Degenerate harmonic structures on fractal graphs.
Konstantinos Tsougkas
Uppsala University Fractals 6, Cornell University
June 16, 2017
1 / 25
Degenerate harmonic structures on fractal graphs. Konstantinos - - PowerPoint PPT Presentation
Degenerate harmonic structures on fractal graphs. Konstantinos Tsougkas Uppsala University Fractals 6, Cornell University June 16, 2017 1 / 25 Analysis on fractals via analysis on graphs. The motivation comes from analysis on fractals. The
Konstantinos Tsougkas
Uppsala University Fractals 6, Cornell University
June 16, 2017
1 / 25
The motivation comes from analysis on fractals. The pointwise formula connects the Laplace operator on the fractal with the discrete graph Laplacians. ∆u(x) = lim
m→∞ λm∆mu(x)
Harmonic functions satisfy ∆u = 0.
Can we get any information on harmonic functions by studying them at the graph level?
2 / 25
3 / 25
We can define the energy of a function on these graphs as Em(u) = 1 r m
y∼mx
(u(y) − u(x))2 and the combinatorial graph Laplacian as ∆mu(x) =
(u(x) − u(y))
∆m = Dm − Am where Dm and Am are the degree and adjacency matrices of the appropriate graph level approximations.
4 / 25
Equivalent criteria for the Dirichlet problem on the graphs: Minimize energy. ∆mh(x) = 0 for x / ∈ V0. Satisfy mean-value property for x / ∈ V0, i.e h(x) = 1 deg(x)
h(y) They have a probabilistic interpretation.
Can we utilize the self-similarity of K to solve the Dirichlet problem algorithmically?
5 / 25
If |V0| = k, define Ai to be the harmonic extension matrices such that h ◦ Fi(q1) . . . h ◦ Fi(qk) = Ai h(q1) . . . h(qk) For the cell FwK we can use Aw = Awm · · · Aw2Aw1. Since constants remain constants, they are stochastic matrices. The second eigenvalue of the matrices of boundary points is the renormalization constant.
6 / 25
h ◦ Fi(q1) h ◦ Fi(q2) h ◦ Fi(q3) = Ai h(q1) h(q2) h(q3) with A1 = 1 5 5 2 2 1 2 1 2 A2 = 1 5 2 2 1 5 1 2 2 A3 = 1 5 2 1 2 1 2 2 5
7 / 25
Are these Ai invertible matrices? Let h a non-constant harmonic function. Define the energy measure
νh(FwK) = r −|w|E(h ◦ Fw) where r is the renormalization constant. If they are not invertible then non-constant harmonic functions can be locally constant on some cell. If such an h exists, then νh gives to a cell zero measure. Energy measures lack self-similarity but have other good properties (Kusuoka measure/Laplacian). In Rd, non-constant harmonic functions cannot be locally constant.
8 / 25
A lot of results in the field require the non-degeneracy condition. Absolute continuity of different energy measures. Radial distribution of coefficients related to Radon-Nikodym averages. Random matrices and derivatives of p.c.f fractals.
9 / 25
Usually not. Such a harmonic structure is called degenerate.
10 / 25
Can we give a criterion for when a finitely ramified self-similar set has a degenerate harmonic structure?
Can we find self-similar sets that have a non-degenerate harmonic-structure? These matrices are created only on the first graph approximation, so it suffices to study them only on these graphs.
11 / 25
“Bad” graph connectivity properties give us degeneracies.
A graph is k-vertex connected if it cannot become disconnected by removing at most k − 1 vertices. Equivalently, for every pair of vertices there are at least k vertex independent paths connecting
connected.
12 / 25
Let K be a finitely ramified self-similar set and ˜ G1 be the modified G1 graph by adding extra edges connecting all boundary points.
If ˜ G1 has vertex connectivity less than |V0|, then the self-similar set K has a degenerate harmonic structure.
If we have connectivity k < |V0| then we can remove vertices v1, . . . , vk making the graph have at least two connected components. A cell must necessarily be included in C ∪ {v1, . . . , vk} where C is a connected component not containing any boundary points. Then since dim R|V0|×1 > dim Rk×1 we can create a non-constant harmonic function which is constant on C ∪ {v1, . . . , vk} and thus on some cell.
13 / 25
SGk has k(k+1)
2
harmonic extension matrices. Can be generalized to higher dimensions.
14 / 25
Figure: The G1 graph of SG2, SG3 and SG6.
15 / 25
For SG2, A1 = 1
5
5 2 2 1 2 1 2 with eigenvalues
1 5, 3 5, 1
For SG3, A1 = 1 15 15 8 4 3 8 3 4 A4 = 1 15 8 4 3 4 8 3 5 5 5 with eigenvalues λ =
1 15, 7 15, 1 and λ = 2 15, 4 15, 1
They are all invertible for all SGk with k ≤ 50 (numerical calculations, Hino 2009).
16 / 25
For all k ≥ 2, the harmonic extension matrices of SGk are invertible.
For every non-constant harmonic function h the energy measure νh is minimally energy dominant. In particular for every two non-constant harmonic functions h1, h2 the energy measures νh1, νh2 are mutually absolutely continuous.
17 / 25
A graph embedding in a surface is a representation of the graph to the surface associating vertices to points and edges to simple arcs so that the arcs do not include any other vertices or intersect with each other except at their common end points. A graph is planar if it can be embedded in the plane. Kuratowski’s theorem states that a graph is planar if and only if it does not contain a subgraph homeomorphic to K5 or K3,3.
How can we draw a planar graph? The answer to this question provides the machinery to answer the conjecture.
18 / 25
Let G a connected planar graph and take S to be the vertices bounding a face of it. Fix the vertices of S in R2. Think of the edges as ideal rubber bands satisfying Hooke’s Law and let the other vertices settle in equillibrium. This is the rubber band representation
19 / 25
Equilibrium has minimum energy. All free vertices are at the barycenter of its neighbors. Each coordinate function is harmonic on V \ S.
If G is a simple 3-connected planar graph, then its rubber band representation with respect to any of its faces is an embedding of G in the plane.
20 / 25
It is obviously planar. Extra boundary edges do not perturb Laplace’s equation and make the graph at least 3-connected. q1 q2 q3 (0, 1) (x1, 0) (x2, α)
21 / 25
(0, 0) (− 1
√ 6, 1 √ 2)
( 1
√ 6, 1 √ 2)
Figure: A barycentric embedding of the second graph approximation of SG2.
Kusuoka measure becomes now renormalized energy of the graph representation.
22 / 25
23 / 25
Hino’s conjecture is also stated for SG d
k for all d ≥ 2.
What conditions along with |V0| or higher vertex connectivity are also sufficient? A similar approach works only for weighted edges.
24 / 25
25 / 25