Effect of Anisotropy on the Localization in a Bifractal System
Extracted source text ยท arXiv cond-mat/9701065
\documentstyle[aps,multicol,epsf]{revtex}
%\documentstyle[aps,pre,preprint]{revtex}
\begin{document}
\preprint{}
\draft
\title{Effect of Anisotropy on the Localization in a Bifractal System}
\author{P. H. Song and Doochul Kim}
\address{Department of Physics and Center for Theoretical Physics, Seoul
National University, Seoul 151--742, Korea}
%\date{\today}
%\ \hspace{5in}{\normalsize SNUTP }\\
\maketitle
\begin{abstract}
Bifractal is a highly anisotropic structure where planar fractals are
stacked to form a 3-dimensional lattice. The localization lengths along
fractal structure for the Anderson model defined on a bifractal are
calculated. The critical disorder
and the critical exponent of the localization lengths are obtained from
the finite size scaling behavior. The numerical results are in a good
agreement with previous results which have been obtained from the
localization lengths along the perpendicular direction.
This suggests that the anisotropy of the embedding lattice structure is
irrelevant to the critical properties of the localization.
\end{abstract}
\pacs{PACS numbers: 71.30.+h, 71.23.An, 71.55.Jv}
\begin{multicols}{2}
\narrowtext
Recently, there has been much attention \cite{bha,qim,roj,abri,xue,zam} to
the critical properties of
the localization in anisotropic system. For a 3-dimensional cubic system
with anisotropic hopping matrix, it seems now that there is a general
agreement \cite{zam} that both the critical disorder ($W_c$) and the
critical exponent ($\nu$) of the localization length are independent of
direction of measurement and that such model belongs to the universality
class of the isotropic Anderson model.
On the other hand, bifractals \cite{shi,sch} are constructed by stacking
planar fractal lattices along the $z$-direction so that they are of
Euclidean structure only in the $z$-direction.
Anisotropy in these systems arises from the lattice structure
itself. Therefore, it is by no means obvious whether the critical properties
obtained from the localization lengths along the $xy$-plane
($\lambda_{xy}$) are the same as those from the localization lengths
along the $z$-axis ($\lambda_z$). There is a possibility that even a
mobility edge, one of useful concepts in the localization theory, may not
exist in this intrinsically anisotropic system.
Therefore, in this Report, we study the critical disorder and the
critical exponent along fractal structure, i.e. along the $xy$-plane,
for a bifractal system. Our results are in excellent agreement with those
obtained from $\lambda_z$'s, which have been reported in previous studies
\cite{sch,gru}, suggesting that the anisotropy of the embedding lattice
structure is irrelevant to the critical properties of the localization. In our
study, one of the bifractals introduced in the Ref. [8] is used as
the model. We consider the Anderson Hamiltonian given as
\begin{equation}
{\cal H} = \sum_i \epsilon_i|i\rangle\langle i| +
\sum_{\langle i,j\rangle} V(|i\rangle\langle j|+|j\rangle
\langle i|),
\end{equation}
where the random site energies $\epsilon_i$ are chosen from a
box distribution of width $W$. The hopping energies $V$ are set to 1
throughout this work and $\langle i,j\rangle$ denotes the sum over nearest
neighbor pairs of sites on the bifractal lattice. In Fig.~1(a), the
lattice is schematically depicted.
The cross section perpendicular to the $z$-axis is a variant of
the Sierpinski gasket and $L$ is the number of the fractal lattices that
have been stacked. The number of the iteration processes for the fractal
lattice is denoted by $n$, e.g. $n$ = 2 for Fig.~1(a). This is
exactly the model called as the Bifractal I in the Ref.
[8]. The Green's function coupling two corner sites of the largest triangle,
{\bf r} and ${\bf r^{\prime}}$, is denoted as $G_{{\bf r},{\bf r'}}(n,L)$.
Then the localization length along the $xy$-plane, $\lambda_{xy}(L)$, can be
defined as follows;
\begin{equation}
\frac{1}{\lambda_{xy}(L)} = -\lim_{n\rightarrow \infty}
\frac{1}{|{\bf r} - {\bf r'}|} \mbox{log}|G_{{\bf r},{\bf r'}}(n,L)|,
\end{equation}
where {\bf r} and ${\bf r^{\prime}}$ have the same $z$-coordinates.
The main point of the calculation is to find $G_{{\bf r},{\bf
r'}}(n,L)$'s for sufficiently large value of $n$, which means that one
should calculate elements of the inverse of a very large
random matrix, $(E - {\cal H})$. This is essentially the same
problem as encountered in the transfer matrix method for
quasi-1-dimensional systems \cite{mac}. However a different recursive
algorithm should be devised since we are considering a ``quasi-2-dimensional
system".
One can handle the problem by decimating recursively the amplitudes of
the sites characterized by the largest iteration number, as shown in
Fig.~1(b). Following scheme is for the case of
$L=1$ but the extension of the method for $L \geq 2$ is straightforward.
