PHYSICAL REVIEW B 83, 085412 (2011)

Probing the surface-to-bulk transition: A closed-form constant-scaling algorithm

for computing subsurface Green functions
Matthew G. Reuter,
*
Tamar Seideman, and Mark A. Ratner
Department of Chemistry, Northwestern University, Evanston, Illinois 60208-3113, USA
(Received 8 November 2010; revised manuscript received 8 January 2011; published 15 February 2011)
Aclosed-formalgorithmfor computing subsurface Green functionsthe blocks of a materials Green function
between the surface and the bulkis presented, where we assume the system satises a common principal-layer
approximation. By exploiting the block tridiagonal and nearly block Toeplitz structure of the Hamiltonian and
overlap matrices, this method scales independently of the system size (constant scaling), allowing studies of large
systems. As a proof-of-concept example, we investigate the decay of surface effects in an armchair graphene
nanoribbon, demonstrating the persistence of surface effects hundreds of atomic layers (0.5 m) away from a
surface. We nally compare the surface-to-bulk transitions of nite and semi-innite systems, nding that nite
systems exhibit amplied surface effects.
DOI: 10.1103/PhysRevB.83.085412 PACS number(s): 73.20.At, 02.60.Dc, 73.22.Dj, 73.90.f
I. INTRODUCTION
The retarded Green function (GF)
1
of a particular Hamil-
tonian,
G(E) = lim
0

[(E i)S H]
1
, (1)
where E is the energy, H is the Hamiltonian matrix, and
S is the overlap matrix, is a useful theoretical construct for
describing both static and dynamic properties of the system.
2
Since obtaining the GFis equivalent to solving the Schr odinger
equation, single-particle formalisms (for example, Hartree-
Fock theory or Kohn-Shamdensity functional theory) are often
used to describe all but the simplest systems.
3
In the context
of condensed-matter physics,
2,4
these single-particle GFs are
useful for their applicability to both bulk and surfaced systems,
to band structures, and to response properties (e.g., electron
transport
5
or magnetoresistance
6,7
).
Since surface effects rapidly decay in three-dimensional
materials (5 atomic layers),
810
most work has historically
focused on the surface and bulk GFs,
2,4
that is, the blocks of
G(E) at and innitely far from a surface, respectively. The
contemporary interest in low-dimensional materials, however,
exposes a need for subsurface GFs [the other blocks of G(E)]
due to the pervasion of surface effects deep into such materials
(50 atomic layers).
1013
Understanding the length scales and
characteristics of these slow progressions from surfacelike to
bulklike environments is important for designing novel devices
that incorporate, for example, carbon nanotubes or graphene
nanoribbons.
Direct, ab initio simulations of these systems become
expensive as the number of atomic layers increases. Division
of the system into principal layers (PLs) is one common
simplication,
14,15
where each PL is a group of atomic
layers that is sufciently large to only interact with its
nearest-neighbor PLs. Within this PL approximation, the
Hamiltonian/Fock/Kohn-Sham (hereafter Hamiltonian) and
overlap matrices become block tridiagonal
16
and, due to crystal
periodicity, also nearly block Toeplitz
17
(deviations from a
block Toeplitz matrix are caused by disorder, including defects
and surface reconstructions). Thus, per Eq. (1), calculation
of the GF is tantamount to inverting a block tridiagonal and
nearly block Toeplitz matrix, and the primary computational
benets of the PL approximation lie in exploiting this
structure.
1821
Such a PLapproach was recently combined with a recursive
algorithm for inverting block tridiagonal matrices
19,22
to
explore the decay of surface effects in carbon nanotubes
12
and graphene nanoribbons.
13
Analysis of this method reveals
that it is only applicable to systems with a nite number of PLs
and, more importantly, that it scales linearly with the number of
PLs. In light of the long surface effect decay lengths exhibited
by low-dimensional materials, linear scaling is undesirable,
and, in this work, we extend the previous method to produce
an algorithm that (i) scales constantly (i.e., its computational
cost is independent of the number of PLs) and (ii) can also
treat semi-innite systems.
The computational improvements reported herein stem
from exploitation of the nearly block Toeplitz structure, as
facilitated by a reinterpretation of several quantities in the
recursive matrix inversion algorithm.
19
In essence, calculating
any block of G(E) reduces to calculating surface GFs of
arbitrarily sized systems (vide infra). Numerous algorithms
have been devised for this task
4
and can be generally cate-
gorized in two ways. First are iterative techniques, including
the aforementioned recursive method
19
and the decimation
method,
15
which use an iterative, PL-by-PL approach to
calculate the surface GF. Second are eigenvalue techniques,
for example, the transfer/companion matrix methods
14,2327
and the M obius transformation method,
28
which utilize the
materials complex band structure to formulate the surface GF.
For calculating subsurface GFs, we seek a surface GF
algorithm that (i) is closed-form (i.e., not self-consistent),
(ii) is constant-scaling, (iii) can handle both semi-innite and
arbitrarily sized nite systems, and (iv) does not impose addi-
tional restrictions on the blocks of [G(E)]
1
. None of the
existing surface GFmethods meet all four criteria. The iterative
methods are closed formand restriction free; however, they are
formally limited to nite systems and are not constant scaling.
Extensions of the iterative methods to semi-innite systems
are made by truncating the system after a sufcient number
of PLs, where sufcient is determined self-consistently
085412-1 1098-0121/2011/83(8)/085412(9) 2011 American Physical Society
MATTHEW G. REUTER, TAMAR SEIDEMAN, AND MARK A. RATNER PHYSICAL REVIEW B 83, 085412 (2011)
(not closed form). Conversely, the eigenvalue methods are
closed form, capable of handling both nite
28
and semi-innite
systems,
14,2325,28
and are constant scaling. Unfortunately, they
usually require the off-diagonal blocks of [G(E)]
1
to be
invertible, a condition which is seldom satised.
26
Here we
construct an acceptable surface GF algorithm for computing
subsurface GFs by augmenting the M obius transformation
method
28
[already achieving (i), (ii), and (iii)] with ideas
from Ref. 27 on how to overcome singular off-diagonal
blocks [(iv)].
The layout of this paper is as follows. Section II rigor-
ously states the PL approximation, summarizes the recursive
algorithm for inverting a block tridiagonal matrix,
19
and intro-
duces our example systema graphene nanoribbonwhich
motivates aspects of the discussion throughout. We proceed,
in Sec. III, to develop our surface GF technique. Section IV
demonstrates proof of concept of our method by calculating
subsurface GFs of the exemplary graphene nanoribbon. The
possible disparity of surface effect decay lengths (so-called
surface depths) between nite and semi-innite systems
is of particular interest. Finally, we conclude in Sec. V and
discuss potential applications of this work beyond calculating
subsurface GFs.
II. GREEN FUNCTIONS WITHIN A PRINCIPAL-
LAYER APPROXIMATION
Within a PL approximation, each PL only interacts with
its nearest-neighbor PLs, making the Hamiltonian and overlap
matrices block tridiagonal. We make two additional assump-
tions, for simplicity. First, the Hamiltonian is Hermitian,
H
mn
= H

