E le c tr o n c o rr e la tio n s in n a rro w e n e rg y b a n d s

B y J. H ubbard

Theoretical Physics Division, A.EHarwell, Didco

{Communicated by B. H. Flowers, — Received 23 April 1963)

I t is p o in ted o u t th a t one of th e m ain effects of correlation p henom ena in d - a n d /-b a n d s is

to give rise to b eh av io u r ch aracteristic of th e atom ic or H e itle r-L o n d o n m odel. To in ­
vestig ate th is situ a tio n a sim ple, a p p ro x im ate m odel for th e in te ra c tio n o f electrons in
narrow energy b an d s is in tro d u ced . T he resu lts of applying th e H a rtre e -F o c k a p p ro x im atio n
to th is m odel are exam ined. U sing a G reen fu n ctio n tech n iq u e a n a p p ro x im ate solution of
th e correlation problem for th is m odel is ob tain ed . T his solution h as th e p ro p e rty of reducing
to th e ex act atom ic solution in th e a p p ro p ria te lim it a n d to th e o rd in a ry u n co rrelated b a n d
pictu re in th e opposite lim it. T he co ndition for ferrom agnetism of th is solution is discussed.
To clarify th e physical m eaning of th e solution a tw o-electron exam ple is exam ined.

1. I n t r o d u c t io n
In recent years much attention has been given to the theory of correlation effects
in the free electron gas (Bohm & Pines 1953; Gell-Mann & Brueckner 1957; Sawada,
Brueckner, Fukuda & Brout 1957; H ubbard 1957, 1958; Pines & Nozieres 1958).
Apart from the intrinsic interest of this problem, the free electron gas serves as
a model for the conduction bands of metals and alloys. Transition and rare-earth
metals have in addition to their conduction bands partly filled d- or/-bands which
give rise to the characteristic properties of these metals. Correlation phenomena
are of great importance in determining the properties of these narrow energy bands,
indeed more important than the corresponding effects in conduction bands. Un­
fortunately, however, the free-electron gas does not provide a good model for these
bands. Rather, one requires a theory of correlations which takes into account
adequately the atomistic nature of the solid. Indeed, in the case of the/-electrons
of rare earth metals it is probable th at for most purposes a purely atomic (sometimes
referred to as a Heitler-London or localized) model will prove satisfactory. The
same cannot be said, however, of the (/-electrons of transition metals. I t is with one
approach to a theory of correlation effects in the (/-bands of transition metals th at
this paper is concerned.
A theory of correlation effects in narrow energy bands is inevitably of a somewhat
different nature from a theory of correlation effects in the free electron gas. The
electron charge density in a (/-band is concentrated near the nuclei of the solid and
sparse between the atoms, making it possible to speak with some meaning of an
electron being ‘on ’ a particular atom. This circumstance gives rise to the possibility
of an atomic description of the (/-band despite its considerable bandwidth. I t is,
in fact, found experimentally th at the (/-electrons of transition metals exhibit
behaviour characteristic of both the ordinary band model and the atomic model.
For example, the occurrence of spin-wave phenomena in ferromagnetic metals
and the strong temperature dependence of the susceptibilities of some transition
metals represent properties which can be understood on the basis of an atomic
[ 238 ]
lectron correlations in narrow energy bands
239
model, while the large d-electron contribution to the low temperature
and the occurrence in ferromagnets of magnetic moments per atom which are far
from integral numbers of Bohr magnetons are properties which are easily explained
by band theory. As will be tried to make plausible below, it is correlation effects
in narrow bands which lead to the atomic behaviour and it is only by taking correla­
tion effects into account th at one can understand how (/-electrons exhibit both
kinds of behaviour simultaneously. Thus a theory of correlations in (/-bands will
be mainly concerned with understanding this situation in greater detail and
determining the balance between bandlike and atomic-like behaviour.
In its most naive form the atomic theory would picture a transition metal as a
collection of (say singly charged) ions immersed in the conduction electron gas and
interacting with each other in much the same way as the corresponding ions in
salts. If, as is generally supposed, the number of (/-electrons per atom is non­
integral this simple picture is untenable. However, it is possible to substitute for it
