Curve
3.1 Dirichlet forms
Definition 3.1. Dirichlet form, on a finite set (We refer to [Kig03])

Let V denote a finite set V , equipped with the usual inner product which, to any pair (u, v) of functions
defined on V , associates:
X
(u, v) = u(P ) v(P )
P ∈V
A Dirichlet form on V is a symmetric bilinear form E, such that:

1. For any real valued function u defined on V : E(u, u) > 0.

2. E(u, u) = 0 if and only if u is constant on V .
3. For any real-valued function u defined on V , if:
u⋆ = min (max(u, 0), 1)

i.e. :

if

1
 u(p) > 1
∀p ∈ V : u⋆ (p) = u(p) si 0 < u(p) < 1
0 if u(p) 6 0


then: E(u⋆ , u⋆ ) 6 E(u, u) (Markov property).

Let us now consider the problem of energy forms on our curve. Such a problem was studied
by U. Mosco [Mos02], who suggested to generalize Riemaniann models to fractals and relate the fractal
analogous of gradient forms, i.e. the Dirichlet forms, to a metric that could reflect the fractal properties
of the considered structure. The link is to be made by means of specific energy forms.

There are two major features that enable one to characterize fractal structures:
i. Their topology, i.e. their ramification.
ii. Their geometry.

The topology can be taken into account by means of classical energy forms (we refer to [Kig89],
[Kig93], [Str99], [Str06]).
As for the geometry, again, things are not that simple to handle. U. Mosco introduces a strictly positive
parameter, δ, which is supposed to reflect the way ramification - or the iterative process that gives
birth to the sequence of graphs that approximate the structure - affects the initial geometry of the
structure. For instance, if m is a natural integer, X and Y two points of the initial graph V1 , and M
a word of length m, the Euclidean distance dR2 (X, Y ) between X and Y is changed into the effective
distance:

(dR2 (X, Y ))δ

This parameter δ appears to be the one that can be obtained when building the effective resistance
metric of a fractal structure (see [Str06]), which is obtained by means of energy forms. To avoid turning
into circles, this means:

7
 i. either working, in a first time, with a value δ0 equal to one, and, then, adjusting it when building
the effective resistance metric ;

ii. using existing results, as done in [FL04].

One may note that in a very interesting and useful remark, U. Mosco puts the light on the relation
that exists between the walk dimension DW of a self-similar set, and δ:

DW = 2 δ
which will enable us to obtain the required value of the constant δ.

Definition 3.2. Walk dimension

Given a strictly positive integer N , the dimension related to a random walk on a self-similar set with
respect to N similarities, the ratio of which is equal to k ∈ ]0, 1[, is given by:

ln EA (τ )
DW = −
ln k
where EA (τ ) is the mean crossing time of a random walk starting from a vertex A of the self-similar
set.

Notation. In the sequel, we will denote by DW (MC) the Walk dimension of the Minkowski Curve.

Remark 3.1. Explicit computation of the Walk dimension of the Minkowski Curve

In the sequel, we follow the algorithm described in [FT13], which uses the theory of Markov chains
(we refer to [KS83]).

To this purpose, we define EX (τ ) to be the mean number of steps a simple random walk needs to reach
a vertex B ∈ V0 when starting at X. We consider the graph MI1 and denote by Xi for i ∈ {1, ..., 7}
the set of vertices of V1 \ V0 :

Figure 5 – The graph MI1

8
First, one has to introduce the adjacency matrix of the graph MI1 :
 
0 1 0 0 0 0 0 0 0

 1 0 1 0 0 0 0 0 0 


 0 1 0 1 0 0 0 0 0 


 0 0 1 0 1 0 0 0 0 

AMI1 =
 0 0 0 1 0 1 0 0 0 


 0 0 0 0 1 0 1 0 0 


 0 0 0 0 0 1 0 1 0 

 0 0 0 0 0 0 1 0 1 
0 0 0 0 0 0 0 1 0

One then build the related stochastic matrix:

 
0 1 0 0 0 0 0 0 0
1 1

2
0 2
0 0 0 0 0 0 
1 1
 
 0 2
0 2
0 0 0 0 0 
1 1
 
 0 0 2
0 2
0 0 0 0 
Astoch 1 1
 
MI1 =
 0 0 0 2
0 2
0 0 0 

1

 0 0 0 0 2
0 1 0 0 

1 1

 0 0 0 0 0 2
0 2
0 

1 1
 0 0 0 0 0 0 2
0 2

0 0 0 0 0 0 0 1 0

By suppressing the lines and columns associated to one may call "cemetary states" (i.e. the fixed
points of the contractions), one obtains the matrix:

0 1 0 0 0 0 0 0
 
1 1
 2 0 2 0 0 0 0 0 
1 1
 
 0 2 0 2 0 0 0 0 
1 1
 
 0 0 2 0 2 0 0 0 
MMI1 = 1 1


 0 0 0 2 0 2 0 0 

1

 0 0 0 0 2 0 1 0 

1 1
 0 0 0 0 0 2 0 2

1
0 0 0 0 0 0 2 0
Starting at P0 , one has:

EP0 (τ ) = EX1 (τ ) + 1 , EP1 (τ ) = 0

1 1 1
EX1 (τ ) = (EP0 + EX2 (τ ))+1 , EX2 (τ ) = (EX1 + EX3 (τ ))+1 , EX3 (τ ) = (EX2 + EX4 (τ ))+1
2 2 2

1 1 1
EX4 (τ ) = (EX3 + EX5 (τ ))+1 , EX5 (τ ) = (EX4 + EX6 (τ ))+1 , EX6 (τ ) = (EX5 + EX7 (τ ))+1
2 2 2
1
EX7 (τ ) =
(EP1 + EX6 (τ )) + 1
2
Let us introduce the vector of expected crossing-times:

EP0 (τ )
 

 EX1 (τ ) 


 EX2 (τ ) 

 EX3 (τ ) 
T= 

 EX4 (τ ) 


 EX5 (τ ) 

 EX6 (τ ) 
EX7 (τ )

9
 From the theory of Markov chains, we have :

T = (EMI1 − MMI1 )−1 1

where 1 := (1, ..., 1)T . By solving the above system, one gets:

64
 

 63 


 60 

 55 
T= 

 48 


 39 

 28 
15
which leads to:

EP1 (τ ) = 64
and:

ln 64
DW (MC) = =3
ln 4
one has thus:

DW (MC) 3
δ= =
2 2

Definition 3.3. Energy, on the graph MCm , m ∈ N, of a pair of functions

Let m be a natural integer, and u and v two real valued functions, defined on the set
n o
Vm = X1m , . . . , XN
m
S
m

of the Nm vertices of MCm .

We introduce the energy, on the graph MCm , of the pair of functions (u, v), as:
NXm −1

!
!
u (Xim ) − u Xi+1m v (Xim ) − v Xi+1 m
EMCm (u, v) =
i=1
dδR2 (X, Y ) dδR2 (X, Y )
NXm −1

42 m δ u (Xim ) − u Xi+1
m
v (Xim ) − v Xi+1 m





=
i=1

For the sake of simplicity, we will write it under the form:

X
EMCm (u, v) = 42 m δ (u(X) − u(Y )) (v(X) − v(Y ))
X∼Y
m

10
Proposition 3.1. Harmonic extension of a function, on the Minkowski Curve - Ramifica-
tion constant

For any integer m > 1, if u is a real-valued function defined on Vm−1 , its harmonic extension,
denoted by ũ, is obtained as the extension of u to Vm which minimizes the energy:
X
EMCm (ũ, ũ) = 42 m δ (ũ(X) − ũ(Y ))2
X∼Y
m

The link between EMCm and EMCm−1 is obtained through the introduction of two strictly positive con-
stants rm and rm−1 such that:
X X
rm 42 (m+1) δ (ũ(X) − ũ(Y ))2 = rm−1 42 (m−1) δ (u(X) − u(Y ))2
X ∼ Y X ∼ Y
m+1 m−1