nm
. Second, all subsurface PLs are identical to the
bulk PL. This last assumption only allows disorder in the
surface PLsprobably surface reconstructionsresulting in
a Hamiltonian matrix that is nearly block Toeplitz. Thus, the
Hamiltonian (and similarly S) has the form
H =

H
L
H

LB
0 0 0
H
LB
H
B
H

BB
0 0
0 H
BB
H
B
0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 H
B
H

BR
0 0 0 H
BR
H
R

, (2)
where H
L
is the Hamiltonian for the left surface PL, H
B
is the
Hamiltonian for the bulk PL, H
R
is the Hamiltonian for the
right surface PL, H
LB
is the coupling between the left surface
PL and a bulk PL, H
BB
is the coupling between two bulk
PLs, and H
BR
is the coupling between a bulk PL and the right
surface PL.
Consider the armchair graphene nanoribbon (ACGNR)
depicted in Fig. 1 as an example of such a system. Here the
bulk PL consists of two atomic layers,
29
whereas the left and
right surface PLs have 12 atomic layers to account for surface-
induced deviations from the bulk geometry. Throughout this
work we are interested in systems of either nite or semi-
innite size. Finite systems, as explicitly described in Eq. (2),
are constructed by inserting N 1 bulk PLs between the left
FIG. 1. The armchair graphene nanoribbon considered through-
out this work. The left and right PLs contain 12 atomic layers, while
the bulk contains two. For reference, one atomic layer in the bulk PL
is highlighted. Finite-sized systems are constructed by inserting the
desired number of bulk PLs between the left and right PLs, whereas
a semi-innite system is simply the left PL and an innite number of
bulk PLs.
and right PLs, such that there are N 2 PLs in total. The
semi-innite system, conversely, has the left PLfollowed by an
innite number of bulk PLs. Computationally, the Hamiltonian
and overlap matrices were obtained using the procedure in
Ref. 13 with the STO-3G basis set (for simple proof-of-
concept) and the unrestricted local density approximation
(LDA) of the density functional theory. The ensuing example
calculations only consider the spin-up electrons and shift the
energy coordinate about the bulk Fermi level, E
F
= 1.87 eV.
Regardless of the size and structure of the Hamiltonian, we
compute G(E) by inverting
M(E) (E i) S H, (3)
where = 1 meV (unless otherwise specied) approximates
the limit 0

. Note that M(E) is not Hermitian. Addi-

tionally, since almost every quantity throughout this work is a
function of E, we suppress such dependence for brevity.
A. The self-energy
Before proceeding to the recursive algorithm for inverting
a block tridiagonal matrix, it is useful to introduce the self-
energy from chemisorption theory.
5,30,31
Since each surface
or subsurface GF is but a small block of the total GF, it is
conceptually convenient to separate the particular PLs degrees
of freedom from the rest of the system and then describe the
entire system in this smaller basis. As a contextual aside, this
reduction of basis size often appears in chemisorption theory
when considering interactions of a relatively small molecule
with a large substrate.
Adding mathematical rigor, let us divide the total system
into three subspaces: the desired subspace for PL n, the left
subspace for all PLs to the left of PL n, and the right subspace
for all PLs right of PLn. Partitioning the matrix Maccordingly,
we have
M=

M
L
M
Ln
0
M
nL
M
n
M
nR
0 M
Rn
M
R

, (4)
where M
ij
are the coupling matrices between two subspaces,
M
i
are the M matrices of the isolated subspaces, and the
subscripts n, L, and R denote the subspaces for PL n, for
the left, and for the right, respectively. From this partitioning
scheme, an effective M matrix in PL ns subspace can be
written as
MM
n