a less restrictive model which nevertheless guarantees most of the characteristic
properties of the atomic model. I t is sufficient to assert that, despite the band
motion of the (/-electrons, the electrons on any atom are strongly correlated with
each other but only weakly with electrons on other atoms; such intra-atomic
correlations are inevitably of such a type as to make the metal behave to some extent
according to the predictions of the atomic model.
I t may be th at this situation can be made clear by considering one or two ex­
amples. Consider first a partly filled (/-band of non-interacting electrons. In such
a system the spin of an atom (that is the total spin of all the electrons on th at atom)
is a quantity winch fluctuates randomly in magnitude and direction, the character­
istic time of fluctuation being of the order of the (/-electron hopping time, i.e. the
time ( ~ fo/A,A — (/-electron bandwidth) in which a (/-electron hops from one atom
to another in performing its band motion. In this situation it is reasonable to think
of the spin being associated with each of the moving (/-electrons.
Let us now inquire what effect one might expect the electron interaction will
have in this situation. As a guide one may note th at H und’s first rule for atoms in­
dicates that the intra-atomic interactions are of such a nature as to aline the
electron spins on an atom, so one may expect a similar effect in a metal. Suppose
now th at the electrons have their spins quantized in what will be called the up and
down directions and th at at some instant a given atom has its total spin in the up
direction. Then the intra-atomic interactions are, according to H und’s rule, of such
a nature that this atom tends to attract electrons with spin up and repel those with
spin down. In this way the property of an atom on having total spin at some instant
tends to be self-perpetuating. If these intra-atomic forces are strong enough to
produce appreciable correlations, then it follows that the state of total spin up
on an atom may persist for a period long compared with the (/-electron hopping
time. This persistence of the atomic spin state is not due to the same up-spin
electrons being localized on the atom. The actual electrons on the atom are always
changing as a result of their band motion, but the electron motions are correlated
in such a way as to keep a preponderance of up-spin electrons on the atom. In these
circumstances (i.e. if the correlations are strong enough) one can think of the spin
240 J . H ubbard
as being associated with the atom rather than with the electrons and the possibility
of an atomic or Heisenberg model emerges.
This example illustrates the possibilities of the situation. Although one may still
suppose the electrons to move rapidly from atom to atom as assumed in the band
model, their motion may be correlated in such a manner as to give properties
characteristic of the atomic theory. In this way one may understand how the
electrons can exhibit both types of behaviour simultaneously. The degree of atomic
behaviour exhibited depends upon the strength of the correlations.
A second example which has been studied by various authors (Slater 1937;
Herring 1952; Thompson i960; Edwards 1962; Kubo, Izuyama & Kim 1962) is
the theory of spin-waves in the band model of ferromagnetic metals. These authors
show that the spin-wave can be regarded as a collective motion which appears
when the electron interactions are taken into account. More precisely, the spin-
wave appears as a bound state of an electron of one spin with a hole of opposite
spin, the relative motion of the electron and hole being such th at they spend most
of their time on the same atom. Now, an electron of one spin and a hole of opposite
spin on the same atom look just like a reversed spin on th at atom, the motion of the
bound electron-hole pair resembling a motion of the reversed spin from atom to
atom, which is just the Heisenberg model picture of a spin-wave. Thus again the
atomic picture emerges as a consequence of correlation effects, this time the corre­
lation between an electron and a hole.
Yet another important example concerns the fluctuation in the number of elec­
trons on a given atom. I t is, of course, one of the more obvious features of the
atomic model th at it assumes th at there are the same number of electrons on each
atom. But one can show that for uncorrelated electrons belonging to a band con­
taining v states per atom th at the probability of finding n electrons on a given atom
is given by the binomial distribution