For the sake of simplicity, we will fix the value of the initial constant: r0 = 1.
1
r=
r1
By induction, one gets:

rm = r1m = r −m
and:
X
Em (u) = r −m 42 m δ (ũ(X) − ũ(Y ))2
X∼Y
m

Since the determination of the harmonic extension of a function appears to be a local problem, on
the graph MCm−1 , which is linked to the graph MCm by a similar process as the one that links MC1
to MC0 , one deduces, for any integer m > 1:

EMCm (ũ, ũ) = r −m 42 m δ EMCm−1 (ũ, ũ)

If v is a real-valued function, defined on Vm−1 , of harmonic extension ṽ, we will write:

X X
Em (u, v) = r −m 42 m δ (ũ(X) − ũ(Y )) (ṽ(X) − ṽ(Y )) = r −m 43 m (ũ(X) − ũ(Y )) (ṽ(X) − ṽ(Y ))
X∼Y X∼Y
m m

The constant r −1 , which can be interpreted as a topological one, will be called ramification constant.
For further precision on the construction and existence of harmonic extensions, we refer to [Sab97].

Definition 3.4. Energy scaling factor

By definition, the energy scaling factor is the strictly positive constant ρ such that, for any inte-
ger m > 1, and any real-valued function u defined on Vm :


EMCm (u, u) = ρ EMCm u|Vm−1 , u|Vm−1

11
Proposition 3.2. The energy scaling factor ρ is linked to the topology and the geometry of the fractal
curve by means of the relation:

42 δ
=8ρ=
8
One may note that a more general calculation of the energy scaling factor can be found in the work
by R. Capitanelli [Cap02].

3.2 Explicit computation of the ramification constant

3.2.1 Direct method
Let us denote by u a real-valued, continuous function defined on V0 = {P0 , P1 }, and by ũ its harmonic
extension to V1 . Let us set: u(P0 ) = A, u(P1 ) = B. We recall that the energy on V0 is given by :

E0 (u) = (A − B)2

We will denote by U1 , U2 , U3 , U4 , U5 , U6 , U7 , U16 , the respective images, by ũ, of the vertices of V1 \V0 .

One has then:

E1 (ũ) = (A − U1 )2 + (U1 − U2 )2 + (U2 − U3 )2 + (U3 − U4 )2 + (U4 − U5 )2 + (U5 − U6 )2 + (U6 − U7 )2 + (U7 − B)2

The minimum of this quantity is such that:

U = A−1 b
where the matrix A is given by:

2 −1 0 0 0 0 0
 

 −1 2 −1 0 0 0 0 


 0 −1 2 −1 0 0 0 

A=
 0 0 −1 2 −1 0 0 


 0 0 0 −1 2 −1 0 

 0 0 0 0 −1 2 −1 
0 0 0 0 0 −1 2
the vectors U and b by:
U1 A
   
U2   0 
   
U3   0 
   
U4 
U= , b= 0
 
 
U5   0 
   
U6   0 
U7 B
One obtains:
 
7A B
8 + 8
3A B
4 + 4
 
 
5A 3B
8 + 8
 
 
A B
U= 2 + 2
 

3A 5B
8 + 8
 
 
A 3B
4 + 4
 
 
A 7B
8 + 8

12
By substituting these values in the expression of the energy expression, one obtains:

(A − B)2 1
E1 (ũ) = = E0 (u)
8 8
Thus:
1
r=
8

3.2.2 A second method, using Einstein’s relation

Definition 3.5. Hausdorff dimension of a self-similar set with respect to N ∈ N⋆ similarities

Given a strictly positive integer N , the Hausdorff dimension of a self-similar set with respect to N
similarities, the ratio of which is equal to k ∈ ]0, 1[, is given by:
ln N
DH = −
ln k

Notation. In the sequel, we will denote by

ln 8
DH (MC) = −
ln k
the Hausdorff dimension of the Minkowski Curve.

Definition 3.6. Spectral dimension of a self-similar set with respect to N ∈ N⋆ similarities

Given a strictly positive integer N , the spectral dimension of a self-similar set with respect to N
similarities, the energy scaling factor of which is equal to ρ, is given by:
2 ln N
DS =
ln(N × ρ)

Notation. In the sequel, we will denote by

6 ln 2
DS (MC) =
ln(8 × ρ)
the spectral dimension of the Minkowski Curve.

Theorem 3.3. Einstein relation

Given a strictly positive integer N , and a self-similar set with respect to N similarities, the ratio of
which is equal to k ∈ ]0, 1[, one has the so-called Einstein relation between the walk dimension, the
Hausdorff dimension, and the spectral dimension of the set:
DS DW
DH =
2

13
Property 3.4. Given a strictly positive integer N , and a self-similar set with respect to N similarities,
the ratio of which is equal to k ∈ ]0, 1[, the energy scaling factor ρ, which solely depends on the topology,
and, therefore, not of the value of the contraction ratio, is given by:
EA (T )
ρ=
N

Figure 6 – The graph MC0

3.3 Specific Dirichlet and energy forms, for the Minkowski Curve
Definition 3.7. Dirichlet form, for a pair of continuous functions defined on the Minkowski
Curve MC

We define the Dirichlet form E which, to any pair of real-valued, continuous functions (u, v) defined
on MC, associates, subject to its existence:

X
r −m 42 m δ u|Vm (X) − u|Vm (Y ) v|Vm (X) − v|Vm (Y )




E(u, v) = lim Em u|Vm , v|Vm = lim
m→+∞ m→+∞
X∼Y
m

Definition 3.8. Normalized energy, for a continuous function u, defined on MC

Taking into account that the sequence Em u|Vm m∈N is defined on

[
V⋆ = Vi
i∈N

one defines the normalized energy, for a continuous function u, defined on MC, by:


E(u) = lim Em u|Vm
m→+∞

14
Property 3.5. The Dirichlet form E which, to any pair of real-valued, continuous functions defined
on MC, associates:

X
r −m 42 m δ




E(u, v) = lim Em u|Vm , v|Vm = lim u|Vm (X) − u|Vm (Y ) v|Vm (X) − v|Vm (Y )
m→+∞ m→+∞
X∼Y
m

satisfies the self-similarity relation:

8
X
−1 2 δ
E(u, v) = r 4 E (u ◦ fi , v ◦ fi )
i=1

Proof.
8
X 8
X

E (u ◦ fi , v ◦ fi ) = lim Em u|Vm ◦ fi , v|Vm ◦ fi
m→+∞
i=1 i=1
X 8
X
r −m 42 m δ



= lim u|Vm (fi (X)) − u|Vm (fi (Y )) v|Vm (fi (X)) − v|Vm (fi (Y ))
m→+∞
X∼Y i=1
m
X 8
X
−m 2 m δ



= lim r 4 u|Vm+1 (X) − u|Vm+1 (Y ) v|Vm+1 (X) − v|Vm+1 (Y )
m→+∞
X ∼ Y i=1
m+1

lim r 4−2 δ Em+1 u|Vm+1 , v|Vm+1

=
m→+∞
= r 4−2 δ E(u, v)

Notation. We will denote by dom E the subspace of continuous functions defined on MC, such that:

E(u) < +∞

Notation. We will denote by dom0 E the subspace of continuous functions defined on MC, which take
the value 0 on V0 , such that:

E(u) < +∞

Proposition 3.6. The space domE, modulo the space of constant function on MC, is a Hilbert space.

15
4 Laplacian, on the Minkowski Curve
Definition 4.1. Self-similar measure, on the domain delimited by the Minkowski Curve

A measure µ on R2 will be said to be self-similar domain delimited by the Minkowski Curve, if there
exists a family of strictly positive pounds (µi )16i68 such that:
8
X 8
X
µ= µi µ ◦ fi−1 , µi = 1
i=1 i=1

For further precisions on self-similar measures, we refer to the works of J. E. Hutchinson (see [Hut81]).

