Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Solid State Communications 124 (2002) 443–447

www.elsevier.com/locate/ssc

Why tight-binding theory?


Walter A. Harrison*
Department of Applied Physics, Stanford University, Stanford, CA 94305, USA
Received 1 July 2002; received in revised form 9 August 2002; accepted 19 August 2002 by M. Cardona

Abstract
In the context of computational physics other methods are more accurate, but tight-binding theory allows very direct physical
interpretation and is simple enough to allow much more realistic treatments beyond the local density approximation. We
address several important questions of this last category: How does the gap enhancement from Coulomb correlations vary from
material to material? Should the enhanced gap be used for calculating the dielectric constant? For calculating the effective mass
in k-dot-p theory? How valid is the scissors approximation? How does one line up bands at an interface? How should we match
the envelope function at interfaces in effective-mass theory? Why can the resulting quantum-well states seem to violate the
uncertainty principle? How should f-shell electrons be treated when they are intermediate between band-like and core-like? The
answers to all of these questions are given and discussed.
q 2002 Elsevier Science Ltd. All rights reserved.
PACS: 71.10.–w; 71.15.Fv; 71.20.Mq
Keywords: A. Semiconductors; A. Surfaces and interfaces; C. Electronic band structure; C. Electron–electron interaction

1. Introduction solids, including semiconductors. This is the realm of


computational physics where the central goal is high
It is most appropriate to raise the question about tight- accuracy and as complete a solution as possible of the
binding theory in this volume honoring the years of quantum-mechanical problem of the many-electron ground
achievement of Professor Salim Ciraci. It was his thesis state, or of excited states. The starting point is generally a
work [1] on the bond-orbital model, in 1974 at Stanford local density approximation (LDA) which reduces the
University, which marked the beginning of the renaissance problem to a one-electron context, though sometimes
of tight-binding theory. This theory had its origins already in corrections to the LDA are made afterward.
the thesis of Felix Bloch [2] in 1928, but was generally Tight-binding theory does not fare well in the context of
considered to be too naive to be worth considering. In a computational physics. It is based on the idea that one-
similar way the nearly-free-electron theory of Bloch’s thesis electron states of molecules and solids can be written as
was not considered meaningful until the advent of combinations of a small number of states of the constituent
pseudopotential theory. In Professor Ciraci’s research it atoms. With such a limited basis it can never be very
became clear that tight-binding theory provided a clear accurate and its use usually involves assumptions and
understanding of the electronic structure of the tetrahedrally approximations of the type that computational physicists
coordinated semiconductors, and a key to simple calcu- wish to avoid. On the other hand, it has the advantage that
lations of the properties of these systems. The subsequent the results and the theory itself are very easy to understand.
developments, including much of Ciraci’s work, have For that reason most results of detailed computational
utilized more accurate methods and more complete treatments of electronic structure are interpreted in terms of
calculations to obtain properties of the entire range of the concepts of tight-binding theory. Further, it has been
possible to use the fact that both tight-binding theory and
* Tel.: þ1-650-723-4224; fax: þ 1-650-725-2189. the opposite extreme, nearly-free-electron theory, are
E-mail address: walt@stanford.edu (W.A. Harrison). meaningful to provide the central parameters for both
0038-1098/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 3 8 - 1 0 9 8 ( 0 2 ) 0 0 5 4 9 - 5
444 W.A. Harrison / Solid State Communications 124 (2002) 443–447

theories [3,4]. Then tight-binding theory becomes an one half electron volt. This not only gives us approximately
independent, though not so accurate, method for predictions the right value, it tells us that it will be about the same for
in all of the areas covered by LDA. GaAs with a dielectric constant of 1 ¼ 11 and similar
Another aspect, upon which we wish to focus here, is that atomic radius, but smaller for germanium with 1 ¼ 16: The
by simplifying the representation of the electronic states so enhancement will be much larger for rock salt with a
greatly it is sometimes possible to treat the many-electron dielectric constant of some 1.7 or for an inert-gas solid with
aspects more realistically than can be done conveniently in 1 near one. (For details, see Ref. 3, 341 ff.) How much use
the context of a full computational treatment of the was it, in contrast, to know that the enhancement arose from
electronic states. We begin with a discussion of the familiar a discontinuity in the energy-dependent exchange and
enhancement of the band gaps over those from LDA. correlation potential?