where s is the mean number of electrons per atom. Thus n fluctuates about its mean
value s, the root-mean-square fluctuation being ^{5(1 —s/v)} and the frequency of
fluctuation of the order of an electron hopping time. Now one general effect of
electrostatic interactions is a tendency to even out the electron charge distribution,
opposing the build-up of an excess of charge in one place and a deficiency in another.
Thus the correlations produced by the interaction will be of such a nature as to
reduce the fluctuation in the electron number on each atom. I t is this type of corre­
lation which is most important in the hypothetical case of a collection of atoms
arranged on a lattice but widely separated from each other. Formally ordinary
band theory is applicable to such a situation, but the correlation effects of the type
discussed above are dominant and make the system behave like a set of isolated
neutral atoms, which is clearly the correct description physically.
I t is clear from the above discussion th at an important requirement of a theory of
correlations in narrow energy bands is that it have the property of reducing to the
atomic solution in the appropriate limit, i.e. when applied to a hypothetical system
 Electron correlations in narrow energy bands 241
of atoms on a lattice but widely separated from each other and interacting only
weakly. I t is one of the purposes of this paper to describe a very simple, approxi­
mate theory having this property. Although one has always in mind the case of
d-electrons, the theory to be described is concerned with the case of an 5-band
having two states per atom (up and down spin states). The advantage of this par­
ticular case is its comparative mathematical simplicity. One may expect th at some
im portant aspects of the real (d-electron) case will be missed in a study of the 5-band
case but may nevertheless hope to obtain some results of general application.
I t might seem th at in view of the fact that no adequate theory of correlations in
free electron gases a t metallic densities exists at the present time th at it is over-
ambitious to attem pt a study of the formally more difficult case of band electrons.
However, it turns out th at in the case of narrow energy bands one can take account
of the atomicity of the electron distribution to introduce a very simple approximate
representation of the electron interactions. This approximate interaction is, in
fact, mathematically much simpler to handle than the Coulomb interaction itself.
This possibility has been well known for many years and has been applied to the
spin-wave problem by the authors mentioned in th at connexion above, but does
not seem to have been exploited hitherto in connexion with the general correlation
problem. In § 2 this approximate interaction and the adequacy of the approxima­
tion involved is discussed.
For the sake of comparison with the results of the theory of correlations developed
later, in § 3 the application of the Hartree-Fock approximation to the simplified
interaction is considered and in particular the condition for ferromagnetism
predicted by Hartree-Fock theory is examined.
In §§ 5 and 6 the approximate correlation theory for an 5-band mentioned above
is developed. To this end a Green function technique of the type described by
Zubarev (i960) is used; to establish the notation the basic definitions and equations
of this technique are briefly reviewed in § 4. In § 5 it is shown how, using this tech­
nique, an exact solution can be obtained in the atomic (zero bandwidth) limit. In
§ 6 the same method is applied to the general (finite bandwidth) case to obtain the
approximate solution. In § 7 the nature and some of the properties of this solution
are discussed.
In § 8 we examine a 2-electron problem which has been studied previously in a
related context (Slater, Statz & Koster 1953) and which throws some light upon the
physical interpretation of the solution obtained in the preceding sections.
Finally in § 9 the condition for ferromagnetism predicted by the new calculation
is discussed. I t is found to be considerably more restrictive than the corresponding
criterion derived from Hartree-Fock theory, and, in fact, can only be satisfied in
rather special circumstances.

2. An a p p r o x im a t e r e p r e s e n t a t io n o f e l e c t r o n in t e r a c t io n s

In this section the approximate model of electron interactions in narrow energy

bands used in later calculations is described. As pointed out in the introduction,
for reasons of mathematical simplicity the case of an 5-band will be considered.
However, when discussing below the validity of the various approximations which
242 J . H ubbard
have gone into the derivation of the model we shall assume we are dealing with
3d-transition metal electrons since this is the case of real interest.
Consider a hypothetical partly-filled narrow 5-band containing n electrons per
atom. The Bloch functions of the band will be denoted by ij/k and the corresponding
energy by ek where k is the wave vector. These wave functions and energies are
assumed to have been calculated in some appropriate Hartree-Fock potential
representing the average interaction of the 5-band electrons with the electrons of
other bands and the n electrons per atom of the 5-band itself. This Hartree-Fock
potential will be assumed to be spin independent so one has the same energies and
wave-functions for both spins.
Now let ck(T,ckfT be the destruction and creation operators for electrons in
Bloch state (k, cr), where cr = + 1 is the spin label. Then the dynamics of the elec­
trons of the band may be described approximately by the Hamiltonian

# =2
Zj ^kcr^ka
kcr

+i 2 2 (kxk 2| l/r |k ; k ') 4 1(Tl4 2(72ck^ 2ckl

ki k2k' k'2 orcr2

- S 2{2(kk'( l/r |k k ') - ( k k '[ l/r |k'k)} c„„ (i)

kk' cr

where the k sums run over the first Brillouin zone (all sums over momenta in this
paper are to be understood in this way) and where

(kl k , | l/r | k jk y = «*J S W A i W ^ O M O d x d x ,. (2)