Property 4.1. Building of a self-similar measure, for the Minkowski Curve

The Dirichlet forms mentioned in the above require a positive Radon measure with full support.
Let us set for any integer i belonging to {1, . . . , 8}:
1
µi = RDH (MC) =
8
This enables one to define a self-similar measure µ on MC as:
8
1 X
µ= µ ◦ fi
8
i=1

Definition 4.2. Laplacian of order m ∈ N⋆

For any strictly positive integer m, and any real-valued function u, defined on the set Vm of the vertices
of the graph MCm , we introduce the Laplacian of order m, ∆m (u), by:
X
∆m u(X) = (u(Y ) − u(X)) ∀ X ∈ V m \ V0
Y ∈Vm , Y ∼X
m

Definition 4.3. Harmonic function of order m ∈ N⋆

Let m be a strictly positive integer. A real-valued function u,defined on the set Vm of the vertices of
the graph MCm , will be said to be harmonic of order m if its Laplacian of order m is null:

∆m u(X) = 0 ∀ X ∈ Vm \ V0

Definition 4.4. Piecewise harmonic function of order m ∈ N⋆

Given a strictly positive integer m, a real valued function u, defined on the set of vertices of MC, is
said to be piecewise harmonic function of order m if, for any word W of length m, u ◦ fW is
harmonic of order m.

16
Definition 4.5. Existence domain of the Laplacian, for a continuous function on MC (see
[BD85])

We will denote by dom ∆ the existence domain of the Laplacian, on the graph MC, as the set of
functions u of dom E such that there exists a continuous function on MC, denoted ∆ u, that we will
call Laplacian of u, such that :
Z
E(u, v) = − v ∆u dµ for any v ∈ dom0 E
D(MC)

Definition 4.6. Harmonic function

A function u belonging to dom ∆ will be said to be harmonic if its Laplacian is equal to zero.

Notation. In the following, we will denote by H0 ⊂ dom ∆ the space of harmonic functions, i.e. the
space of functions u ∈ dom ∆ such that:

∆u = 0
Given a natural integer m, we will denote by S (H0 , Vm ) the space, of dimension 8m , of spline functions "
of level m", u, defined on MC, continuous, such that, for any word W of length m, u ◦ TW is harmonic,
i.e.:

∆m (u ◦ TW ) = 0

Property 4.2. For any natural integer m:

S (H0 , Vm ) ⊂ dom E

Property 4.3. Let m be a strictly positive integer, X ∈ m ∈ S (H , V )

/ V0 a vertex of the graph MC, and ψX 0 m
a spline function such that:

1 if X = Y
 
m δXY ∀ Y ∈ Vm
ψX (Y ) = , where δXY =
0 ∀ Y ∈ / Vm 0 else
Then, since X ∈ m ∈ dom E.
/ V0 : ψX 0
m leads to:
For any function u of dom E, such that its Laplacian exists, definition (4.5) applied to ψX

Z Z
m m −m m m
E(u, ψX ) = Em (u, ψX ) = −r ∆m u(X) = − ψX ∆u dµ ≈ −∆u(X) ψX dµ
D(MC) D(MC)

since ∆u is continuous on MC, and the support of the spline function ψXm is close to X:

Z Z
m m
ψX ∆u dµ ≈ −∆u(X) ψX dµ
D(MC) D(MC)

By passing through the limit when the integer m tends towards infinity, one gets:

17
 Z Z
m m
lim ψX ∆m u dµ = ∆u(X) lim ψX dµ
m→+∞ D(MC) m→+∞ D(MC)

i.e.:
Z !−1
−m 2 m δ m
∆u(X) = lim r 4 ψX dµ ∆m u(X)
m→+∞ D(MC)

Remark 4.1. As it is explained in [Str06], one has just to reason by analogy with the dimension 1, more
particulary, the unit interval I = [0, 1], of extremities X0 = (0, 0), and X1 = (1, 0). The functions ψX1
and ψX2 such that, for any Y of R2 :

ψX1 (Y ) = δX1 Y , ψX2 (Y ) = δX2 Y

are, in the most simple way, tent functions. For the standard measure, one gets values that do not
depend on X1 , or X2 (one could, also, choose to fix X1 and X2 in the interior of I) :

1
Z Z
ψX1 dµ = ψX2 dµ =
I I 2
(which corresponds to the surfaces of the two tent triangles.)

Figure 7 – The graphs of the spline functions ψX1 and ψX2 .

In our case, we have to build the pendant, we no longer reason on the unit interval, but on our trian-
gular or square cells.

Given a natural integer m, and a point X ∈ Vm , the spline function ψX m is supported by two m-cells.

It is such that, for every m-line cell fW (MC) the vertices of which are X, Y 6= X :
m
ψX + ψYm = 1

Thus:
1
Z
m
(ψX + ψYm ) dµ = µ(fW (MC)) =
fW (MC) 8m

18
By symmetry, all three summands have the same integral. This yields:

1
Z
m
ψX dµ =
fW (MC) 2 × 8m
Taking into account the contributions of the remaining m-square cells, one has:

1
Z
m
ψX dµ = m
MC 8
which leads to:
Z −1
m
ψX dµ = 8m
MC
Since:

r −m = 8m
this enables us to obtain the point-wise formula, for u ∈ dom ∆:

∀ X ∈ MC : ∆µ u(X) = lim 64m ∆m u(X)

m→+∞

Theorem 4.4. Let u be in dom ∆. Then, the sequence of functions (fm )m∈N such that, for any natural
integer m, and any X of V⋆ \ V0 :
Z !−1
−m 2 m δ m
fm (X) = r 4 ψX dµ ∆m u(X)
D(MC)

converges uniformly towards ∆ u, and, reciprocally, if the sequence of functions (fm )m∈N converges
uniformly towards a continuous function on V⋆ \ V0 , then:

u ∈ dom ∆

Proof. Let u be in dom ∆. Then:

Z
m
Z !−1 ∆ u ψX dµ
D(MC)
r −m 42 m δ m
ψX dµ ∆m u(X) = Z
D(MC) m
ψX dµ
D(MC)

Since u belongs to dom ∆, its Laplacian ∆ u exists, and is continuous on the graph MC. The uniform
convergence of the sequence (fm )m∈N follows.

Reciprocally, if the sequence of functions (fm )m∈N converges uniformly towards a continuous function
on V⋆ \ V0 , the, for any natural integer m, and any v belonging to dom0 E:

19
 X
r −m 42 m δ



Em (u, v) = u|Vm (X) − u|Vm (Y ) v|Vm (X) − v|Vm (Y )
2 , X∼Y
(X,Y ) ∈ Vm
m
X
r −m 42 m δ



= u|Vm (Y ) − u|Vm (X) v|Vm (Y ) − v|Vm (X)
2 , X∼Y
(X,Y ) ∈ Vm
m
X X
r −m 42 m δ


= − v|Vm (X) u|Vm (Y ) − u|Vm (X)
X ∈ Vm \V0 Y ∈ Vm , Y ∼X
m
X X
−m 2 m δ


− r 4 v|Vm (X) u|Vm (Y ) − u|Vm (X)
X ∈ V0 Y ∈ Vm , Y ∼X
m
X
= − r −m 42 m δ v(X) ∆m u(X)
X ∈ Vm \V0
Z ! Z !−1
X
= − v(X) m
ψX dµ r −m 42 m δ m
ψX dµ ∆m u(X)
X ∈ Vm \V0 D(MC) D(MC)

Let us note that any X of Vm \ V0 admits exactly two adjacent vertices which belong to Vm \ V0 , which
accounts for the fact that the sum
X X
r −m 42 m δ


v(X) u|Vm (Y ) − u|Vm (X)
X ∈ Vm \V0 Y ∈ Vm \V0 , Y ∼X
m

has the same number of terms as:

X
r −m 82 m δ u|Vm (Y ) − u|Vm (X) v|Vm (Y ) − v|Vm (X)

(X,Y ) ∈ (Vm \V0 )2 , X∼Y

m

For any natural integer m, we introduce the sequence of functions (fm )m∈N such that, for any X
of Vm \ V0 :
Z !−1
−m 2 m δ m
fm (X) = r 4 ψX dµ ∆m u(X)
D(MC)

The sequence (fm )m∈N converges uniformly towards ∆ u. Thus:

 
Z  X 
m
Em (u, v) = − v|Vm (X) ∆ u|Vm (X) ψX dµ
D(MC)  X ∈ Vm \V0


5 Normal derivatives
Let us go back to the case of a function u twice differentiable on I = [0, 1], that does not vanish in 0
and : Z 1 Z 1
(∆u) (x) v(x) dx = − u′ (x) v ′ (x) dx + u′ (1) v(1) − u′ (0) v(0)
0 0
The normal derivatives:

∂n u(1) = u′ (1) et ∂n u(0) = u′ (0)

appear in a natural way. This leads to:

20
 Z 1 Z 1 X
(∆u) (x) v(x) dx = − u′ (x) v ′ (x) dx + v ∂n u
0 0 ∂ [0,1]

One meets thus a particular case of the Gauss-Green formula, for an open set Ω of Rd , d ∈ N⋆ :
Z Z Z
∇ u ∇ v dµ = − (∆u) v dµ + v ∂n u dσ
Ω Ω ∂Ω
where µ is a measure on Ω, and where dσ denotes the elementary surface on ∂ Ω.

In order to obtain an equivalent formulation in the case of the graph MC, one should have, for a pair
of functions (u, v) continuous on MC such that u has a normal derivative:
Z X
E(u, v) = − (∆u) v dµ + v ∂n u
Ω V0

For any natural integer m :

Em (u, v) X
r −m 42 m δ u|Vm (Y ) − u|Vm (X) v|Vm (Y ) − v|Vm (X)



=
2 , X∼Y
(X,Y ) ∈ Vm
m
X X
r −m 82 m δ


= − v|Vm (X) u|Vm (Y ) − u|Vm (X)
X ∈ Vm \V0 Y ∈ Vm , Y ∼X
m
X X
−m 2 m δ


− r 4 v|Vm (X) u|Vm (Y ) − u|Vm (X)
X ∈ V0 Y ∈ Vm , Y ∼X
m
X
= − v|Vm (X) r −m 42 m δ ∆m u|Vm (X)
X ∈ Vm \V0
X X
r −m 82 m δ v|Vm (X) u|Vm (X) − u|Vm (Y )


+
X ∈ V0 Y ∈ Vm , Y ∼X
m

We thus come across an analogous formula of the Gauss-Green one, where the role of the normal
derivative is played by:
X X
r −m 42 m δ


u|Vm (X) − u|Vm (Y )
X ∈ V0 Y ∈ Vm , Y ∼X
m

Definition 5.1. For any X of V0 , and any continuous function u on MC, we will say that u admits a
normal derivative in X, denoted by ∂n u(X), if:
X
lim r −m 42 m δ


u|Vm (X) − u|Vm (Y ) < +∞
m→+∞
Y ∈ Vm , Y ∼X
m

We will set:
X
lim r −m 42 m δ


∂n u(X) = u|Vm (X) − u|Vm (Y ) < +∞
m→+∞
Y ∈ Vm , Y ∼X
m

21
Definition 5.2. For any natural integer m, any X of Vm , and any continuous function u on MC, we
will say that u admits a normal derivative in X, denoted by ∂n u(X), if:
X
lim r −k 42 k δ


u|Vk (X) − u|Vk (Y ) < +∞
k→+∞
Y ∈ Vk , Y ∼X
k

We will set:
X
∂n u(X) = lim r −k 42 k δ


u|Vk (X) − u|Vk (Y ) < +∞
k→+∞
Y ∈ Vk , Y ∼X
k

Remark 5.1. One can thus extend the definition of the normal derivative of u to MC.

Theorem 5.1. Let u be in dom ∆. The, for any X of MC, ∂n u(X) exists. Moreover, for any v
of dom E, et any natural integer m, the Gauss-Green formula writes:
Z X
E(u, v) = − (∆u) v dµ + v ∂n u
D(MC) V0

6 Spectrum of the Laplacian

6.1 Spectral decimation
In the following, let u be in dom ∆. We will apply the spectral decimation method developed
by R. S. Strichartz [Str06], in the spirit of the works of M. Fukushima et T. Shima [?]. In order to
determine the eigenvalues of the Laplacian ∆ u built in the above, we concentrate first on the eigen-
values (−λm )m∈N of the sequence of graph Laplacians (∆m u)m∈N , built on the discrete sequence of
graphs (MCm )m∈N . For any natural integer m, the restrictions of the eigenfunctions of the continu-
ous Laplacian ∆ u to the graph MCm are, also, eigenfunctions of the Laplacian ∆m , which leads to
recurrence relations between the eigenvalues of order m and m + 1.

We thus aim at determining the solutions of the eigenvalue equation:

−∆ u = λ u on MC
as limits, when the integer m tends towards infinity, of the solutions of:

−∆m u = λm u on Vm \ V0
We will call them Dirichlet eigenvalues (resp. Neumann eigenvalues) if:

resp. ∂n u|∂MC = 0


u|∂MC = 0

Given a strictly positive integer m, let us consider a (m − 1)−cell, with boundary vertices X0 , X1 .
We denote by Y1 , Y2 , Y3 , Y4 , Y5 , Y6 , Y7 the points of Vm \ Vm−1 (see Fig. ):

22
 Figure 8 – The m-cell

The discrete equation on MC leads to the following system:



 (2 − λm ) u(Y1 ) = u(X0 ) + u(Y2 )
 (2 − λm ) u(Y2 ) = u(Y1 ) + u(Y3 )



 (2 − λm ) u(Y3 ) = u(Y2 ) + u(Y4 )


(2 − λm ) u(Y4 ) = u(Y3 ) + u(Y5 )
(2 − λm ) u(Y5 ) = u(Y4 ) + u(y6 )




(2 − λm ) u(Y6 ) = u(Y5 ) + u(Y7 )




(2 − λm ) u(Y7 ) = u(Y6 ) + u(X1 )


Under matrix form, one has:

Am u = u0
where:

2 − λm −1 0 0 0 0 0
 
 −1 2 − λm −1 0 0 0 0 
 
 0
 −1 2 − λm −1 0 0 0  
Am =
 0 0 −1 2 − λm −1 0 0 

 0
 0 0 −1 2 − λm −1 0  
 0 0 0 0 −1 2 − λm −1 
0 0 0 0 0 −1 2 − λm

u(Y1 ) u(X0 )
   
u(Y2 )  0 
   
u(Y3 )  0 
   
u= u(Y4 ) quad, u0 =  0 
  
u(Y5 )  0 
   
u(Y6 )  0 
u(Y7 ) u(X1 )

23
6.1.1 Direct method
n √ p √ o
By assuming λm 6= 2, 2 + ε 2, 2 + ε 2 + ε 2 , ε ∈ {−1, 1} for which the matrix Am is not in-
vertible, we can solve the system using Gauss algorithm applied on the augmented matrix Ãm :
2 − λm −1 0 0 0 0 0 u(X0 )
 
 −1 2 − λm −1 0 0 0 0 0 
 
 0
 −1 2 − λm −1 0 0 0 0  
Ãm = 
 0 0 −1 2 − λm −1 0 0 0  .
 0
 0 0 −1 2 − λ m −1 0 0 

 0 0 0 0 −1 2 − λm −1 0 
0 0 0 0 0 −1 2 − λm u(X1 )
Thus:

 
− u(X0 )(λm(λ−4)λm ((λm −4)λm ((λm −4)λm +7)+14)+7u(X0 )+u(X1 )
2
m −2)((λm −4)λm +2)((λm −4)λm (λm −2) +2)
u(Y1 )
  
u(X0 )(λm −4)λm (λm −2)2 +3u(X0 )+u(X1 )

 
u(Y2 )   ((λm −4)λm +2)((λm −4)λm (λm −2)2 +2) 
(λm −4)λm (a(λm −4)λm +5u(X0 )+u(X1 ))+5u(X0 )+3u(X1 ) 
  
u(Y3 )  − (λm −2)((λm −4)λm +2)((λm −4)λm (λm −2)2 +2)


 
u(Y4 ) =  u(X0 )+u(X1 )
.