Let {\bf x} ({\bf y}) be a vector the elements of which are the
amplitudes of the sites represented by the solid (empty) circles in
Fig.~1(b). Then a matrix equation for {\bf x} and {\bf y}
can be constructed in the form,
\begin{equation}
\left( \begin{array}{cc}
H_x(E) & V_{xy}(E)\\
V_{xy}^t(E) & H_y(E)\\
\end{array} \right)
\left( \begin{array}{c}
{\bf x}\\
{\bf y}\\
\end{array} \right)
=
\left( \begin{array}{c}
{\bf z}\\
{\bf 0}\\
\end{array} \right)
\end{equation}
where $H_x(E), H_y(E)$ and $V_{xy}(E)$ are matrices and contributions from
remaining sites other than shown in Fig.~1(b) are contained in a vector
{\bf z}. The vector {\bf 0} in the right hand side of Eq.~(3) represents the
fact that {\bf y} is directly coupled only with {\bf x}. Initially, when
the sites have been indexed as in Fig.~1(b), the explicit forms of the three
matrices are as follows; (i) $H_{x,ij} = \delta_{ij} (E-\epsilon_{x_i})$,
(ii) $H_{y,ii} = E-\epsilon_{y_i}, H_{y,12} = H_{y,21} = H_{y,13}
= H_{y,31} = H_{y,25} = H_{y,52} = H_{y,34} = H_{y,43} = H_{y,46} = H_{y,64}
= H_{y,56} = H_{y,65} = -V$ and $H_{y,ij} = 0$ otherwise, (iii) $V_{xy,11} =
V_{xy,12} = V_{xy,23} = V_{xy,24} = V_{xy,35} = V_{xy,36} = -V$ and $V_{xy,ij}
= 0$ otherwise. Eliminating the amplitudes of the internal sites,
i.e. {\bf y}, we get
\begin{equation}
(H_x-V_{xy}H_y^{-1}V_{xy}^t)\ {\bf x} = {\bf z}.
\end{equation}
By performing the decimation process of Eq.~(4) for every triangle consisting
of nine sites, the number of the whole eigenvalue equations reduces by a factor
of three. The $3 \times 3$ matrix in the left hand side of Eq.~(4) defines the
renormalized hopping energies within the smallest triangles of the new lattice
and the renormalized on-site terms. Since the hopping energies are modified only
within the smallest triangle one can cast the matrix equation for the remaining
sites again in the form of Eq.~(3). It should be also noted that at the first
step of the iteration, the elements of $V_{xy}$ are independent of $E$ as
can be seen from iii), while after the decimation process, i.e. Eq.~(4),
they become functions of $E$, in general. Therefore, the problem has been reduced
to another eigenvalue equation problem on the Sierpinski gasket with the
iteration number smaller than the original by 1. Now we can iterate the
above procedure until 3 linear equations for the amplitudes of the 3 outermost
sites are left in case of $L = 1$. For general $L$, we have $3 \times L$
linear equations instead of 3. Then the inverse of the matrix constructed
by the coefficients of the amplitudes is the Green's function for the outermost
sites.
In principle, this algorithm can be iterated infinitely. However,
one of the technical problems in real calculation is that some off-diagonal
elements of the matrix becomes very small compared with, say, ${\cal O}(1)$
diagonal elements, as the iteration proceeds. In general, these small
elements contain the essential information for our problem since
they are directly related to the Green's function connecting two different
sites far away from each other. For example, let's assume that the
localization length
is of order unity and the distance between two sites is, say, 1000 $\times$
(lattice constant). Then the Green's function connecting the two sites is
of order $e^{-1000} \sim 10^{-430}$ while the diagonal elements of the
Green's function are of order unity. Therefore one is investigating an
asymptotic behavior of vanishingly small matrix elements while the order of
magnitudes of some other elements of the matrix is far much larger than them.
Direct manipulation of such matrix leads to loss of information on the smaller
elements.
One of possible techniques to overcome this difficulty is to decompose the
matrix into two parts as
\begin{equation}
A = A_0 + e^{-\alpha} A^{\prime},
\end{equation}
where the matrix elements of both $A_0$ and $A^{\prime}$ are within the
range safely handled by computers. $\alpha$ is a number and should be
modified whenever the matrix is manipulated. When a matrix is decomposed
as the above, the inverse of the matrix can be calculated by the formula
\begin{equation}
(A_0 + e^{-\alpha} A^{\prime})^{-1} = A_0^{-1} + A_0^{-1} \sum_{n=1}^\infty
(-e^{-\alpha} A^{\prime} A_0^{-1})^n.
\end{equation}
For our calculation, it turns out to be sufficient to retain terms up to
second order in $e^{-\alpha}$.
We calculate $\lambda_{xy}(L)$'s for $E = 0$ and several
values of $W$ in the range $4.0 \leq W \leq 9.0$.