The first term of H represents the ordinary band energies of the electrons, the
second their interaction energy. The last term subtracts the potential energy of the
electrons in th at part of the H artree-Fock field arising from the electrons of the
5-band itself. This term has to be subtracted off to avoid counting the interactions
of the electrons of the band twice, once explicitly in the Hamiltonian and also
implicitly through the Hartree-Fock field determining the ek. The are the assumed
occupation numbers of the states of the band in the H artree-Fock calculation; it
has been assumed th at up and down spin states are occupied equally.
I t is convenient to transform the Hamiltonian of (1) by introducing the Wannier
functions
k(x ), (3)

where N is the number of atoms. One can then write

^ k(x) = N ~i 2 e**-®*0(x —R J, (4)

where the sum runs over all the atomic positions R/. Introducing the creation and
destruction operators c\a and cirrfor an electron of spin crin the orbital state <^>(x —R,:),
one can also write
ck(T = N~* 2 eik-R<ci(T, = JV-i 2 e_ik Ri 4<x- (5)
i i
 Electron correlations narrow energy bands 243
These results can now be used to rewrite the Hamiltonian of (1) as
H = 2 S
i , j cr ijk l O'er'

- U S {2 (ijl/r\
| kl) - l/r | (6)
ij k l or

where T{j = (7)

(ij\ l/r |kl) = ( 8)

and vfl= iV-1 ^ rkeik-(Rj-R/). (9 )

k
I t is now possible to make the essential simplifying approximation. Since one is
dealing with a narrow energy band the Wannier functions ^ will closely resemble
atomic 5-functions. Furthermore, if the bandwidth is to be small these 5-functions
must form an atomic shell which has a radius small compared with the inter-atomic
spacing. From (8) it may be seen th at in this circumstance the integral (m | l/r =I
will be much greater in magnitude than any of the other integrals (8), suggesting
th at a possible approximation is to neglect all the integrals (8) apart from I. If
this approximation, the validity of which is discussed in greater detail below, is
made, then the Hamiltonian of (6) becomes
Tjj 4<r Cja+ ¥ 2 ni*ni, -c r -I 2
i,j or i,cr i, or
m (9), vu = N~x ^ i'k = \n, so the last term of (10) reduces
k
to — \ I n Yinicr— —\I N n 2 = constant and may be dropped. Equation (10) gives
i, or
the approximate Hamiltonian used in the later sections of this paper.
Obviously many approximations, explicit and implicit, have gone into the
derivation of the simplified Hamiltonian of (10). We will next try to assess the
validity of these approximations when applied to the case of transition metal
3d-electrons.
The most obvious approximation has been the neglect of all the interaction terms
in (6) other than the (i|l/r | i)et rm. For the sa
I has the order of magnitude 20 eV for 3^-electrons in transition metals. The largest
of the neglected terms are those of the type (ij\ l/r | where i and j are nearest
neighbours. From (9) these integrals can be estimated to have the order of magnitude
(2/12) Ry ~ 6 eV (R — interatomic spacing in Bohr units). Actually this figure
should be reduced appreciably to allow for the screening of the interactions of
electrons on different atoms by the conduction electron gas. This screening effect
may be allowed for approximately by multiplying the above estimate by a factor
e-KR where k is an appropriate screening constant. In the case of 3d transition
metals e~Kli ~ reducing the ji( |l/r | term to the o
3 eV. For the case in which i and j are now nearest neighbours
9q— k|R/—
Rj I
(ij\ l/r | ij) ~ Ry
244 J . H ubbard
which falls off rapidly with increasing |R^ —R?-| on account of the exponential
factor. Thus the term , £ ^ ^ 1/r {ij)n^ (n )
i, j a, a-'

in (6) is quite appreciable, but can, perhaps be neglected compared to / as a first