R
, (5)
085412-2
PROBING THE SURFACE-TO-BULK TRANSITION: A . . . PHYSICAL REVIEW B 83, 085412 (2011)
where

L
= M
nL
M
1
L
M
Ln
(6)
is the self-energy of the interface between PL ns subspace and
the left subspace and

R
= M
nR
M
1
R
M
Rn
(7)
is similarly the self-energy of the interface between PL ns
subspace and the right subspace. FromEqs. (6) and (7),
L
and

R
depend on both the coupling between the subspaces and the
GF of the left or right subspace (M
1
L(R)
). These self-energies
essentially change the boundary conditions of PL ns subspace
to account for the left and right subspaces, which are no longer
explicitly considered.
B. Inverting a block tridiagonal matrix
Examination of Eqs. (1), (2), and (3) shows, as foreshad-
owed, that the problem at hand is inverting a block tridiagonal
matrix. For convenience, we index our system such that the
left surface PL is 0, the right surface PL is N 1, and
the subsurface PLs are 1,2, . . . ,N. Under these conditions
(the Hamiltonian is block tridiagonal, Hermitian, and nite),
Ref. 19 shows how to recursively calculate all blocks of the
GF. Note that this algorithm does not require the nearly block
Toeplitz structure.
Generalizing the formulation of Ref. 19,
32
we start with the
diagonal blocks,
G
n,n
= [M
n,n
X
n
Y
n
]
1
, (8)
for n = 0,1, . . . ,(N 1). X
n
and Y
n
are computational
intermediates and are calculated recursively:
X
N1
= 0, (9)
X
n1
= M
n1,n
[M
n,n
X
n
]
1
M
n,n1
, 0 < n (N 1),
(10)
Y
0
= 0, (11)
Y
n1
= M
n1,n
[M
n,n
Y
n
]
1
M
n,n1
, 0 n < (N 1).
(12)
Finally, the off-diagonal blocks of G are calculated from the
diagonal blocks,
G
n,m
= [M
n,n
X
n
]
1
M
n,n1
G
n1,m
, n > m, (13)
G
n,m
= [M
n,n
Y
n
]
1
M
n,n1
G
n1,m
, n < m. (14)
While Eqs. (8)(14) present a complete computational
picture, further analysis facilitates both a faster algorithm and
the ability to treat semi-innite systems. A comparison of
Eqs. (9) and (10) with Eq. (7) reveals that X
n
is precisely
R
,
the self-energy between PL n and all PLs to its right. In a
similar fashion, Y
n
is
L
, the self-energy between PL n and
all PLs to its left. Application of Eq. (5) additionally shows that
[M
nn
X
n
]
1
is the (left) surface GF for the isolated system
of PLs n, . . . ,N, and, similarly, [M
nn
Y
n
]
1
is the (right)
surface GF for PLs 0, . . . ,n.
III. CALCULATING SURFACE GREEN FUNCTIONS
This new interpretation of the quantities X
n
and Y
n
, when
combined with Eq. (8), shows that calculating PL ns subsur-
face GF essentially requires the self-energies for coupling the
isolated PL n to all PLs on both its left (Y
n
) and its right (X
n
).
Extensions to semi-innite materials are straightforwarduse
the self-energies of semi-innite systemsand such a result
has been previously employed to study tight-binding models.
10
Furthermore, Eqs. (6), (7), (10), and (12) reveal that, within the
PL approximation, calculating any one of these self-energies
is practically equivalent to computing a surface GF.
Numerous methods for calculating surface GFs have been
developed in the last 50 years; see Ref. 4 for a review.
For use in computing subsurface GFs, we digress to seek
a surface GF algorithm that (i) is closed form, (ii) scales
constantly, (iii) applies to both nite and semi-innite systems,
and (iv) makes no additional assumptions about the blocks
of M. We furthermore neglect any surface reconstructions
throughout this section, returning to their inclusion later.
In this case, since the Hamiltonian and overlap matrices
are block Toeplitz, M is also block Toeplitz, and we
write the diagonal, superdiagonal, and subdiagonal blocks of
M as
M
00
(E) = (E i)S
B
H
B
, (15)
M
01
(E) = (E i)S

BB
H

BB
, (16)
and
M
10
(E) = (E i)S
BB
H
BB
, (17)
respectively.
Finally, the block Toeplitz (periodic) structure of M
encourages the use of Blochs theorem; before proceeding, we
reviewsome key elements of complex band structure theory. A
materials complex band structure encompasses all (possibly
complex) Bloch k vectors that satisfy the Schr odinger equation
for a given energy. The eigenvalue surface GF methods obtain
and use these Bloch vectors in various ways; commonly, how-
ever, they rewrite the Schr odinger equation as an eigenvalue
equation in , where
= e
ika
(18)
and a is the lattice constant.
26
States with [[ = 1 (i.e., real k)
propagate into the bulk, whereas states with [[ ,= 1 exponen-
tially grow and decay into the material at rates proportional
to Im(k). Furthermore, the use of innitesimal imaginary
energies, E i, precludes the existence of propagating
states,
23,28
although such states will have [[ 1 as 0

.
Finally, since the difference between growth and decay is a
matter of directional orientation, Bloch vectors come in pairs,
{k,k