2. Energy-gap enhancement
3. The susceptibility and the effective mass
The correction to the band gap provides a very clear
Should we use this enhanced gap in the expressions for
illustration of the contrast between the two different
theoretical goals. In 1983 it was pointed out by Sham and dielectric susceptibility and effective mass, both of which
Schlüter [5] and by Perdew and Levy [6] that LDA, if contain the gap? This question was addressed recently by
correctly carried out, leads to band gaps which are Lew Yan Voon and Harrison [7]. The dielectric suscepti-
significantly smaller than the experimental ones. This bility arises primarily from polarization of individual bonds,
came as a shock to many of us since computational admixture to each bond orbital of an antibonding orbital in
physicists at the time had been adjusting some parameter the same site. However, there is no extra Coulomb repulsion
in their LDA calculations to obtain the observed gap, when the admixture is from the same site, just as there is no
concealing this fundamental difficulty. From the papers by correction to the LDA eigenvalues in transfer of electrons
Sham and Schlüter and by Perdew and Levy it was seen that between orbitals in the same atom. The main term in the
there was a discontinuity in the energy-dependent exchange dielectric susceptibility should therefore be calculated using
and correlation potential which enhanced the gap. It was the LDA bands without enhanced gaps. It is not easy to
subsequently incorporated by computational physicists in confirm this by comparison of susceptibilities from LDA
their calculations. calculations and experimental values because of other
It is significant that this description gives no idea how approximations, such as the neglect of local-field correc-
large the correction is, one half an electron volt in silicon, or tions, in the calculations. The finding that the unenhanced
whether it would be larger or smaller in germanium, or what gap is appropriate for calculating susceptibilities would
it might be in rock salt or other systems. The computational suggest to us that the enhancement should not be included in
physicist can tell you the answer. So can tight-binding the band calculation itself, unless one is willing to restrict
theory, but it can also allow you to understand why. discussion to quasiparticle excitations.
An LDA calculation of the electronic structure of a The shift in energy of excitation of U/1 would seem to be
silicon atom will give an energy 1p for the valence p-state, independent of wavenumber, suggested also by the idea of a
approximately equal to the 8 eV energy required to remove discontinuity of the exchange and correlation potential in
the corresponding electron from the free atom, or the energy the gap. Thus we might expect we could use a ‘scissors
gained in adding an electron to the ion, the atom with one operation’ in the gap to uniformly raise the energy of LDA
electron removed. Clearly if we were instead to add an conduction bands to obtain the corresponding excitation
electron to the neutral atom, we gain a very much smaller energies. On the other hand, the effective mass in the
electron affinity because of the extra Coulomb repulsion conduction band can be obtained from k·p theory, with a
from the additional electron in the atom. That repulsion is squared matrix element of the gradient operator between
about valence and conduction-band states, divided by the band
gap. In order to learn which gap to use, Lew Yan Voon and
U ¼ e2 =ra < 7 eV; ð1Þ Harrison [7] returned to a simple tight-binding chain of
atoms, with s-like states forming a conduction band and p-
with ra an atomic radius of some 2 Å. Similarly if we have a like states forming a valence band. When the band gaps are
collection of well-separated silicon atoms, removing a p- small, so the k·p shifts in the mass are large, the dominant
electron from one atom and placing it in a p-state of a distant terms arise from coupling between p-states and neighboring
atom costs an extra energy U, though the eigenvalue 1p is the s-states [3,4], a coupling which transfers electronic charge
same. It is also true that if these atoms are part of a solid with from p-states on one atom to s-states on neighboring atoms
dielectric constant, 1 < 12 for silicon, the potential in Eq. with fully occupied p-states. The coupling to s-states on the
(1) is reduced to e 2/(1r ) so the energy required exceeds the same site is much smaller [3,4]. Thus the enhanced gap
difference in the LDA eigenvalues by U/1 < 7/12, about should be used, as concluded by Cardona’s group [8] on the
W.A. Harrison / Solid State Communications 124 (2002) 443–447 445