approximation.
The next biggest terms neglected are those of the types:
(m | 1/r | ij) R
q~ y ~ |eV,
('y I l lr\ik) ~ ig R y ~ xoeV>
(ii\ l/r | j j )~ (ij\ 1/r | )ij ~g 2Ry ~ xVeV,
where i, j and k are all nearest neighbours and q ~ is the overlap c
of e) between two 3d-electrons on nearest neighbour atoms. All the other inter­
action terms in (6) which have been neglected are smaller still than these which
one sees are already small compared to those of (11).
A different type of approximation th at has been made is to assume th at only the
interactions of importance are those between the 3d-electrons (actually between the
electrons of the s-band in the equations above), the interactions with electrons of
other bands being represented only through the Hartree-Fock field. One question
concerning this point is raised at once by the fact th at in estimating the order of
magnitude of the terms of (11) allowance was made for the screening effect of the
conduction electron gas on the interactions. I t might therefore be inquired whether
there is not a similar screening effect reducing the magnitude of I. There is, in fact,
such an effect. Because the speed at which d-electrons move from atom to atom is
slow compared with the velocity of a typical conduction electron the latter can
correlate efficiently with the d-electrons and screen their fields. Thus, if a given atom
has an extra d-electron its negative charge will repel conduction electrons producing
a correlation hole about th at atom in the conduction electron gas. The presence of
this correlation hole reduces the electrostatic potential at the atom (and therefore
at each of its d-electrons) by about 5 V, which is equivalent to reducing by 5 eV.
This reduction is appreciable but does not change the order of magnitude of
I t might also be thought th at I will be reduced by the screening of the interactions
of the d-electrons by the core electrons and by the d-electrons themselves. This is
not expected to be a big effect, however, because the kinetic energies of the orbital
motion of the d-electrons are large compared to I. In fact, one may estimate the
reduction in I due to this effect by noticing th at a similar effect will occur in free
atoms. In the case of free atoms it has been found th at these effects make the
F 2(3d, 3 )dnad F 4(3d, 3d) parameters (using the notation of Condon & Shortley
193 5) determined from experiment about 10 to 20 % smaller than those calculated
from Hartree-Fock wave functions (see Watson i960) so one may expect a reduction
in I of a similar order of magnitude.
I t would seem from the above discussion, although it may be more realistic to
use in the Hamiltonian of (10) an ‘effective’ I ( ~ 10eV) rather than th at given by
the integral (8), the approximations involved in (10) are otherwise not so poor as
to make it an unreasonable starting-point for a theory of correlations when suitably
 Electron correlations in narroiv energy bands 245
generalized from the 5-band to the d-band case. I t may, perhaps, be hoped th at the
terms omitted in going from (6) to (10) may be treated as perturbations on solutions
obtained from (10).

3. T h e H a r t r e e - F o c k a p p r o x im a t io n

For the sake of comparison with the results of the correlation theory developed
in later sections it is convenient now to investigate the results obtained by applying
the H artree-Fock approximation to the Hamiltonian of (10). Actually, we shall
not make an exhaustive study of all possible H artree-Fock solutions, but will
restrict attention to a particularly simple class of solutions which may represent
non-magnetic or ferromagnetic states but not more complicated spin arrangements.
A similar restriction applies also to the correlated solutions investigated in later
sections.
As is well known, one may obtain the effective Hartree-Fock Hamiltonian by
‘linearizing’ the interaction terms in the true Hamiltonian. In the case of the
Hamiltonian of (10) this amounts to simply replacing the term _a by
n i<r(n i , - a ') + n i >- c r ( n i<r) where (nia) is the average of the expectation of ni(T over
a canonical ensemble at some temperature 0. Dropping the last term of (10) which
has been shown to be a constant, the Hartree-Fock Hamiltonian is found to be
Hhf = 2 2 V U v + 7 2 (12)
i 9j or i,cT

Attention will now be restricted to the class of solutions for which

(nia) = n0. for alH. (13)
Then (12) becomes
Hhf = S S E j c\<tcjo- + -^2 n_ac\aci(T, (14)
i , j or icr

or, transforming back to the operators cla, ck(T

E fif 2 S {^k 0-} Cfco-, (I 'd
k cr

which is simply the Hamiltonian for a collection of non-interacting electrons with

a slightly modified band structure, the energy of the (k, ) state now being ek + In_a.
I t follows at once that if P(E) is the density of states per atom corresponding to
the band structure ek, then the densities of states Pa(E), where = + 1, for the
electrons described by the Hamiltonian of (15) are
P(r(E) = -E
(P In_a)= P(E I
where the last step follows from
= n. (17)
If fi is the chemical potential of the electrons, then at the absolute zero of tem­
perature one will have
[/l Pt7(E )d E = rP iE -In +
J —OO J —00
The pair of equations (18) must now be solved together with (17) for n ^ ,n ^ and ju.

16 Vol. 276. A.
246 J . H ubbard
One possible solution of ( 18) is that for which
= raj, = \n, ( 19)
which represents a non-magnetic state of the system: ft is determined by

\n= f P(E —\In) dE.