], with corresponding eigenvalues {,1/

].
A. The M obius transformation method
The M obius transformation method (MTM)
28
straddles
both classes of surface GF algorithms, combining the
materials complex band structure with a PL-by-PL approach.
As such, it immediately satises three of our four criteria: it
is closed form, constant scaling, and applicable to both nite
and semi-innite systems. Unfortunately, the MTM requires
085412-3
MATTHEW G. REUTER, TAMAR SEIDEMAN, AND MARK A. RATNER PHYSICAL REVIEW B 83, 085412 (2011)
M
01
and M
10
to be nonsingular, a rarely met condition.
26
Thus,
after introducing the MTM in this section, we proceed to use
it in constructing our desired surface GF algorithm.
The MTM derives its utility from a matrix M obius
transformation,
28
which is a generalization of the M obius
(bilinear) transformation from complex variables
33
to matri-
ces. Given M M matrices a, b, c, d, and z and a 2M 2M
matrix
A =

a b
c d

, (19)
we dene a matrix M obius transformation as
A z (az b) (cz d)
1
. (20)
Furthermore, it is easily veried that this matrix M obius
transformation retains associativity fromthe canonical M obius
transformation,
A (A
/
z) = (AA
/
) z, (21)
where A
/
is an arbitrary matrix M obius transformation and
AA
/
is the standard matrix product.
The elegance of the MTM results from its ability to
accelerate the PL-by-PL deposition of identical PLs into
a constant-scaling procedure. Suppose that G
N1,L
is the left
surface GF for a material with N 1 PLs. Then, from Eqs. (5)
and (7),
G
N,L
= (M
00
M
01
G
N1,L
M
10
)
1
= T
L
G
N1,L
, (22)
where
T
L
=

0 M
1
10
M
01
M
00
M
1
10

. (23)
Applying Eq. (22) recursively and exploiting associativity of
the matrix M obius transformation, we nd
G
N,L
=

T
N1
L

G
1,L
,
=

P
L

N1
L
P
1
L

G
1,L
, (24)
where G
1,L
= M
1
00
is the surface GF of an isolated PL,
L
is the diagonal eigenvalue matrix of T
L
, and P
L
is the unitary
eigenvector matrix of T
L
.
Equation (24) is a powerful result and deserves several
comments. First, it presents a constant-scaling approach for
calculating the surface GF of a material with an arbitrary
number of PLssimply compute T
L
, diagonalize it, and
apply the aggregate matrix M obius transformation. Second,
the eigenvalues of T
L
are the desired values.
28
If we order
these eigenvalues such that [
1
[ [
2
[ [
2M
[, we nd
from Eq. (20) that
lim
N

N
L
z = 0 (25)
for an arbitrary z since half of the eigenvalues satisfy [[ < 1
and the other half [[ > 1. Thus, the surface GF of a semi-
innite material is
28
G
,L
= P
L
0. (26)
Third, a similar matrix M obius transformation,
T
R
=

0 M
1
01
M
10
M
00
M
1
01

, (27)
exists for computing right surface GFs. Writing T
R
=
P
R

R
P
1
R
,
G
N,R
=

P
R

N1
R
P
1
R

G
1,R
, (28)
G
,R
= P
R
0. (29)
Finally, as foreshadowed, Eqs. (23) and (27) show that the
MTM fails when M
01
or M
10
is singular.
B. Eliminating singularities in M
01
and M
10
Singular (or numerically singular) inter-PL coupling ma-
trices have troubled the eigenvalue surface GF methods since
their conception. Physically, singularity signies evanescent
states in the material that grow and decay very quickly,
that is, states with [Im(k)[ 0.
27,34
Analytically singular
matrices, corresponding to states with [Im(k)[ = , have been
encountered for some model systems
28,34
and generally appear
only for lamentable choices of model parameters. Moreover,
techniques for addressing these cases have been reported.
34
Realistic systems, conversely, may exhibit numerically sin-
gular coupling matrices, indicating the presence of states
that grow/decay rapidly (0 _[Im(k)[ < ).
27
While these
evanescent states may be physical, they can also be introduced
through unnecessarily large PLs.
26
For instance, if the left side
of a PL has negligible interaction with the right side of the
right-neighbor PL, entire rows of the coupling matrix will be
numerically zero.
Regardless of the source of numerical singularity, numeri-
cally singular coupling matrices are, unfortunately, common.
26
Two types of repairs have been devised for semi-innite
systems within transfer/companion matrix formalisms,
26,27
and both use the singular value decomposition (SVD) of M
10
(or similarly M
01
; only M
10
is discussed) to isolate and manage
the near-singular values. The rst type of repair performs aux-
iliary calculations on the near-singular values;
26,27
however,
these procedures are specic to the transfer/companion matrix
methods, and extensions to the MTM are not obvious. On
the other hand, the second repair type approximates M
10
with
a numerically nonsingular effective coupling matrix, M
eff
10
.
27
Such an effective coupling matrix approximation has proven
to be effective for semi-innite systems,
27
and here we explore
its generalization to nite systems.
The effective coupling matrix approximation exploits the
observations that (i) small singular values of M
10
loosely
translate to states with [Im(k)[ 0 and that (ii) states with
[Im(k)[ 0 tend to have vanishingly small contributions to
quantities of interest.
27
As discussed below, this approximation
amounts to articially inating the small singular values of
M
10
, thus reducing the large [Im(k)[ values. In this manner, the
important states ([Im(k)[ 0) are still computed accurately,
while the troublesome states are approximated by states that
grow/decay less rapidly.
27
The tradeoff of this technique lies
in determining the proper amount of singular value ination:
Too little leaves M
eff
10
numerically singular, whereas too much
085412-4
PROBING THE SURFACE-TO-BULK TRANSITION: A . . . PHYSICAL REVIEW B 83, 085412 (2011)
a
b
12.10
12.06
12.02
11.98
0.8
0.7
0.6
0.5
l
o
g

