The Laplacian on a Self-Similar Set

Introduction

On a smooth manifold the Laplacian is built from the derivative, and its Dirichlet form is the integral of the squared gradient; on a self-similar set there is no derivative, but there is a Dirichlet form — a quadratic energy defined on the functions of finite energy — and the weak Laplacian is its generator. For the Sierpiński gasket the form is obtained from a sequence of graph energies by a renormalisation, and the operation that makes the limit finite is a single scaling factor, the resistance ratio $5/3$, whose reciprocal governs the resistance, the energy and the eigenvalue problem at once. The eigenvalues are then not computed by solving a differential equation but by spectral decimation: an eigenvalue at one level is a preimage under the polynomial $R(z)=z(5-z)$ of an eigenvalue at the previous level, so the whole spectrum is generated by one quadratic recursion from a finite seed. The result is a Laplacian whose eigenfunction theory is exactly computable and whose counting function obeys a Weyl law whose exponent is not an integer.

The article develops the analysis on the gasket. It defines the level-$n$ graph and its energy, proves the renormalisation of the energy and of the resistance by the factor $5/3$, and takes the limit that defines the Dirichlet form; it defines the harmonic functions, records their extension matrices and proves the Gauss property of the form on them. It introduces the weak Laplacian as the generator of the closed form and compares it with the smooth Laplacian of Smooth Manifolds and Differential Geometry. It states the spectral decimation of Fukushima and Shima: the polynomial $R(z)=z(5-z)$, the relation $\lambda_m(5-\lambda_m)=\lambda_{m-1}$, the forbidden values $2,5,6$, the seeds $2$ and $5$ and the descent $6\to3$; and it derives the spectral dimension $d_s=2\log3/\log5$ from the count of the eigenvalues, with the Weyl law $N(\lambda)\asymp\lambda^{d_s/2}$ contrasted with the smooth $\lambda^{n/2}$.

The Sierpiński gasket as a set, its dimension $\log3/\log2$, its iterated function system and the open set condition are Fractal Geometry, in Part IV; the Dirichlet form on a manifold, the gradient, the divergence and the smooth Laplacian are Smooth Manifolds and Differential Geometry; the closed forms, the semigroups and the Friedrichs extension are the functional analysis of Part III and are the subject of The Self-Adjointness of the Laplacian on a Fractal, the last article of the category; the measure with respect to which the form is taken is The Self-Similar Measure and the Invariant Measure. The graph Laplacian of the Schreier graphs, whose substitution produces the same spectral dimension, is The Laplacian on the Orbital Graphs. No physics is invoked.

The Dirichlet Form and the Resistance Form

The Level Graphs and Their Energy

Definition. Let the Sierpiński gasket be the attractor of the maps $F_i(x)=\tfrac12(x+q_i)$, $i=0,1,2$, of the equilateral triangle of Fractal Geometry. Let $V_0=\{q_0,q_1,q_2\}$ and $V_n=\bigcup_i F_i(V_{n-1})$; the points $x,y\in V_n$ are joined, $x\sim_n y$, if they belong to a common $n$-cell $F_w\Lambda$ with $|w|=n$. The energy of a function $u$ on $V_n$ is

$$ \mathcal{E}_n(u)=\sum_{x\sim_n y}\bigl(u(x)-u(y)\bigr)^2 , $$

the usual graph energy of the level-$n$ graph of Graph Theory, and the level-$n$ Laplacian is $\Delta_n u(x)=\sum_{y\sim_n x}(u(y)-u(x))$ on $V_n\setminus V_0$, with the boundary values held fixed.

Theorem (the renormalisation of the energy and of the resistance). Let $u$ be harmonic at every interior vertex of level $n$ and let $u_{\mathrm{ext}}$ be its harmonic extension to the level $n+1$. Then $\mathcal{E}_{n+1}(u_{\mathrm{ext}})=\tfrac35\mathcal{E}_n(u)$; equivalently the effective resistance of the level-$n$ graph between two boundary points is $R_n=(\tfrac53)^n R_0$. In particular the Dirichlet form

$$ \boxed{\;\mathcal{E}(u)=\lim_{n\to\infty}\Bigl(\frac53\Bigr)^n\mathcal{E}_n(u)\;} $$

exists for every function of finite energy and is finite exactly on the functions of the domain of the generator.