J — 00

If I is sufficiently large it may also be possible to find ferromagnetic solutions

for which =}=Wj,. In this case equation ( 18) must have two distinct solutions which
are such that they can satisfy ( 17). The condition th at ferromagnetism is just
possible can now easily be seen to be the condition th at ( 19) and (20) give a double
solution of ( 18). But this condition can at once be found from ( 18) to be
1= IP {fi—\Iri).(21)
Thus, if for any E the condition IP(E) > 1 is satisfied, then for some and f i
determined by (20) and (21) Hartree-Fock theory predicts th a t the system will
become ferromagnetic. I t will be found th at when correlation effects are taken into
account one obtains a somewhat more restrictive condition for ferromagnetism.

4 . A Green function technique

In the next two sections an approximate solution of the correlation problem for
the Hamiltonian of (10) is derived. The method of calculation is based upon the
Green function technique described by Zubarev (i960). In order to establish the
notation, the principal definitions and basic equations of this technique are briefly
reviewed in this section.
Let X be any operator. Then define
<X> = Z -1tr {X z = tr (e-tf*-/**)}, (22)
where PI is the Hamiltonian and N the total number operator, /? = 1/x 0 , k = Boltz­
m ann’s constant, 0 = absolute temperature and = chemical potential of the
electrons.
Now let A(t) = eiflr<H(0) e~iHt (in units in which 1) and B(t') be two operator
Then retarded (+ ) and advanced ( —) Green functions may be defined by
<(A(t); S((')»<±> = + i0 {± ((-*')} (23 )
where [A, IP] = A B —tjBA, tj =± 1 (whichever is the more co
is the step function 6{x) =1 if x 0>, = 0 o
shown to satisfy the equation of motion

i W ) ) > w = # ( t - t ' K [ A ( t ) , B ( % ) + (([A(t),H]; j 3(f')»<±>. (24)

Since « A (0 ; ./?(0 )>(±) are functions of t —t' only, one can define for real E the
Fourier transforms
1 f 00
((A; B))& = — ((B(0)))<±>e‘^
(25)d t .
 Electron correlations in narrow energy bands 247
In the case of the retarded ( + ) function the integral (25) converges also for
complex E provided J>E>0, so A
( B')')(
; j^
of E in the upper half of the complex .27-plane. Similarly, ((A; l ? ) ) ^ is a regular
function in the lower half of the complex .27-plane. One may now define
((A ; B } \ = ((A; if 0,
if J E < 0, (26)
which will be a function regular throughout the whole complex .27-plane except on
the real axis. From (24) it can be shown th at « B}}E satisfies
E { { A - B }) e = 1l2n{[A,B]v) + {([A,H]; B ))b . (27)
I t can be shown (Zubarev i960) that

(B(f)A(t)) = i Bm [« A B )) b +u28)
(

Equations (27) and (28) together with the method of approximation described by
Zubarev (i960) form the essential basis of calculations with these Green functions.

5. T h e e x a c t s o l u t io n i n t h e a t o m ic l im it

In this section the application of the technique described in the preceding section
to the Hamiltonian (10) in the limiting case of zero bandwidth is discussed. This
limit corresponds to the situation in which the wave functions on different atoms
have only a negligible overlap, in which case one knows th at the atomic theory gives
the exact solution. I t will be shown th at in this case the Green function technique
also leads to an exact solution. Of course, for the Hamiltonian (10) the results of
these calculations are rather trivial, but they do serve to reveal the essential trick
required to make more elaborate theories go over into the exact atomic solution
in the appropriate limit.
All effort will be concentrated on obtaining an expression for the Green’s function.
(29)
since, as is well known, a knowledge of this Green function enables one to calculate
pseudo-particle energies, the Fermi energy, free energy, etc. For example, sub­
stituting (29) into (28), putting j = Jc, t —t' =0 and summ
the mean number na of electrons per atom of spin cr the expression
na = N - 12 (c}acjo.)
j
* /* dF
- ^ Mm S j [<%(Jg + ie) - Gl( E - ie)]+ 1 (3 0 )

from which one may infer that

p,(E) = L lim 2 [G%(E + - Oft(E - ie)] (31)