M
1
0
L
P
L
4 2 0 2 4
E E
F
eV

SVD 10
7
10
9
10
11
FIG. 2. Relevant system properties for using the effective cou-
pling matrix approximation to model the ACGNR system. (a) The
condition number of M
10
, showing that the matrix is numerically
singular across the entire spectrum. (b) L, the minimum number of
PLs recommended for using M
eff
10
, when the condition number of
M
eff
10
is
1
SVD
. Systems with fewer than ceiling (L) PLs should use an
iterative method with M
10
, whereas M
eff
10
is a good approximation for
larger systems.
jeopardizes the states negligibility (small [Im(k)[ resemble
propagating states).
The condition number of a matrix, , is a useful metric for
quantifying this compromise. By denition, d
max
/d
min
,
where d
max
and d
min
are the largest and smallest singular values
of the matrix, respectively. Furthermore, is related to the
precision lost when inverting the matrix; = 1 is ideal and
1 indicates a numerically singular matrix.
35
To justify
this discussion, Fig. 2(a) shows that (M
10
) 10
12
for the
ACGNR system, indicating numerical singularity.
Following Ref. 27, construction of M
eff
10
requires the SVD
of M
10
,
M
10
= UDV

, (30)
where U and V are unitary matrices and D is a diagonal
matrix with the singular values, {d
n
], along its diagonal.
By convention, the singular values are sorted descendingly,
d
n
d
n1
, so that (M
10
) = d
1
/d
M
. Choosing a tolerance
parameter, 0 <
SVD
_1, we build the effective singular value
matrix, D
eff
, by inating M
eff
small singular values in D to
d
1

SVD
. Specically,
d
eff
n
= max(d
n
,d
1

SVD
). (31)
Finally,
M
eff
10
= UD
eff
V

, (32)
such that
SVD
controls the condition number of M
eff
10
,
(M
eff
10
) =
1
SVD
. The error introduced by this approximation
is of the order d
1

SVD
.
27
Having replaced M
eff
singular values in the construction
of M
eff
10
, the M
eff
smallest and M
eff
largest [
eff
[ correspond
to states with exaggerated decay lengths. Furthermore, if we
order the eigenvalues as before ([
eff
1
[ [
eff
2
[ [
eff
2M
[),

eff
M
eff
and
eff
2MM
eff
1
are the approximated eigenvalues with
magnitudes closest to one (smallest [Im(k)[), leading to the
longest exaggerated decay lengths. Note that the
eff
values
are obtained from an effective matrix M obius transformation,
T
eff
L
=

0

M
eff
10

1
M
01
M
00

M
eff
10

, (33)
and similarly for T
eff
R
.
If the size of the system is smaller than, or comparable to,
the longest decay length of an approximated state, M
eff
10
will
likely be a poor approximation of M
10
. The natural question
is then, After how many PLs does M
eff
10
become a good
approximation of M
10
? From the discussion of the MTM,
adding a PL to the system has the effect of multiplying by

eff
M
eff
. Given a numerical accuracy threshold > 0, we want
to nd the smallest number of PLs, L, such that [
eff
M
eff
[
L
< .
After some rearrangement,
L >
log()
log

eff
M
eff

. (34)
Figure 2(b) depicts the minimal L, showing that, for the sample
system, very few PLs are needed for M
eff
10
to be a reasonable
approximation. In practice, L is the ceiling of the right-hand
side of Eq. (34). The parameter = 10
4
is used throughout.
Only systems with more than L PLs should be handled
with the MTM (as modied); systems with fewer than L PLs
should use M
10
with an iterative method. In this manner, we
have succeeded in devising a surface GF algorithm that (i) is
closed form, (ii) is constant scaling, even though there is an
induction period for systems with fewer than L PLs, (iii) can
handle both nite and semi-innite systems, and (iv) remains
intact when M
10
is singular. To demonstrate this point, we have
calculated surface GFs for small, nite systems (two to four
PLs) using both the modied MTMand the recursive technique
with M
10
. Figure 3 shows the root-mean-square absolute error
between the two methods. Although the errors are sufciently
small for all displayed choices of
SVD
, the tradeoff between
too much and too little singular value ination is evident. Thus,
the effective coupling matrix approximation is viable for nite
systems, and
SVD
= 10
9
is used for the remainder of this
work.
IV. EXAMPLE: AN ARMCHAIR GRAPHENE
NANORIBBON
The modied MTM developed in the last section provides
the nal component of our algorithmfor calculating subsurface
GFs. For PL n, we rst compute the surface GFs for all
PLs to the left and right (which are readily transformed into
the self-energies Y
n
and X
n
, respectively) and then calculate
the subsurface GF with Eqs. (8), (13), and (14). The only
remaining obstacles are the left and right surface PLs, which
are allowed to vary from the bulk PLs. These surface PLs are
easily incorporated with the recursive technique: We compute
X
N
using Eqs. (9) and (10) and Y
1
using Eqs. (11) and (12),
and then switch to the modied MTM for all subsequent X
n
and Y
n
. Note that X
N
is only needed for nite systems.
We now proceed to demonstrate proof of concept of this
method by examining the subsurface GFs of the sample
ACGNR system. Of particular interest is the progression from
085412-5
MATTHEW G. REUTER, TAMAR SEIDEMAN, AND MARK A. RATNER PHYSICAL REVIEW B 83, 085412 (2011)
a
b
c
10
4
10
6
10
8
10
4
10
6
10
8
10
4
10
6
10
8
R
M
S
A
b
s
o
l
u
t
e
E
r
r
o
r
i
n
G
e
V
1
4 2 0 2 4
E E
F
eV