basis of experimental measurements of the effective mass. difficulty comes in the finding of realistic boundary
Use of the enhanced gap is contrary to the use of the scissors conditions.
operation since it changes the band curvature rather than The immediate suggestion for matching effective-mass
shifting the band as a whole. Again, it appears best to keep envelope functions at an interface would be to match c and
the LDA bands, and correct each property as needed. 7c at the boundary, as in ordinary quantum theory.
However, we may see immediately that this is not correct
if the mass differs on the two sides [10]. With a current
operator given by 2ðe"=imp Þcp 7c the current is then not
4. Matching bands and states at interfaces continuous across the interface. Of course one can use
continuity of the current as one condition, but the second
In discussing the electronic structure of interfaces remains uncertain, the most common choice currently
between semiconductors, the question arises immediately being to match c and (1/m p)7c, sometimes called the
of how to set the energy scales, which are generally arbitrary BenDaniel – Duke [11] form or the Bastard [12] boundary
in a band calculation, relative to each other on the two sides. condition.
Tight-binding theory would at first seem to answer the We can also use tight-binding theory to find the correct
question immediately since the band structures are obtained form [13], though this is disappointing at first. We can
relative to the same set of atomic energy levels for all construct simple tight-binding s-bands for each side of an
constituent atoms, giving a ‘natural’ scale for the bands. interface with nearest-neighbor coupling Vsss on the two
However, more careful thought [9] indicates that this does sides to correspond to different effective masses. We can
not set the correct scale at an interface. If the two hybrid couple them at the interface and see what matching
orbitals forming the interface bond between the two conditions on the envelope function would give the correct
result. The disappointment is that we find the c on the right
semiconductors have different energy, a polar bond will be
is given by a constant times c on the left, plus another
formed. This produces a dipole which shifts the bands on the
constant times 7c on the left. 7c is similarly written with
two sides relative to each other. It is not difficult to see [9]
two more constants. The only way simple conditions will be
that this has the effect of reducing the difference in hybrid
obtained is if we assume that the Vsss across the interface is
energies by a factor of the reciprocal of the large
given by the geometric mean of the Vsss on the two sides
semiconductor dielectric constant. Thus the bands shift till
[13]. That is a reasonable choice and the conditions become
the hybrid energies are nearly the same on both sides. As a
a continuous (1/m p)1/2c and a continuous (1/m p)1/27c [13,
consequence, we can make a table (Table 19-1 of Ref. 3) of
14] We are not guaranteed that this is a correct matching, but
the valence-band maxima relative to the hybrid energy (the
if there is a correct simple choice this is it, and other forms-
average hybrid energy in the case of compounds such as
such as the BenDaniel – Duke, or Bastard, form-are certainly
GaAs, equal to the average of the energy of the eight valence incorrect at least in this case where it could be tested.
orbitals for the system), and the offset of the valence bands It is interesting to explore the consequences when the
is obtained by subtracting these tabulated values. The effective-mass change is large, as for a very small mass
conduction bands are then given using the enhanced gap on outside a quantum well. Then the envelope wavefunction for
each side of the interface if we wish to discuss quasiparticle the lowest quantum-well state becomes nearly constant
excitations. This offset of the valence-band orbitals could within the well and drops discontinuously to a very small
have been obtained from a full multilayer LDA band value outside. The state is analogous to the standing sound
calculation, carried out self-consistently with alternate wave in an open organ pipe. In this case the energy of the
layers of the two semiconductors. If the layers were confined state is very near zero in the effective-mass band
sufficiently thick, the offset should be obtainable from the and there is no confinement energy as we might anticipate
resulting states. from the uncertainty principle for a state localized in a
The use of the effective mass concept is valid and useful thickness L with a mass m p, an energy "2 k2 =2mp with k ¼
when we work with states near the band edge in a p=L: Of course there is no difficulty. There is plenty of
semiconductor, and where the dynamics of wave packets kinetic energy in the periodic part of the Bloch function but
is accurately describable by an effective-mass Hamiltonian. we have been misled by applying a vague ‘principle’ to an
It has been widely used in treating quantum wells and other unfamiliar circumstance. We have confirmed this lack of
systems with interfaces between epitaxially grown semi- confinement energy for an easily soluble illustration, a
conductors. Some workers have questioned the validity of Kronig – Penney model for a finite chain, with strong
effective-mass theory in quantum wells so that construction attractive delta functions [14]. The lowest state is a sum of
of a packet within the well would require almost the entire states m1=2 e2mlx2xj l with nearly equal coefficients. The
band. However, tight-binding theory shows that the effective mass inside is large and free-electron outside, and
effective-mass bands remain valid, and the boundaries the envelope function is nearly constant inside. The apparent
only affect what wavenumbers enter, even for wells violation of the uncertainty principle could be very useful in
containing as few as two atomic planes. The only possible some application and is missed completely if one uses an
446 W.A. Harrison / Solid State Communications 124 (2002) 443–447