In our calculation, the Sierpinski gasket with $n = 11$ ($3^{12}$
sites/cross section) is used and $L$ is varied within the range
$3 \leq L \leq 15$. For a given set of parameters, configurational
averages are performed over 4 $\sim$ 70 different realizations to control
the uncertainty of $\lambda_{xy}(L)$ within 1 $\%$. Periodic boundary
condition is imposed in the $z$-direction. The results are
shown in Fig.~2. As $L$ increases, the renormalized localization
length, $\lambda_{xy}(L)/L$, increases for smaller values of $W$,
while for larger values it keeps decreasing. This implies that in the
macroscopic limit, i.e., $L \rightarrow \infty$, there exists a
transition from an extended state to a localized state as $W$ varies,
e.g., from 5.0 to 6.0. Estimates of $W_c$
and $\nu$ are obtained by fitting the data to the scaling form
\begin{equation}
\log(\lambda_{xy}(L)/L) = a + b (W-W_c) L^{1/\nu}.
\end{equation}
where $a$ and $b$ are constants.
Several sets of data for $4.5 \leq W \leq 7$ and $7 \leq L \leq 15$,
have been fitted to the equation (7) and finally we get $W_c = 5.79 \pm
0.04$ and $\nu = 2.92 \pm 0.14$. The errors are the
dispersions between different sets of data. The scaling plot with these
values is shown in Fig.~3. Our numerical results are in good agreement
with those obtained from the analysis of $\lambda_z$'s, {\it i.e.}
$W_c = 5.8$ and $\nu = 3.0 \pm 0.2$\cite{gru} for the same model. This
supports the idea that $W_c$ and $\nu$ are independent of the direction
of measurement for this bifractal system.
That $W_c$'s are the same along the two directions indicates that there
exists a well-defined mobility edge in spite of the intrinsic anisotropy in
this model. In addition we can expect that the
anisotropy is irrelevant to the localization transition in a somewhat
strong sense, that is, {\it the critical properties of the localization
is insensitive not only to the anisotropy in the energy parameters but
also to that of the embedding lattice structure}.
We are thankful to M. Schreiber and H. Grussbach for sending their
numerical data. This work has
been supported by the Korea Science and Engineering Foundation through
the Center for Theoretical Physics and by the Ministry of Education through
BSRI both at Seoul National University.
\begin{references}
\bibitem{bha}
R. N. Bhatt, P. W\"{o}lfle and T. V. Ramakrishnan, Phys. Rev. B {\bf 32},
569 (1985).
\bibitem{qim}
Qiming Li, C. M. Soukoulis, E. N. Economou and G. S. Grest, Phys. Rev.
{\bf 40}, 2825 (1989).
\bibitem{roj}
A. G. Rojo and K. Levin, Phys. Rev. B {\bf 48}, 16861 (1993), and references
therein.
\bibitem{abri}
A. A. Abrikosov, Phys. Rev. B {\bf 50}, 1415 (1994).
\bibitem{xue}
W. Xue, P. Sheng, Q. J. Chu and Z. Q. Zhang, Phys. Rev. Lett. {\bf 63},
2837 (1989); Phys. Rev. B {\bf 42}, 4613 (1990); Q. J. Chu and Z. Q. Zhang,
Phys. Rev. B {\bf 48}, 10761 (1993).
\bibitem{zam}
I. Zambetaki, Qiming Li, E. N. Economou and C. M. Soukoulis, Phys. Rev.
Lett. {\bf 76}, 3614 (1996).
\bibitem{shi}
Y. Shi and C. Gong, Phys. Rev. E {\bf 49}, 99 (1994).
\bibitem{sch}
M. Schreiber and H. Grussbach, Phys. Rev. Lett. {\bf 76}, 1687 (1996).
\bibitem{gru}
H. Grussbach, Ph.D. thesis, Johannes Gutenberg-Universit\"{a}t,
1995 (unpublished).
\bibitem{mac}
See B. Kramer and A. MacKinnon, Rep. Prog. Phys. {\bf 56}, 1469 (1993)
and references therein.
\end{references}
\begin{figure}
\centerline{\epsfxsize=7cm \epsfbox{fig1.epsi}}
\vspace{4mm}
\caption{
(a) Schematic diagram of the lattice with the iteration
number $n$ = 2 and $L$ = 5. Each vertex is a lattice site of the
Hamiltonian given by
Eq.~(1). (b) Schematic diagram of the decimation process. The amplitudes of the
sites represented by the empty circles are eliminated to modify the forms of
the equations for the sites represented by the solid circles.
}
\end{figure}
\noindent
\begin{figure}
\centerline{\epsfxsize=9cm \epsfbox{fig2.epsi}}
\caption{
Log-log plot of the renormalized localization length
along the Sierpinski gasket as a function of $L$ for various values
of $W$. The uncertainty of each data point is less than the symbol
size.
}
\end{figure}
\noindent
\begin{figure}
\centerline{\epsfxsize=9cm \epsfbox{fig3.epsi}}
\caption{
Scaling behavior of $\log(\lambda_{xy}/L)$ versus
$(W-W_c)\ $ $ L^{1/\nu}$ with $W_c$ = 5.79 and $\nu$ = 2.92.
}
\end{figure}
\noindent
\end{multicols}
\end{document}