SVD 10
7
10
9
10
11
FIG. 3. Root-mean-square (RMS) absolute error between all
elements of surface GFs calculated with the recursive technique
(using M
10
) and the modied MTM (M
eff
10
). The surface GFs are
for systems with (a) two bulk PLs, (b) three bulk PLs, and (c) four
bulk PLs. When
SVD
is too small (gray line), the effective coupling
matrix remains numerically singular, incurring error. Increasing
SVD
improves the effective coupling matrix approximation (black line);
however, increasing
SVD
too much (dashed black line) worsens it by
modifying physically important states. While the optimal choice of

SVD
is most likely dependent on both the system and the numerical
precision (e.g. double precision), the errors here are small over a large
range of
SVD
values.
the surface GF to the bulk GF, which, for the ACGNR, is
expected to require many atomic layers.
13
We also want to
examine any disparities in this progression between nite and
semi-innite systems.
The local density of states (LDOS) for atomic layer m will
be used for these purposes and is calculated as
2

m
=
1

Im[Tr
m
(SG)], (35)
where Tr
m
denotes a trace over the desired atomic layer.
Furthermore, if the desired atomic layer is contained in PL
n, we can write

m
=
1

Im[Tr
m
(S
n,n
G
n,n
S
n,n1
G
n1,n
S
n,n1
G
n1,n
)]
(36)
since Sis block tridiagonal. If n = 0, the second terminside the
trace of Eq. (36) should be omitted, and likewise for n = N 1
and the third term when the system is nite. For reference, the
bulk LDOS,
B
, is also calculated with Eq. (36); in this case
there are an innite number of bulk PLs to both the left and
the right.
Figure 4 compares the LDOSs for atomic layers centered
in nite ACGNRs to the bulk LDOS. Note that this central
atomic layer is the mth atomic layer from one surface, but
the (m1)th from the other since each bulk PL contains
13 20
25 50
100 200
500 1000
1500 2500
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3

e
V
1
4 2 0 2 4 2 0 2 4
E E
F
eV
Subsurface Layer Bulk Layer
FIG. 4. LDOSs for the central atomic layers in nite ACGNRs.
The number inset in the bottom-left corner of each plot is the
minimum depth (in atomic layers) from either surface. Discrete
levels are evident in small systems, and these levels coalesce into
bands as the ACGNRs become longer. After 100 atomic layers, the
bands appear as uctuations around the bulk limit. These uctuations
dampen as the depth from the surfaces increases, and the 2500th
atomic layer effectively mimicks the bulk ACGNR. Two surface states
persist in the band gap for 25 atomic layers.
two atomic layers. Unsurprisingly, small nite systems do not
resemble the bulk.
13
First, there are surface states in the band
gap which decay over the rst 25 atomic layers. Second, the
LDOSs are collections of discrete states. Bands appear to form
between atomic layers 25 and 50, although the broadening
of each state (as caused by ) is insufcient to reproduce
the smooth bulk line shape. The further addition of states in
larger systems evens out these uctuations (beyond atomic
layer 100), until the bulk LDOS is recovered around atomic
layer 2500. Since each bulk PL is 0.4 nm in length, this
suggests that surface effects persist for 0.5 m; such an
ACGNR would need to be 1 m in length to exhibit
a bulklike interior. Note that these surface-depth estimates
are heavily reliant on the specic property used to assess
convergence.
Atomic layers in the semi-innite ACGNR display many
of the same qualitative trends as those in nite ACGNRs, as
085412-6
PROBING THE SURFACE-TO-BULK TRANSITION: A . . . PHYSICAL REVIEW B 83, 085412 (2011)
1 10
25 50
100 200
500 1000
1500 2500
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3
10
1
10
1
10
2
10
3

e
V
1
4 2 0 2 4 2 0 2 4
E E
F
eV
Subsurface Layer Bulk Layer
FIG. 5. LDOSs for various atomic layers in a semi-innite
ACGNR. The number inset in the bottom-left corner of each plot
is the depth from the surface (in atomic layers). As with the nite
ACGNRs, uctuations are observed around the bulk limit, after 10
atomic layers. However, unlike the nite systems, these uctuations
are seemingly milder and the 1500th atomic layer resembles the
bulk. The sole surface state decays in 25 atomic layers.
shown in Fig. 5. As before, the surface state decays in the rst
25 atomic layers, and, once past 10 atomic layers, the
m
uctuate around the bulk limit. The most noticeable difference
between the nite and semi-innite systems is the smoothness
of these uctuations in the semi-innite system. Owing to
the innite number of bulk PLs on one side, we never see
the discrete levelsexcepting the surface statein the semi-
innite ACGNR. Furthermore, the uctuations in the semi-
innite system disappear sooner than in the nite systems; the
1500th atomic layer resembles the bulk. Interestingly, this
suggests that the surface depth is 0.3 m for this ACGNR,
not 0.5 m as estimated for the nite systems. The presence
of a second surface in a nite system seems to exacerbate
surface effects.
To quantify these disparities in surface depths, we introduce
the metric
10