iv e-M)+ j
gives the density of (pseudo-particle) states per atom of spin cr.
Define T0 = A -12 ek; (3 2 )
k
16-2
248 J . H ubbard
T0is the mean band energy. In the limit of zero bandwidth ek = T0 for all k, whence
it follows that = T0 and the Hamiltonian (10) becomes
H S n i<T +
i,cr i,cr
from which it follows that
[ci(T>H ] = T0ci<r+ Ici<rnin<
-,(34)
so equation (27 ) gives
m m = ^ S ii + T0G^(E) + i r - j(E),

where ^(E ) = _aci(T\c p ))^ =

Now, [»<,_trc<(r,cJ0.]+ = 8tjni'_(r-, also from (33) [nia,II] = 0, so (27 ) gives for
T ij(E) the equation
ET%j(E) = — dy (ni,-(r}+ To + 1 <Xc](T')')e

At this point one may notice th at since n% — the last term of (37) is just IT ^ ,
so (37) can be solved at once to give
n m = 2n ijE - T 0- I ' (38)

The usual infinite sequence of equations of the type (27) involving higher and
higher order Green functions has been avoided here by the simple observation
th a t the last term of (37) can be expressed in terms of T?-. I t is this possibility th at
enables one to obtain an exact solution in the present case and solutions reducing
to the correct solution in the zero bandwidth limit in the general case discussed in
the next section.
It follows from the definition (22) and the symmetry of the problem th at (ni(T)
is independent of i and a, so one has at once (nia) \n . Using this result and
substituting (38) into (35) one obtains

<L f1
oW

(39)
J—1>

GUE)
+
t?3

which gives at once from (31)

PAE ) = (! To) + \n8(E —Tq— (40)
Thus the calculation shows that the system behaves as though it has two energy
levels T() and T0+ I containing 1 —\n and \n states per atom re
electrons are added to the band, initially the Fermi-energy will be fixed a t fi = T0
whilst the lower level fills up. The lower level will become full when = 1 —\n ,
i.e. n = 1, and the chemical potential y then jumps to y = T0+1 whilst the remain­
ing electrons are added. This is just the correct result. As electrons are added they
will (at 0 = 0) distribute themselves on different atoms giving y = until when
n —1 this is no longer possible and any further electrons added have to go on atoms
which have already one electron so y jumps to T0+ I. A similar discussion can be
given for other properties such as specific heat, behaviour in a magnetic field, at
0 = 0 and at finite temperatures; in all cases the Green function solution yields
the correct result for a collection of isolated atoms.
 Electron correlations in narrow energy bands 249

6. A n a p p r o x im a t e s o l u t io n o f t h e c o r r e l a t io n p r o b l e m

In this section the correlation problem for the Hamiltonian (10) will be studied
in the finite bandwidth case by the same technique used in the last section. In the
present calculation, however, certain additional terms appear which have to be
treated approximately to obtain a solution.
Returning to the Hamiltonian (10), one finds
[Cicr>E] = 2 C+ In it_
i
[nicr’E] = 2 ^u(c! <TCjtr(42)
1

so the equation for Gy defined by (29) becomes

EGUE ) = ^ % + S Tik Gl, + 1T%„ (43)

where T - is again defined by (36), but now satisfies the equation

ET 'ij(E)= ^ (ni, -a) + % Y?j + I T -

+ 2 Ek -O-C'ka; C1'o-))e
k+i
+ S c]-a.')}E —c]<t)).e}' (44)
k

The term jhas been obtained by using

%
IF — ni,-a’ while the term T? j
has been obtained by separating out the = part of the fourth term. The first
three terms are identical with those of (37) while the latter pair of terms vanish in
the zero band-width limit. Thus, whatever approximations are made in the last
pair of terms of (44), one will obtain a theory th at goes over into the exact solution
in the zero bandwidth limit.
In order to break off the sequence of Green function equations an approximate
expression will be substituted for the last pair of terms in (44). These approximations
are obtained by the methods indicated by Zubarev, and are given by

C^ » E - (45)
-<rCk,-<TCic
r>Jct) ) e — ( CI,
< (4 6

( ( Ck-< rCi,- o - Cicr’ C/< r ) ) s — ( c k -< r Y 'i,-c r ')@ ij(E )- (4 7 )