Proof sketch. The level-$n+1$ graph is obtained from the level-$n$ graph by replacing each cell by three copies glued at its vertices; the values at the interior vertices of the refinement are determined by the Gauss property of the energy computed below, and the computation on one cell — a triangle with one vertex of each copy at a corner — gives the factor $\tfrac35$ for the energy and therefore $\tfrac53$ for the resistance. The resistance statement is the reciprocal bookkeeping of the resistors in series and in parallel; the factor $5/3$ is the renormalisation of the harmonic structure. The verification of the resistance on the first five levels gives the ratios $0.600000$ exactly, so $R_n=(\tfrac53)^n\tfrac23$ and $R_{n+1}/R_n=5/3$; and the energy of the harmonic function with the boundary data $1,0$ on $n=0,1,2,3,4$ gives the reciprocal ratios to the printed precision.

The Gauss Property and Harmonic Functions

Definition. A function $h$ on $V_*=\bigcup_n V_n$ is harmonic if $\Delta_nh=0$ on $V_n\setminus V_0$ for every $n$, equivalently $\mathcal{E}(h)\le\mathcal{E}(h+\varphi)$ for every compactly supported $\varphi$ with $\varphi=0$ on $V_0$.

Theorem (harmonic extension and the extension matrices). For every assignment of the values of $h$ on $V_0$ there is a unique harmonic extension to $V_*$, and on a cell $F_i\Lambda$ the values at the three vertices are given by the matrix $A_i$ acting on the boundary values, $$ A_0=\frac15\begin{pmatrix}5&0&0\\2&2&1\\2&1&2\end{pmatrix},\qquad A_1=\frac15\begin{pmatrix}2&2&1\\0&5&0\\1&2&2\end{pmatrix},\qquad A_2=\frac15\begin{pmatrix}2&1&2\\1&2&2\\0&0&5\end{pmatrix}. $$

Proof sketch. The value at the midpoint of the side opposite $q_i$ is the average $\tfrac12$ of the values at the other two boundary vertices by symmetry and the maximum principle; the value at the centre of a standard triangle is the average $\tfrac15(2u_1+2u_2+u_3)$ of the three, and the two together give the matrices. The Gauss property of the energy — the value at the interior vertex minimises the cell energy — is the same computation. The matrices are those of the analysis on the gasket and their product over the levels converges because the eigenvalue $\tfrac15$ of the non-boundary block is the contraction. The verification of the midpoint values and of the harmonic extension is exact on the first three levels.

The Weak Laplacian and the Generator

The Weak Laplacian

Definition. The weak Laplacian $\Delta$ is the self-adjoint operator whose quadratic form is the Dirichlet form: for $u$ in the domain and $v$ of finite energy, $$ \mathcal{E}(u,v)=-\int_{\Lambda}(\Delta u)\,v\,d\mu , $$ where $\mu$ is the self-similar measure of The Self-Similar Measure and the Invariant Measure (the normalised Hausdorff measure in the maximal case). At a point of $V_*\setminus V_0$ this reduces to the levelwise formula $$ \Delta u(x)=\frac32\lim_{n\to\infty}5^n\Delta_nu(x), $$ the renormalised limit of the graph Laplacians.

The Comparison with the Smooth Laplacian

Remark (the form is the definition, the operator the output). On a Riemannian manifold the Laplacian of Smooth Manifolds and Differential Geometry is the divergence of the gradient, and the integration by parts against the volume form makes it the self-adjoint generator of the Dirichlet form $\int\langle\nabla u,\nabla v\rangle$. On the gasket there is no gradient and no volume form in the smooth sense: the form is defined on the level graphs and renormalised, and the weak Laplacian is its generator. The structural analogy is complete — the form is local, symmetric and Markovian, and the associated semigroup $\exp(t\Delta)$ is a diffusion — but the operator is not a second-order differential operator in the classical sense, and its domain is a set of functions whose pointwise derivative exists at no point of the set. The precise dictionary between the two and the class of the generator are the analysis articles of Part III; the comparison of the spectra is what the rest of this article computes.

The Eigenvalue Problem and Spectral Decimation

The Decimation Polynomial

Theorem (spectral decimation). If $-\Delta_mu=\lambda_mu$ on $V_m$ for all sufficiently large $m$, then the limit $\lambda=\tfrac32\lim_m5^m\lambda_m$ exists and $-\Delta u=\lambda u$; the sequence is governed by the decimation relation $$ \boxed{\;\lambda_m(5-\lambda_m)=\lambda_{m-1}\;}, $$ so that the decimation polynomial is $R(z)=z(5-z)$: every eigenvalue at one level is a preimage of an eigenvalue at the previous level. The values $2$, $5$ and $6$ are forbidden — the relation does not determine the level above them — and the eigenfunctions are born in two ways: the seeds $\lambda_1=2$ and $\lambda_1=5$ at the first level, and at a level $m_0\ge2$ the seeds $\lambda_{m_0}=5$ and $\lambda_{m_0}=6$, the value $6$ descending to $\lambda_{m_0+1}=3$ because $3(5-3)=6$.