incorrect matching condition, such as the BenDaniel – Duke in estimating it from LDA and certainly it would be
form. inappropriate to decide that at some point the states were
‘core-like’ and to drop them from the calculation. We would
like to carry these answers over to important solid-state
5. Correlations in f-shell metals cases, such as f-shell metals, where we know that electron
correlations become central.
The principal defect of the LDA arises from the The difficulty is that we cannot extend this exact solution
introduction of self-direct and self-exchange interactions, to the state of many electrons. On the other hand, it is often
as in Hartree– Fock Theory (e.g. Ref. 3, 593ff). In Hartree – fruitful even within LDA to simplify the electronic structure
Fock these cancel exactly and provide a major simplification to something like a two-level system. One step in the
in yielding the same Hamiltonian for every electron state. In calculation of the total energy in any system is the
LDA this very large self-interaction is added to the much summation of the energies of occupied electronic states.
smaller exchange between different orbitals and approxi- This can also be done in a Friedel approximation in which
mated by a function of electron density. When Coulomb the density of states for the f-bands is replaced by a
effects are large, this makes the LDA results very inaccurate. rectangular distribution of 14 states per atom. Then the only
Modifying the density functional, or adding gradient parameter needed is the width Wf in energy of the
corrections, does not even address the question. Making distribution. If we are willing to make that approximation
self-interaction corrections [15] can rectify the errors in in the context of LDA, the energy gain per atom due to
principle, while retaining the more accurate aspects of the formation of f-bands for the Zf electrons per atom is
LDA but we prefer to proceed with tight-binding theory simply [3]
which makes it possible to omit the self interactions from
the outset. Wf
Etot ¼ 2Zf ð14 2 Zf Þ ð4Þ
A familiar case where Coulomb correlations become 28
dominant is directly describable in tight-binding theory. It is
a pair of atoms, such as lithium, in which a single valence s- with Wf playing the same role that Vsss played in our two-
orbital of energy 1s on each atom is important. The two level example. By comparing results for the two limits we
orbitals are coupled by a Vsss which depends on the can see that the effect ofqffiffiffiffiffiffiffiffiffiffiffi
the Coulomb
ffi correlations is to
internuclear distance d and if the two valence electrons are replace Wf in Eq. (4) by Uf2 þ Wf2 2 Uf ; where Uf is the
placed on the same atom there is an excess energy. extra Coulomb energy from placing two electrons in an
atomic f-state rather than having one in an atomic s-state [3].
U p ¼ U 2 e2 =d: ð2Þ We have sacrificed some accuracy in the LDA aspects of the
We have actually stripped this system down to the point that electronic structure, but may have been able to make a very
we may find an exact solution of the correlated tight-binding realistic description of the central many-electron aspects.
problem. There are only six two-electron states, spin-up and We see that the correlations effectively broaden the bands,
spin-down on atom one, spin-up on atom one and spin-down but reduce their effect in holding the atoms together.
on atom two, etc. With reflection symmetry between atoms This has provided a novel view of the nature of f-shell
we may reduce the solution of the six-by-six Hamiltonian to metals [3]. The electron states are separated into a simple-
quadratic equations, leading to a total ground-state energy of metal part and an f-electron part, the latter given by Eq. (4)
(Ref. 3, p. 595) with the modified Wf. The simple-metal part of the energy is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi minimum at a large spacing and the f-electron part at a small
 spacing. In the rare earths and heavy actinides, the simple-
Up Up 2
1total ¼ 21s þ 2 þ4Vss2 s : ð3Þ metal terms lead to a large spacing and, at that spacing, Wf is
2 2
so small compared to Uf that the f-levels are indeed core-
This leads to the LDA result, 21s þ U p =2 þ 2Vsss (note like, with the occupied levels well below the Fermi energy
Vsss , 0) at very small spacing where V2sss is much greater and the empty levels Uf higher, well above the Fermi level.
than U p2, including the U p/2 from the 50% probability of The resulting states with local moments are a result of the
both bond electrons lying on the same atom. This result also large spacing, not the cause of it. In the light actinides, Wf
properly segregates the electrons as 21s at large spacing. becomes so large that the partially filled f-band contracts the
There is an additional repulsive interaction between the spacing just as the corresponding d-bands cause the high-
atoms arising from the nonorthogonality of s-states on density structures of the transition metals, only slightly
neighboring atoms [3] which we need not consider here. weakened by the electron correlations. In between these
This two-atom result carries some important qualitative extremes, in plutonium for the actinides and cerium for the
and quantitative messages. We note that the force between rare earths, the two minima are distinct and competitive.
the two atoms, the negative of the derivative of energy with Both high-density and low-density structures exist depend-
respect to d, is reduced at all spacings, but only partially, and ing upon pressure and temperature, and the properties of
it does not drop discontinuously. We would not do so badly each form follow from the density which occurs.
W.A. Harrison / Solid State Communications 124 (2002) 443–447 447