   (λm −4)λm (λm −2)2 +2
u(Y5 ) 
− (λm −4)λ(λmm(u(X 0 )+u(X1 )(λm −4)λm +5u(X1 ))+3u(X0 )+5u(X1 )
 

−2)((λm −4)λm +2)((λm −4)λm (λm −2)2 +2)
 
u(Y6 )  
 u(X0 )+u(X1 )(λm −4)λm (λm −2)2 +3u(X1 ) 
u(Y7 )
 
 ((λm −4)λm +2)((λm −4)λm (λm −2)2 +2) 
u(X0 )+u(X1 )(λm −4)λm ((λm −4)λm ((λm −4)λm +7)+14)+7u(X1 )
− (λm −2)((λm −4)λm +2)((λm −4)λm (λm −2)2 +2)

Let us now compare the λm−1 −eigenvalues on Vm−1 , and the λm −eigenvalues on Vm . To this purpose,
we fix X0 ∈ Vm−1 \ V0 .
One has to bear in mind that X0 also belongs to a (m − 1)−cell, with boundary points X0 , X1′ and
interior points Z1 , Z2 , Z3 , Z4 , Z5 , Z6 , Z7 .
Thus:

(2 − λm−1 ) u(X0 ) = u(X1 ) + u(X1′ ) , (2 − λm ) u(X0 ) = u(Y1 ) + u(Z1 )

and:
2u(X0 )((λm − 3)(λm − 2)λm − 1)(λm ((λm − 7)λm + 14) − 7) + u(X1 ) + u(X1′ )
 
u(Y1 ) + u(Z1 ) = −
(λm − 2)((λm − 4)λm + 2) ((λm − 4)λm (λm − 2)2 + 2)

Then :

2u(X0 )((λm − 3)(λm − 2)λm − 1)(λm ((λm − 7)λm + 14) − 7) + u(X1 ) + u(X1′ )
 
(2 − λm ) u(X0 ) = −
(λm − 2)((λm − 4)λm + 2) ((λm − 4)λm (λm − 2)2 + 2)
 
2((λm − 3)(λm − 2)λm − 1)(λm ((λm − 7)λm + 14) − 7) + (2 − λm−1 )
= u(X0 ) −
(λm − 2)((λm − 4)λm + 2) ((λm − 4)λm (λm − 2)2 + 2)

Finally:

λm−1 = −(λm − 4)(λm − 2)2 λm ((λm − 4)λm + 2)2

One may solve:
r q p
λm = 2 + ε1 2 + ε2 2 + ε3 4 − λm−1 , ε1 , ε2 , ε2 , ∈ {−1, 1}

24
Let us introduce:
Z !−1
−m 2 m δ m
λ = lim r 4 ψX dµ λm = lim 64m λm
m→∞ D(MC) m→∞

One may note that the limit exists, since, when x is close to 0:
r
√ x
q
+ O x2


2− 2+ 2+ 4−x=
64

6.1.2 Recursive method

It is interesting to apply the method outlined by the second author ; the sequence (u (Yk ))06i68 satisfies
a second order recurrence relation, the characteristic equation of which is:

r 2 + {λm − 2} r + 1 = 0
The discriminant is:

δm = {λm − 2}2 − 4 = ωm
2
, ωm ∈ C
The roots r1,m and r2,m of the characteristic equation are the scalar given by:

2 − λm − ω m 2 − λm + ω m
r1,m = , r2,m =
2 2
One has then, for any natural integer k of {0, . . . , 8} :

k k
u (Yk ) = αm r1,m + βm r2,m
where αm and βm denote scalar constants.
In the same way, X0 and X1 belong to a sequence that satisfies a second order recurrence relation, the
characteristic equation of which is:

{Λm−1 − 2} r = −1 − r 2
and discriminant:

δm−1 = {Λm−1 − 2}2 − 4 = ωm−1

2
, ωm−1 ∈ C
The roots r1,m−1 and r2,m−1 of this characteristic equation are the scalar given by:

2 − Λm−1 − ωm−1 2 − Λm−1 + ωm−1

r1,m−1 = , r2,m−1 =
2 2
From this point, the compatibility conditions, imposed by spectral decimation, have to be satisfied:

u (Y0 ) = u (X0 )
u (Y8 ) = u (X1 )
i.e.:

αm + βm = αm−1 + βm−1 Cm
8 8
αm r1,m + βm r2,m = αm−1 r1,m−1 + βm−1 r2,m−1 C2,m
where, for any natural integer m, αm and βm are scalar constants (real or complex).

Since the graph MI m−1 is linked to the graph MI m by a similar process to the one that links MI 1
to MI 0 , one can legitimately consider that the constants αm and βm do not depend on the integer m:

25
 ∀ m ∈ N⋆ : αm = α ∈ R , βm = β ∈ R
The above system writes:

8 8
α r1,m + β r2,m = α r1,m−1 + β r2,m−1
and is satisfied for:
8

r1,m = r1,m−1
8
r2,m = r2,m−1
i.e.:

2 − λm − ω m 8
  
2 − λm−1 − ωm−1
=


2 2

2 − λm + ω m 8
 
 2 − λm−1 + ωm−1

 =
2 2
This lead to the recurrence relation:

∀ m N, n > 2 : λm−1 = −(λm − 4)(λm − 2)2 λm (2 + (λm − 4)λm )2

which is the same one as in the above.

6.2 A spectral means of determination of the ramification constant

In the sequel, we present an alternative method, that enables one to compute the ramification constant
using spectral decimation (we refer to Zhou [Zho07] for further details). Given a strictly positive
integer m, let us denote by Hm the Laplacian matrix associated to the graph MCm , and by L the set
of linear real valued functions defined on MCm . One may write:
T
 
Tm Jm
Hm =
Jm Xm
where Tm ∈ L(V0 ), Jm ∈ L(V0 , Vm \ V0 ) and Xm ∈ L(Vm \ V0 ).

Let D be the Laplacian matrix on MC0 , and M a diagonal matrix with Mii = −Xii .

Definition 6.1. The Laplacian is said to have a strong harmonic structure if there exist rational
functions of the real variable λ, respectively denoted by KD and KT such that, for for any real
number λ satisfying det (X + λ M ) 6= 0:
T + J T (X + λM )−1 J = KD (λ) D + KT (λ) T
Let us set:
F = {λ ∈ R : KD (0) = 0 or det(X + λM ) = 0}
and:
Fk = λ ∈ F λ is an eigenvalue of the normalized laplacian


The elements of F will be called forbidden eigenvalues. As for the elements of Fk , they will be
called forbidden eigenvalues at the kth step.
The map R such that, for any real number λ:
λ − KT (λ)
R(λ) =
KD (λ)
will be called spectral decimation function.