m
=

dE[
m
(E)
B
(E)[, (37)
a 10 meV
b 1 meV
c 0.1 meV
1
10
1
10
2
10
3
1
10
1
10
2
5.0
2.0
1.0
0.5

m
1 10 100 1000
m
Semi Infinite Finite
FIG. 6. Deviations of atomic layer m from the bulk,
m
, for
m 7500. At a particular distance from the surface, the atomic
layers in semi-innite systems more closely resemble the bulk than
do those from a nite system. Furthermore, for small m,
m
follows
a power law,
m
a/m
b
. This approximation fails for large m due
to broadening, . For large or reasonable amounts of broadening
[panels (a) and (b), respectively], surface effects seem to decay more
rapidly in nite systems; however, the effects decay at the same rate
for small [panel (c)]. The best-t power laws are (a) semi-innite:
a = 3.3, b = 0.57, R
2
= 0.97, nite: a = 35, b = 0.94, R
2
= 0.99;
(b) semi-innite: a = 2.7, b = 0.28, R
2
= 0.98, nite: a = 14, b =
0.37, R
2
= 0.99; (c) semi-innite: a = 2.3, b = 0.15, R
2
= 0.93,
nite: a = 8.9, b = 0.13, R
2
= 0.99.
where E

and E

are taken to be the lower and upper bounds

of our spectra. Clearly,
m
0 as
m

B
. Figure 6(b)
displays
m
for m 7500 atomic layers and furthermore
shows that
m
m
b
when m is small. Interestingly, b is
larger for the nite systems, indicating that surface effects
decay more quickly, even though
semi-innite
m
<
nite
m
. As m
increases, here for m 400 atomic layers, this power law
breaks down and
m
0 more rapidly. At these large distances
from the surface(s), broadening from causes the atomic
layers LDOSs to become numerically indistinguishable from
the bulk LDOS.
To further demonstrate the effects of broadening, Figs. 6(a)
and 6(c) show the
m
sequences for alternative values of . As
before,
semi-innite
m
<
nite
m
, and the sequences obey power laws
for small m. As expected, more broadening [panel (a)] results
in smaller
m
values, as well as quicker convergence to the
bulk, whereas less broadening [panel (c)] exhibits the opposite
effects. The exponents of the power laws are also functions of
the broadening. Surface effects appear to decay at faster rates
in nite systems with more broadening [panels (a) and (b)];
however, surface effects deteriorate at comparable rates in both
types of systems as the broadening lessens [panel (c)]. Clearly,
the decay of surface effects depends on the experimental or
085412-7
MATTHEW G. REUTER, TAMAR SEIDEMAN, AND MARK A. RATNER PHYSICAL REVIEW B 83, 085412 (2011)
computational resolution, and it is likely that perfect precision
( 0

) would yield an innite surface depth.