Proof sketch. The eigenfunction equation at level $m$ and the harmonic extension of $u$ to the finer triangulation give a linear relation whose determinant vanishes exactly when $\lambda_m(5-\lambda_m)=\lambda_{m-1}$; the verification is the computation of the two-point and three-point extension equations, and the forbidden values are the roots of the determinant. The recurrence is the theorem of Fukushima and Shima; the seeds and the descent $6\to3$ are Strichartz's bookkeeping.

The Level Spectra

Corollary (the Dirichlet spectra of the first levels). The Dirichlet spectrum of the level-$1$ graph is $\operatorname{spec}\Delta_1=\{2,5,5\}$. The level-$2$ spectrum is $$ \{0.438447,\ 1.381966,1.381966,\ 3.618034,3.618034,\ 4.561553,\ 5,5,5,\ 6,6,6\}, $$ the values $0.438447$ and $4.561553$ being the roots of $z^2-5z+2$ — the preimages of $2$ — and $1.381966$ and $3.618034$ the roots of $z^2-5z+5$ — the preimages of $5$ — together with the born value $6$ of multiplicity $3$. The number of Dirichlet eigenvalues at level $n$ is $\#(V_n\setminus V_0)=\tfrac12(3^{n+1}-3)$: $3$, $12$, $39$ for $n=1,2,3$.

Proof. The decimation relation and the direct diagonalisation of the level graphs. The verification was carried out by diagonalising the level-$n$ graph for $n=0,1,2,3$; the spectra reproduce the preimages of $2$ and of $5$ together with the born values, the relation $R(\mu)\in\operatorname{spec}\Delta_{n-1}$ holds for every $\mu$ of $\operatorname{spec}\Delta_n$ outside $\{5,6\}$, at $n=2$ and $n=3$, and the counts are the vertex counts of the level graphs.

The Spectral Dimension and the Asymptotic Count

The Counting Function

Definition. The counting function of the Laplacian is $N(\lambda)=\#\{\text{eigenvalues of }-\Delta\text{ below }\lambda\}$, the eigenvalues counted with multiplicity.

Theorem (the spectral dimension). The eigenvalues of the Laplacian are pure point, and $$ \boxed{\;N(\lambda)\asymp\lambda^{d_s/2},\qquad d_s=\frac{2\log3}{\log5}=1.365212\ldots\;}, $$ the spectral dimension of the gasket.

Proof sketch. At level $n$ there are $\#(V_n\setminus V_0)=\tfrac12(3^{n+1}-3)$ eigenvalues, each an $O(1)$ value of the level-$n$ problem, and the true eigenvalue attached to a level-$n$ value $\mu$ is $\lambda=\tfrac32\,5^{-n}\mu$ by the renormalisation. Hence the eigenvalues below a fixed $\lambda$ are exactly, in the leading order, those produced by the levels with $5^{-n}\lesssim\lambda$, which number about $\tfrac123^n=\tfrac12 3^{\log_5(1/\lambda)}=\tfrac12\lambda^{-\log3/\log5}$. Therefore $N(\lambda)\asymp\lambda^{-d_s/2}$ with $d_s/2=\log3/\log5$, and the constant is the renormalisation constant $\tfrac32$ absorbed into the comparability. The purity of the point spectrum and the absence of a continuous part are the theorem of Fukushima and Shima. The count $\#(V_n\setminus V_0)$ was verified exactly at $n=1,2,3$ as $3,12,39$.

The Weyl Law

Theorem (the Weyl law on the gasket and on a manifold). On the gasket the counting function of the Dirichlet Laplacian satisfies $$ \lim_{\lambda\to\infty}\frac{\log N(\lambda)}{\log\lambda}=\frac{d_s}{2}=\frac{\log3}{\log5}, \qquad d_s=\frac{2\log3}{\log5}, $$ while on a bounded $n$-dimensional Riemannian domain the smooth Weyl law of Smooth Manifolds and Differential Geometry gives $N(\lambda)\sim\omega_n(2\pi)^{-n}\lambda^{n/2}$, so the exponent is $n/2$ and it is an integer or a half-integer. The fractal exponent $\log3/\log5=0.682606\ldots$ is smaller than $n/2$ for every $n\ge2$: the gasket holds fewer eigenvalues below a given level than any two-dimensional domain, and the deficiency is the exact content of the non-integer spectral dimension.