26
Remark 6.1. In the case of the Minkowski Curve, one may check that:

#(Fi (V0 ) ∩ V0 ) 6 1 ∀ i ∈ {1, 2, 3, 4, 5, 6, 7, 8}

   
−1 1 −1 0
D= , T =
1 −1 0 −1
−2 1 0 0 0 0 0
 
 1 −2 1 0 0 0 0 
 
 0
 1 −2 1 0 0 0 

X= 0
 0 1 −2 1 0 0 
 0
 0 0 1 −2 1 0 
 0 0 0 0 1 −2 1 
0 0 0 0 0 1 −2
2 0 0 0 0 0 0 1 0
   
 0 2 0 0 0 0 0   0 0 
   
 0 0 2 0 0 0 0   0 0 
   
M =
 0 0 0 2 0 0 0  , J = 0 0
  

 0 0 0 0 2 0 0   0 0 
   
 0 0 0 0 0 2 0   0 0 
0 0 0 0 0 0 2 0 1

(2λ−1) (4λ2 −6λ+1)(8(λ−2)(λ−1)λ(2λ−3)+1)

 
1
− 8(λ−1)(2(λ−2)λ+1)(8(λ−2)λ(λ−1)2 +1)
− 8(λ−1)(2(λ−2)λ+1)(8(λ−2)λ(λ−1) 2 +1)
T +J T (X+λM )−1 J =  (2λ−1)(4λ2 −6λ+1)(8(λ−2)(λ−1)λ(2λ−3)+1)

1
− 8(λ−1)(2(λ−2)λ+1)(8(λ−2)λ(λ−1)2 +1) − 8(λ−1)(2(λ−2)λ+1)(8(λ−2)λ(λ−1)2 +1)

Thus, the Minkowsky Curve has a strong harmonic structure, and:

1 λ(4λ(λ(2λ − 7) + 7) − 7)
T + J T (X + λM )−1 J = − D+ T
8(λ − 1)(2(λ − 2)λ + 1) (8(λ − 2)λ(λ − 1)2 + 1) 8(λ − 2)λ(λ − 1)2 + 1
Then :
1
=8
KD (0)
And we can verify again the spectral decimation function is :

R(λ) = −(λ − 4)(λ − 2)2 λ((λ − 4)λ + 2)2

6.3 A detailed study of the spectrum

6.3.1 First case: m = 1.
The Minkowski graph, with its eleven vertices, can be seen in the following figures:

27
 Figure 9 – The line after the first iteration.
Let us look for the kernel of the matrix A1 in the case of forbidden eigenvalues i.e.
√ √
 q 
λ1 ∈ 2, 2 + ε 2, 2 + ε 2 + ε 2, , ε ∈ {−1, 1}

For λ1 = 2, we find the one dimensional Dirichlet eigenspace:

V21 = Vect {(−1, 0, 1, 0, −1, 0, 1)}

√
For λ1 = 2 + ε 2, ε ∈ {−1, 1}, we find the one-dimensional Dirichlet eigenspace:

1
n √ √ o
V2+ε∈{−1,1}√ = Vect (−1, ε 2, −1, 0, 1, −ε 2, 1)
2
p √
For λ1 = 2 + ε ∈ {−1, 1} 2 + 2, ε ∈ {−1, 1}, we find the one-dimensional Dirichlet eigenspace:
n √ √ q √
√ √ o
V1
p p
√ √ = Vect 1, −ε 2+ 2, 1 + 2, −ε 2 2+ 2 , 1 + 2, −ε 2 + 2, 1
2+ε ∈ {−1,1} 2+ 2
p √
For λ1 = 2 + ε 2− 2, ε ∈ {−1, 1}, we find the one-dimensional Dirichlet eigenspace
n √ √ q √
√ √ o
V1
p p
√ √ = Vect 1, −ε 2 − 2, 1 − 2, ε 2 2 − 2 , 1 − 2, −ε 2 − 2, 1
2+ε∈{−1,1} 2− 2

One may easily check that:

#(V1 \ V0 ) = 7
Thus, the spectrum is complete.

6.3.2 Second case: m = 2

Let us now move to the m = 2 case.

Figure 10 – The cell F1 (V1 )

Let us denote by Zji := fi (Xj ) the points of V2 \ V1 that belongs to the cell fi (V0 ).

One has to solve the following systems, taking into account the Dirichlet boundary conditions (u(P0 ) = u(P1 ) = 0):

u(Z71 ) + u(Z12 ) (2 − λm ) u(Z11 ) u(X0 ) + u(Z21 )

 
 (2 − λm ) u(X1 ) =  =
u(Z72 ) + u(Z13 ) (2 − λm ) u(Z21 ) u(Z11 ) + u(Z31 )
 


 (2 − λm ) u(X2 ) = 

 =
u(Z73 ) + u(Z14 )  (2 − λm ) u(Z31 ) u(Z21 ) + u(Z41 )
 
 (2 − λm ) u(X3 ) = =

 

(2 − λm ) u(X4 ) = u(Z74 ) + u(Z15 ) , (2 − λm ) u(Z41 ) = u(Z31 ) + u(Z51 )
(2 − λm ) u(X5 ) = u(Z75 ) + u(Z16 ) (2 − λm ) u(Z51 ) = u(Z41 ) + u(Z61 )

 


 

 (2 − λm ) u(X6 ) = u(Z76 ) + u(Z17 )  (2 − λm ) u(Z61 ) = u(Z51 ) + u(Z71 )

 

 
(2 − λm ) u(X7 ) = u(Z77 ) + u(Z18 ) (2 − λm ) u(Z61 ) = u(Z61 ) + u(X1 )
 

28
 (2 − λm ) u(Z12 ) u(X1 ) + u(Z22 ) (2 − λm ) u(Z13 ) u(X2 ) + u(Z23 )
 
 =  =
(2 − λm ) u(Z22 ) u(Z12 ) + u(Z32 ) (2 − λm ) u(Z23 ) u(Z13 ) + u(Z33 )
 


 = 

 =
 (2 − λm ) u(Z32 ) u(Z22 ) + u(Z42 )  (2 − λm ) u(Z33 ) u(Z23 ) + u(Z43 )
 
= =

 

(2 − λm ) u(Z42 ) = u(Z32 ) + u(Z52 ) , (2 − λm ) u(Z43 ) = u(Z33 ) + u(Z53 )
(2 − λm ) u(Z52 ) = u(Z42 ) + u(Z62 ) (2 − λm ) u(Z53 ) = u(Z43 ) + u(Z63 )

 


 

 (2 − λm ) u(Z62 ) = u(Z52 ) + u(Z72 )  (2 − λm ) u(Z63 ) = u(Z53 ) + u(Z73 )

 

 
(2 − λm ) u(Z72 ) = u(Z62 ) + u(X2 ) (2 − λm ) u(Z73 ) = u(Z63 ) + u(X3 )
 

(2 − λm ) u(Z14 ) u(X3 ) + u(Z24 ) (2 − λm ) u(Z15 ) u(X4 ) + u(Z25 )

 
 =  =
(2 − λm ) u(Z24 ) u(Z14 ) + u(Z34 ) (2 − λm ) u(Z25 ) u(Z15 ) + u(Z35 )
 


 = 

 =
 (2 − λm ) u(Z34 ) u(Z24 ) + u(Z44 )  (2 − λm ) u(Z35 ) u(Z25 ) + u(Z45 )
 
= =

 

(2 − λm ) u(Z44 ) = u(Z34 ) + u(Z54 ) , (2 − λm ) u(Z45 ) = u(Z35 ) + u(Z55 )
(2 − λm ) u(Z54 ) = u(Z44 ) + u(Z64 ) (2 − λm ) u(Z55 ) = u(Z45 ) + u(Z65 )

 


 

(2 − λm ) u(Z64 ) = u(Z54 ) + u(Z74 ) (2 − λm ) u(Z65 ) = u(Z55 ) + u(Z75 )

 


 

(2 − λm ) u(Z74 ) = u(Z64 ) + u(X4 ) (2 − λm ) u(Z75 ) = u(Z65 ) + u(X5 )
 

(2 − λm ) u(Z16 ) u(X5 ) + u(Z26 ) (2 − λm ) u(Z17 ) u(X6 ) + u(Z27 )