10
V. CONCLUSIONS
Progressions from surfacelike to bulklike environments
occur rapidly in three-dimensional materials, whereas surface
effects persist deep into low-dimensional materials (including
carbon nanotubes and graphene nanoribbons). Subsurface
GFs, the blocks of Gbetween the surface and the bulk, provide
one approach for investigating these surface-to-bulk transitions
and have received minimal attention when compared to surface
and bulk GFs. Within a common PL approximation, this work
establishes a framework for computing subsurface GFs that
(i) is closed form, (ii) scales constantly (that is, it does not
depend on the size of the system), and (iii) can handle both
nite and semi-innite systems. The key advantage to this
technique comes from exploiting the nearly block Toeplitz, in
addition to the block tridiagonal, structure of the Hamiltonian
and overlap matrices in condensed-matter systems. As shown
in Sec. II, computing any subsurface GF reduces to calculating
surface GFs, and, in this context, we developed a suitable
surface GF method by combining the MTM with an effective
coupling matrix approximation (Sec. III).
As a proof-of-concept example, we applied this formalism
to an ACGNR (Sec. IV), showing that surface effects may
extend 0.5 m into such a material. A comparison of nite
and semi-innite systems furthermore revealed that surface
effects are, unsurprisingly, more pronounced in nite systems.
While these results should only be interpreted qualitatively,
due to the simple computational methods used in obtaining
the Hamiltonian and overlap matrices, this example system
typies the importance of surface effects and, perhaps, the
illusory nature of bulk in low-dimensional systems.
Several applications for this algorithm are easily envi-
sioned. First, this work limited disorder to surface reconstruc-
tions: All deviations from a block Toeplitz structure occurred
in the surface PLs. Other types of disorder, particularly point
defects, should be easily includable. Abstractly, the modied
MTM presented here accelerates the PL-by-PL deposition
of any number of identical PLs into a single step, and the
presence of defective PLs should not be troublesome. Second,
this formalism may be pertinent to the study of topological
insulators,
36
where the surface is conducting and the bulk
has a band gap. Investigating the possibility of subsurface
states
37
and the decay of surface effects in these materials
may provide additional insight for the conception of novel
devices.
ACKNOWLEDGMENTS
We are grateful to Oded Hod, W. Scott Thornton, Thorsten
Hansen, and Judith C. Hill for helpful conversations and
to the NSF Chemistry Division (Grant No. CHE-0616927)
for their generous support. M.G.R. also thanks the Depart-
ment of Energy Computational Science Graduate Fellowship
Program (Grant No. DE-FG02-97ER25308) for a graduate
fellowship. Finally, this work was partly supported by the
Non-Equilibrium Energy Research Center (NERC), which
is an Energy Frontier Research Center funded by the US
Department of Energy, Ofce of Science, Ofce of Basic
Energy Sciences under Award No. DE-SC0000989.
*
mgreuter@u.northwestern.edu
1
We focus on the retarded GF in this work. Extensions to the
advanced GF are straightforward.
2
E. N. Economou, Greens Functions in Quantum Physics, 3rd ed.
(Springer-Verlag, Heidelberg, 2006).
3
Many-body effects are important in some cases, and many-body
GFs often build upon the single-particle GF.
2
4
J. Velev and W. Butler, J. Phys. Condens. Matter 16, R637 (2004).
5
S. Datta, Electronic Transport in Mesoscopic Systems (Cambridge
University Press, Cambridge, 1995).
6
S. Sanvito, C. J. Lambert, J. H. Jefferson, and A. M. Bratkovsky,
Phys. Rev. B 59, 11936 (1999).
7
J. Mathon and A. Umerski, Phys. Rev. B 63, 220403(R)
(2001).
8
H.-J. Brocksch and K. H. Bennemann, Surf. Sci. 161, 321 (1985).
9
W. T. Geng, M. Kim, and A. J. Freeman, Phys. Rev. B 63, 245401
(2001).
10
M. G. Reuter, J. Chem. Phys. 133, 034703 (2010).
11
Y. Xue and M. A. Ratner, Phys. Rev. B 70, 205416 (2004).
12
O. Hod, J. E. Peralta, and G. E. Scuseria, e-print
arXiv:physics/0609091 (to be published).
13
O. Hod, J. E. Peralta, and G. E. Scuseria, Phys. Rev. B 76, 233401
(2007).
14
D. H. Lee and J. D. Joannopoulos, Phys. Rev. B 23, 4988 (1981).
15
M. P. L opez Sancho, J. M. L opez Sancho, and J. Rubio, J Phys. F
15, 851 (1985).
16
Note that these matrices depend on the in-PL momentum,

k
|
, for
higher-dimensional materials; this

k
|
dependence is suppressed for
notational brevity.
17
Block Toeplitz matrices repeat the same block along a particular
diagonal. In the case of block Toeplitz and block tridiagonal
matrices, there are only three unique blocks: the superdiagonal,
the diagonal, and the subdiagonal.
18
G. Labahn, D. K. Choi, and S. Cabay, SIAM J. Comput. 19, 98
(1990).
19
E. M. Godfrin, J. Phys. Condens. Matter 3, 7843 (1991).
20
B. V. Minchev, J. Comput. Appl. Math. 156, 179 (2003).
21
R.-S. Ran and T.-Z. Huang, Appl. Math. Comput. 179, 243
(2006).
22
O. Hod, J. E. Peralta, and G. E. Scuseria, J. Chem. Phys. 125,
114704 (2006).
23
D. H. Lee and J. D. Joannopoulos, Phys. Rev. B 23, 4997 (1981).
24
D. H. Lee and J. D. Joannopoulos, J. Vac. Sci. Technol. 19, 355
(1981).
25
Y.-C. Chang and J. N. Schulman, Phys. Rev. B 25, 3975 (1982).
26
J. K. Tomfohr and O. F. Sankey, Phys. Rev. B 65, 245105
(2002).
27
I. Rungger and S. Sanvito, Phys. Rev. B 78, 035407 (2008).
085412-8
PROBING THE SURFACE-TO-BULK TRANSITION: A . . . PHYSICAL REVIEW B 83, 085412 (2011)
28
A. Umerski, Phys. Rev. B 55, 5266 (1997).
29
Our bulk PLs are half the size of those used in Ref. 13 due to the
absence of diffuse basis functions. No Hamiltonian matrix element
between second nearest-neighbor PLs was more than 0.8 meV in
magnitude, and only 18 matrix elements of 5476 were larger than
0.1 meV.
30
D. M. Newns, Phys. Rev. 178, 1123 (1969).
31
V. Mujica, M. Kemp, and M. A. Ratner, J. Chem. Phys. 101, 6849
(1994).
32
The formulation in Ref. 19 requires the systemto be nite, in which
case the limit 0

in Eq. (1) is unnecessary. Taking = 0,

M becomes Hermitian, and Ref. 19 makes use of this structure.
In the present work M has no such structure, requiring trivial
modications to the algorithm.
33
J. W. Brown and R. V. Churchill, Complex Variables and Applica-
tions, 7th ed. (McGraw-Hill, New York, 2004).
34
T. B. Boykin, Phys. Rev. B 54, 7670 (1996); 54, 8107
(1996).
35
M. T. Heath, Scientic Computing: An Introductory Survey, 2nd ed.
(McGraw-Hill, New York, 2002).
36
C. L. Kane, Nat. Phys. 4, 348 (2008); J. E. Moore, Nature (London)
464, 194 (2010).
37
S. V. Eremeev, Y. M. Korote, and E. V. Chulkov, JETP Lett. 92,
161 (2010).
085412-9