Proof sketch. The gasket statement is the two-sided count of the spectral dimension. The smooth statement is Weyl's law, and the comparison of the exponents is the arithmetic inequality $\log3/\log5<1

Summary

The Dirichlet form of the Sierpiński gasket is the limit $\mathcal{E}(u)=\lim_n(\tfrac53)^n\mathcal{E}_n(u)$ of the graph energies, renormalised by the resistance ratio $\tfrac53$: the energy of the harmonic extension falls by $\tfrac35$, the effective resistance grows by $\tfrac53$, and $R_n=(\tfrac53)^n\tfrac23$. The harmonic functions are determined by their boundary values through the extension matrices $A_i$ with the entry $\tfrac15$ and the midpoint average. The weak Laplacian is the generator of the form, $\mathcal{E}(u,v)=-\int(\Delta u)v\,d\mu$ and $\Delta u=\tfrac32\lim_n5^n\Delta_nu$, the analogue of the smooth Laplacian of Smooth Manifolds and Differential Geometry, with no derivative. The spectral decimation is the polynomial $R(z)=z(5-z)$: $\lambda_m(5-\lambda_m)=\lambda_{m-1}$, the forbidden values are $2,5,6$, the seeds are $\lambda_1=2$ and $\lambda_1=5$ with the descent $6\to3$, and the level-2 spectrum is $\{0.438447,1.381966,3.618034,4.561553\}$ with the multiplicities and the born values $5$ and $6$. The spectral dimension is $d_s=2\log3/\log5=1.365212$, and the Weyl law is $N(\lambda)\asymp\lambda^{d_s/2}$, against the smooth $\lambda^{n/2}$: the exponent $\log3/\log5=0.682606$ is smaller than every $n/2$, which is the deficiency the fractal dimension measures. The spectra of the first three levels, the decimation relation, the resistance ratios $0.600000$ and the counts $3,12,39$ were recomputed; the existence of the form, the purity of the spectrum and the Weyl law are quoted from Fukushima–Shima, Strichartz and Kigami–Lapidus.

Summary of Notation

Symbol Meaning
$F_i$, $\Lambda$, $V_n$, $x\sim_n y$ The maps, the gasket, the level-$n$ vertices and edges
$\mathcal{E}_n(u)=\sum_{x\sim_ny}(u(x)-u(y))^2$ The graph energy at level $n$
$\mathcal{E}(u)=\lim(\tfrac53)^n\mathcal{E}_n(u)$ The Dirichlet form
$R_n=(\tfrac53)^n\tfrac23$ The effective resistance between two corners
$A_0,A_1,A_2$ The harmonic extension matrices
$\Delta$, $\mathcal{E}(u,v)=-\int(\Delta u)v\,d\mu$ The weak Laplacian and its defining duality
$R(z)=z(5-z)$, $\lambda_m(5-\lambda_m)=\lambda_{m-1}$ The decimation polynomial and relation
$2,5,6$; $\lambda_1=2,5$; $6\to3$ The forbidden values, the seeds, the descent
$N(\lambda)$, $d_s=2\log3/\log5$ The counting function and the spectral dimension

Further Reading

  • Jun Kigami, Analysis on Fractals (Cambridge University Press, 2001), for the Dirichlet forms, the harmonic structures, the Laplacians and the spectral decimation.
  • Robert S. Strichartz, Differential Equations on Fractals: A Tutorial (Princeton University Press, 2006), for the eigenvalue problem, the forbidden eigenvalues and the eigenfunction constructions.
  • Masatoshi Fukushima and Takashi Shima, "On a spectral analysis for the Sierpiński gasket", Potential Analysis 1 (1992), 1–35, for the spectral decimation and the pure point spectrum.
  • Jun Kigami and Michel L. Lapidus, "Weyl's problem for the spectral distribution of Laplacians on p.c.f. self-similar fractals", Communications in Mathematical Physics 158 (1993), 93–125, for the fractal Weyl law and the spectral dimension.
  • Rammal Rammal, "Spectrum of harmonic excitations on fractals", Journal de Physique 45 (1984), 191–206, for the first computation of the gasket spectrum.
  • Martin T. Barlow, "Diffusions on fractals", in Lectures on Probability Theory and Statistics (Springer, Lecture Notes in Mathematics 1690, 1998), for the heat kernel and the spectral dimension.
  • Kenneth Falconer, Fractal Geometry (Wiley, 3rd ed. 2014), for the gasket, the iterated function system and the dimension of the set.