 
 =  =
(2 − λm ) u(Z26 ) u(Z16 ) + u(Z36 ) (2 − λm ) u(Z27 ) u(Z17 ) + u(Z37 )
 


 = 

 =
 (2 − λm ) u(Z36 ) u(Z26 ) + u(Z46 )  (2 − λm ) u(Z37 ) u(Z27 ) + u(Z47 )
 
= =

 

(2 − λm ) u(Z46 ) = u(Z36 ) + u(Z56 ) , (2 − λm ) u(Z47 ) = u(Z37 ) + u(Z57 )
 (2 − λm ) u(Z56 ) = u(Z46 ) + u(Z66 )  (2 − λm ) u(Z57 ) = u(Z47 ) + u(Z67 )

 

 
(2 − λm ) u(Z66 ) = u(Z56 ) + u(Z76 ) (2 − λm ) u(Z67 ) = u(Z57 ) + u(Z77 )

 


 

(2 − λm ) u(Z76 ) = u(Z66 ) + u(X6 ) (2 − λm ) u(Z77 ) = u(Z67 ) + u(X7 )
 

(2 − λm ) u(Z18 ) u(X7 ) + u(Z28 )


 =
(2 − λm ) u(Z28 ) u(Z18 ) + u(Z38 )



 =
 (2 − λm ) u(Z38 ) u(Z28 ) + u(Z48 )

=


(2 − λm ) u(Z48 ) = u(Z38 ) + u(Z58 )
 (2 − λm ) u(Z58 ) = u(Z48 ) + u(Z68 )



8 u(Z58 ) + u(Z78 )
 (2 − λm ) u(Z6 ) =



 8
(2 − λm ) u(Z7 ) = u(Z68 ) + u(P1 )
The system can be written as: A2 x = 0 and we look for the kernel of A2 for forbidden eigenvalues.

For λ2 = 2, the eigenspace is one dimensional, generated by the vector:

(0, 0, 0, 0, 0, 0, 0, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0,
− 1, 0, 1, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0, −1, 0, 1, −1, 0, 1, 0, −1, 0, 1)
√
For λ2 = 2 + ε 2, ε ∈ {−1, 1}, the eigenspace is one dimensional, and is generated by:

√ √ √ √ √ √ √
(0, 0, 0, 0, 0, 0, 0, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1,
√ √ √ √ √ √ √ √ √
− ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2, 1, −1, ε 2, −1, 0, 1, −ε 2,
p √
For λ2 = 2 + ε 2 + 2, ε ∈ {−1, 1}, the eigenspace has dimension eight, and is generated by:

29
 √ √ √  √ √ √ √
q r  q q
(0, 0, 0, 0, 0, 0, 0, −1, ε 2 + 2, −1 − 2, ε 2 2 + 2 , −1 − 2, 2 + 2, −1, 1, −ε 2 + 2, 1 + 2,
√  √ √ √ √ √  √ √
r  q q r  q
− ε 2 2 + 2 , 1 + 2, − 2 + 2, 1, −1, ε 2 + 2, −1 − 2, ε 2 2 + 2 , −1 − 2, ε 2 + 2,
√ √ √  √ √ √ √ √ 
q r  q q r 
− 1, 1, −ε 2 + 2, 1 + 2, −ε 2 2 + 2 , 1 + 2, −ε 2 + 2, 1, −1, ε 2 + 2, −1 − 2, ε 2 2 + 2 ,
√ √ √ √ √  √ √ √
q q r  q q
− 1 − 2, ε 2 + 2, −1, 1, −ε 2 + 2, 1 + 2, −ε 2 2 + 2 , 1 + 2, −ε 2 + 2, 1, −1, ε 2 + 2,
√ √  √ √ √ √ √  √
r  q q r 
− 1 − 2, ε 2 2 + 2 , −1 − 2, ε 2 + 2, −1, 1, −ε 2 + 2, 1 + 2, −ε 2 2 + 2 , 1 + 2,
√
q
− ε 2 + 2, 1)

p √
For λ2 = 2 + ε 2 − 2, ε ∈ {−1, 1}, the eigenspace has dimension eight, and is generated by:

√ √ √  √ √ √ √
q r  q q
(0, 0, 0, 0, 0, 0, 0, −1, ε 2 − 2, 2 − 1, −ε 2 2 − 2 , 2 − 1, ε 2 − 2, −1, 1, −ε 2 − 2, 1 − 2,
√  √ √ √ √ √  √ √
r  q q r  q
ε 2 2 − 2 , 1 − 2, −ε 2 − 2, 1, −1, ε 2 − 2, 2 − 1, −ε 2 2 − 2 , 2 − 1, ε 2 − 2, −1, 1,
√ √ √  √ √ √ √ √  √
q r  q q r 
− ε 2 − 2, 1 − 2, ε 2 2 − 2 , 1 − 2, −ε 2 − 2, 1, −1, ε 2 − 2, 2 − 1, −ε 2 2 − 2 , 2 − 1,
√ √ √ √  √ √ √ √
q q r  q q
ε 2 − 2, −1, 1, −ε 2 − 2, 1 − 2, ε 2 2 − 2 , 1 − 2, −ε 2 − 2, 1, −1, ε 2 − 2, 2 − 1,
√  √ √ √ √ √  √ √
r  q q r  q
− ε 2 2 − 2 , 2 − 1, ε 2 − 2, −1, 1, −ε 2 − 2, 1 − 2, ε 2 2 − 2 , 1 − 2, −ε 2 − 2, 1)

From every forbidden eigenvalue λ1 , the spectral decimation leads to eight eigenvalues. Each of these
eigenvalues has multiplicity 1.

One may easily check that:

#V2 \ V0 = 63 = 7 × 1 + 8 × 7
Thus, the spectrum is complete.

6.3.3 General case

Let us now go back to the general case. Given a strictly positive integer m, let us introduce the
respective multiplicities Mm (λm ) of the eigenvalue λm .

One can easily check by induction that:

#Vm \ V0 = 8m − 1

Theorem 6.1. To every forbidden eigenvalue λ is associated an eigenspace, the dimension of which
is one: Mm (λ) = 1.

30
Proof. The theorem is true for m = 1 and m = 2. Recursively, we suppose the result true for m − 1.
At m There are :
m−1
X
8i × 7 = 8m − 8 = #Vm \ V0 − 7
i=1

eigenvalues generated using the continuous formula of spectral decimation. The remaining eigenspace
has dimension 7.
We can verify that every forbidden eigenvalue λ is an eigenvalue for m. If we consider the eigenfunction
null everywhere except the points of Vm \ Vm−1 where it takes the values of Vλ1 in the interior of every
m-cell.
We conclude that the eigenspace of every forbidden eigenvalue is one dimensional.

Figure 11 – The graph of the first eigenfunction

Figure 12 – The graph of the second eigenfunction

31
 Figure 13 – The graph of the third eigenfunction

Figure 14 – The graph of the fourth eigenfunction

7 Metric - Towards spectral asymptotics

Definition 7.1. Effective resistance metric, on MC
Given two points (X, Y ) of MC2 , let us introduce the effective resistance metric between X
and Y :
 −1
RMC (X, Y ) = min E(u)
{u | u(X)=0,u(Y )=1}

In an equivalent way, RMC (X, Y ) can be defined as the minimum value of the real numbers R such
that, for any function u of dom ∆:

|u(X) − u(Y )|2 6 R E(u)

Definition 7.2. Metric, on the Minkowski Curve MC

Let us define, on the Minkowski Curve MC, the distance dMC such that, for any pair of points (X, Y )
of MC2 :

32
 Figure 15 – The graph of the fifth eigenfunction

Figure 16 – The graph of the sixth eigenfunction

 −1
dMC (X, Y ) = min E(u, u)
{u | u(X)=0,u(Y )=1}

Remark 7.1. One may note that the minimum

min E(u)
{u | u(X)=0,u(Y )=1}

is reached for u being harmonic on the complement set, on MC, of the set

{X} ∪ {Y }

(One might bear in mind that, due to its definition, a harmonic function u on MC minimizes the
sequence of energies (EMCm (u, u))m∈N .