6. Conclusion point of view which contains the essential truth, but is


simple enough for our minds to encompass.
We return to our starting question: why tight-binding
theory? Here it replaced the discontinuity in the energy-
dependent exchange and correlation potential by a U/1 from References
which we could estimate the gap enhancement for any
system. It is approximately correct to use that enhanced gap [1] W.A. Harrison, S. Ciraci, Phys. Rev. B 10 (1974) 1516.
in calculating m p with the k·p formula, but the unenhanced [2] F. Bloch, Z. Physik 52 (1928) 555.
gap is better for calculating the dielectric susceptibility. It [3] W.A. Harrison, Elementary Electronic Structure, World
aligned the two sets of bands at an interface by matching Scientific, Singapore, 1999.
hybrid energies. If the effective mass changes across an [4] W.A. Harrison, Electronic Structure and the Properties of
interface, we should share the discontinuity as (1/m p)1/2 Solids, W.H. Freeman, San Francisco, 1980, reprinted by
between the value and the slope, though this gives quantum- Dover, New York, 1989.
well states which appear to violate the uncertainty principle. [5] L.J. Sham, M. Schlüter, Phys. Rev. Lett. 51 (1983) 1888.
The tight-binding view of the f-shell metals was simple [6] J.P. Perdew, M. Levy, Phys. Rev. Lett. 51 (1983) 1884.
[7] L.C. Lew Yan Voon, W.A. Harrison, submitted for publi-
enough that we could make an exact calculation of the
cation.
effects of Coulomb correlations within that representation. It [8] L.C. Lew Yan Voon, M. Willatzen, M. Cardona, N.E.
gave us a new view of the nature of f-shell systems, quite Christensen, Phys. Rev. B 53 (1996) 10703 and references
different from the prevalent view that electron localization therein.
removes the f-shell electrons from the bonding role, [9] W.A. Harrison, J. Tersoff, J. Vac. Sci. Technol. B4 (1986)
producing the large atomic volumes of the rare earths and 1068 discussed also in Ref. 3.
heavy actinides. We can even understand why imposing that [10] W.A. Harrison, Phys. Rev. 123 (1961) 85.
prevalent assumption on an LDA computational study will [11] D.J. BenDaniel, C.B. Duke, Phys. Rev. 152 (1966) 683.
give essentially the right volume, even if for the wrong [12] G. Bastard, Wave Mechanics Applied to Semiconductor
reason. Heterostructures, Halsted Press, New York, 1988.
[13] W.A. Harrison, A. Kozlov, Matching conditions in effective
If it were possible to fully incorporate Coulomb
mass theory, in: P. Jiang, H.-Z. Zheng (Eds.), Proceedings of
correlations in an otherwise LDA approach, it would
the 21st International Conference on the Physics of Semi-
certainly lead to the correct volumes for f-shell metals, but conductors, Beijing, China, August, 1992, vol. 1, World
it is not clear it would tell us why-any more than experiment Scientific, Singapore, 1992, p. 341.
does. It would also give the correct gap enhancement, m p, [14] L.C. Lew Yan Voon, W.A. Harrison, in preparation.
susceptibility, and quantum-well states, but not the relations [15] R.D. Cowan, Phys. Rev. 163 (1967) 54 and a number of more
between them. Perhaps the meaning of ‘understanding’ is a recent references given in Ref. 3, p. 598.

You might also like