33
 Figure 17 – The graph of the seventh eigenfunction

Definition 7.3. Dimension of the Minkowski Curve MC, in the resistance metrics

The dimension of the Minkowski Curve MC, in the resistance metrics, is the strictly positive num-
ber dMC such that, given a strictly positive real number r, and a point X ∈ MC, for the X−centered
ball of radius r, denoted by Br (X):

µ (Br (X)) = r dMC

Property 7.1. Given a natural integer m, and two points (X, Y ) of MC2 such that X ∼ Y :
m
1
min E(u) . r m 4−2 m δ = m
{u | u(X)=0,u(Y )=1} 8
1
Let us denote by µ the standard measure on MC which assigns measure 2m to each quadrilateral m−
4
cell. Let us now look for a real number dMC such that:
 m dMC
1 1
= 2m
8 4
One obtains:
2
dMC =
3
Given a strictly positive real number r, and a point X ∈ MC, one has then the following estimate, for
the X−centered ball of radius r, denoted by Br (X):

µ (Br (X)) = r dMC

Definition 7.4. Eigenvalue counting function

We introduce the eigenvalue counting function N MC such that, for any real number x:

N MC (x) = # {λ eigenvalue of −∆ : λ 6 x}

34
Property 7.2. According to J. Kigami [Kig98], one has the modified Weyl formula:
   
ln x
MI
N (x) = G + o(1) xDS (MI)
2

8 From the Minkowski Curve, to the Minskowski Island

By connecting four Minkowski Curves, as it can be seen on the following figure:

Figure 18 – The Minskowski Island.

one obtains what is called "the Minskowski Island".

We expose, in the sequel, our results (we refer to Zhou [Zho07] for further details).
In the case of the Minkowski Curve, one may check that:

#(Fi (V0 ) ∩ V0 ) 6 1 ∀ i ∈ {1, 2, 3, 4, 5, 6, 7, 8}

   
−2 1 0 1 −2 0 0 0
 1 −2 1 0   0 −2 0 0 
D=  , T = 
 0 1 −2 1   0 0 −2 0 
1 0 1 −2 0 0 0 −2
Let us set:
−2 1 0 0 0 0 0
 

 1 −2 1 0 0 0 0 


 0 1 −2 1 0 0 0 

B=
 0 0 1 −2 1 0 0 


 0 0 0 1 −2 1 0 

 0 0 0 0 1 −2 1 
0 0 0 0 0 1 −2
and:

35
 2 0 0 0 0 0 0
 

 0 2 0 0 0 0 0 


 0 0 2 0 0 0 0 

C=
 0 0 0 2 0 0 0 


 0 0 0 0 2 0 0 

 0 0 0 0 0 2 0 
0 0 0 0 0 0 2
Then:
 
B 0 0 0
 0 B 0 0 
X= 
 0 0 B 0 
0 0 0 B
and:
 
1 0 0 0

 0 0 0 0 


 0 0 0 0 


 0 0 0 0 


 0 0 0 0 


 0 0 0 0 

0 1 0 0
 
 
0 1 0 0
 
 
0 0 0 0
 
 
0 0 0 0
 
 
 

 0 0 0 0 

 

 0 0 0 0 

C 0 0 0 
 0 0 0 0 

 0 C 0 0   0 0 1 0 
M = 
 0 0 C 0  , J = 

 0 0 1 0 

0 0 0 C 
 0 0 0 0 


 0 0 0 0 


 0 0 0 0 


 0 0 0 0 


 0 0 0 0 

0 0 0 1
 
 
0 0 0 1
 
 
0 0 0 0
 
 
0 0 0 0
 
 
0 0 0 0
 
 
 

 0 0 0 0 

 0 0 0 0 
1 0 0 0
One may check that the Minkowsky Island has a strong harmonic structure, and that:
1 λ (4 λ(λ(2λ − 7) + 7) − 7)
T + J T (X + λM )−1 J = − D+ T
8(λ − 1) (2 (λ − 2) λ + 1) (8 (λ − 2)λ (λ − 1)2 + 1) 8 (λ − 2)λ (λ − 1)2 + 1

We obtain, as previously:
1
=8
KD (0)

36
The spectral decimation function is given, for any real number λ, by:

R(λ) = −(λ − 4)(λ − 2)2 λ((λ − 4)λ + 2)2

which leads to the same spectrum as the one of the Curve.

References
[BD85] A. Beurling and J. Deny. Espaces de Dirichlet. i. le cas élémentaire. Annales scientifiques de
l’É.N.S. 4 e série, 99(1):203–224, 1985.

[Cap02] R. Capitanelli. Nonlinear energy forms on certain fractal curves. J. Nonlinear Convex Anal.,
3(1):67–80, 2002.

[FL04] U. R. Freiberg and M. R. Lancia. Energy form on a closed fractal curve. Journal for Analysis
and its Applications, 23(1):115–137, 2004.

[FS92] M. Fukushima and T. Shima. On a spectral analysis for the Sierpiński gasket. Potential Anal,
1:1–3, 1992.

[FT13] U. R. Freiberg and C. Thäle. Exact Computation and Approximation of Stochastic and
Analytic Parameters of Generalized Sierpinski Gaskets. C. Methodol. Comput. Appl. Probab.,
15(3):485–509, 2013.

[Har93] J. Harrison. StokesŠ theorem for nonsmooth chains. Bull. of AMS, 29(2):235–242, 1993.

[HN91] J. Harrison and A. Norton. Geometric integration on fractal curves in the plane. Bull. of
AMS, 402:567–594, 1991.

[HN92] J. Harrison and A. Norton. The Gauss-Green theorem for fractal boundaries. Duke Math. J.,
673:575–586, 1992.

[Hut81] J. E. Hutchinson. Fractals and self similarity. Indiana University Mathematics Journal,
30:713–747, 1981.

[Kig89] J. Kigami. A harmonic calculus on the Sierpiński spaces. Japan J. Appl. Math., 8:259–290,
1989.

[Kig93] J. Kigami. Harmonic calculus on p.c.f. self-similar sets. Trans. Amer. Math. Soc., 335:721–
755, 1993.

[Kig98] J. Kigami. Distributions of localized eigenvalues of laplacians on post critically finite self-
similar sets. Journal of Functional Analysis, 156(1):170–198, 1998.

[Kig03] J. Kigami. Harmonic analysis for resistance forms. Journal of Functional Analysis, 204:399–
444, 2003.

[KS83] J. G. Kemeny and J. Laurie Snell. Finite Markov chains. Springer, 3rd printing 1983 edition,
1983.

[Man77] B. B. Mandelbrot. Fractals: form, chance, and dimension. San Francisco: Freeman, 1977.

[Mos02] Umberto Mosco. Energy functionals on certain fractal structures. Journal of Convex Analysis,
9(2):581–600, 2002.

[Sab97] C. Sabot. Existence and uniqueness of diffusions on finitely ramified self-similar fractals.
Annales scientifiques de l’É.N.S. 4 e série, 30(4):605–673, 1997.

37
[Str99] R. S. Strichartz. Analysis on fractals. Notices of the AMS, 46(8):1199–1208, 1999.

[Str06] R. S. Strichartz. Differential Equations on Fractals, A tutorial. Princeton University Press,

2006.

[UD14] R. Uthayakumar and A. N. Nalayini Devi. Some properties on Koch curve. In C. Bandt, M.
Barnsley, R. Devaney, K. Falconer, V. Kannan, P. B. Kumar (eds), Fractals, Wavelets, and
their Applications. Springer Proceedings in Mathematics & Statistics, 2014.

[Zho07] Denglin Zhou. Spectral Analysis of Laplacians on Certain Fractals. PhD thesis, University
of Waterloo, 2007.

38