By making these approximations one obtains what is practically the crudest

theory possible consistent with the condition that it reduces to the correct zero
bandwidth limit. One shortcoming of the theory which arises from these approxi­
mations is pointed out in § 8.
Other important physical effects neglected as a consequence of these approxi­
mations are associated with collective motions of the spin-wave type (see authors
cited in the Introduction) and zero-sound type (Landau 1957).
250 J . H ubbard
With the approximations (46), (47), the last term of (44) vanishes as a consequence
of translational symmetry, since

2 T i k ( Ci, - < r Ck , - ( f ) —
7c=H i, k

= -„ew > = Z T a <cl-„c{,_„),

i, k k=£i

where Tik =Tki which follows from ek = e_k has been used, and 2* means t
double sum with the term i —komitted.
In (45) we will now put <»,v> = » „ (48)

which follows from translational symmetry. Strictly speaking it follows from the
symmetry of the problem th at na is independent of cr and therefore equal to \n .
However, it would seem intuitively evident that when solutions of the equations
with 4= Wj, exist, these solutions are connected with the possible ferromagnetism
of the system (provided they have lower energy than the non-magnetic solution).
One can, perhaps, rationalize this situation by imagining th at a minute magnetic
field is applied to the system; this field destroys the symmetry between up and
down spin, but is so small th at it can be neglected in the calculations. To discuss
this situation in detail would go far beyond the scope and intent of the present paper.
I t may be remarked th at a similar situation exists in the zero bandwidth case dis­
cussed in the preceding section, but that nothing new would be found there by
considering solutions for which =)= • One might also inquire whether meaning
can be assigned to solutions (forbidden by symmetry) for which (nia) is not in­
dependent of ior even for which quantities like aCj do not vanish. The ans
would seem to be that these solutions correspond to the possibilities of antiferro­
magnetism, spiral spin arrangements, etc., but only solutions falling under (46)
will be investigated here.
Substituting the approximations (45) to (47) into (44) one obtains

= n_(T~_+ (Tq+ 1) r + n_a 2 (49)

Z7T k=¥i

whence TW = S w ) , (50)

which when substituted into (43) gives

EGU E) = W + | i + l^ r r ) g + S ^ < 4 <«>

This equation may be solved by Fourier transformation. Writing

Gij(E ) = N - 1£ Gff(q, E) exp [iq . (R* - R;)] (52)
q

and using (7), one obtains from (51)

E G ^ E ) = T„ffqq, E) + {l + ^ J + («, - T„) G*(q, J ) | , (53)

 Electron correlations in narrow energy bands 251
whence 1 E-
G"(q,E) (54)
2nN (E —e^) (E —T0—I) + neq) ’
which gives the approximate solution to the correlation problem which has been
sought. The properties of this solution are discussed in the next section.

7. P r o p e r t ie s o p t h e a p p r o x im a t e s o l u t io n

The general nature of the solution given by (54) will next be investigated. The
expression (54) for Ga(E)is a rational function of E and may be resolved
fractions according to
1 1 &*>-T0- I ( 1-»_„))
(55)
W’ ' 2*N E -E $ /’
where E $ < EfJ are the two roots of
{ E - e ^ { E - T Q- I ) + n_(TI(TQ -e^ = 0.
It can be shown th at E $ < T0+ I ( l —n_^)< so (56) has
1 l d (1) d (2> \
^ > = 2 + <57>
with ^4^, > 0. If one had A (^ = A (^ = 1, then the expression (57) would be
A
the Green function appropriate to a band structure having two bands with the
dispersion laws E = E^J and E — E (*J. The effect of the factors ^ cannot be given
any very simple interpretation beyond saying that they reduce the density of states
in each band in such a way th at the total number of states per atom in both bands
together is just 1 and not 2 as it would be if A $ = = 1. One may see this
directly by noticing that from (31) and (52) one has
pa(E) = i lim 'El{Ga(q,E + ie) —G<T(q,E —ie)}, (58)
e->0+ q

which gives on substitution of the expression (57) for G<T(q,E)

p„(E) = + (59)
q
and finally noting that A $ + A $ = l. (60)
The general form of the band structure E$given by
In the limit I-> 0, the lower E (*J curve goes over into A P X and at the same tim
A (*J -> 1 along A P and A ^ -> 0 along P X . Similarly, the upper c
into B P Y, A $ 1 along P Y. Thus as the interaction is switched off the two por­
tions ^4P and P Y combine to make up the unperturbed band structure, the other
parts disappearing. That Ga(E) goes over into the unperturbed Green function
(27 tN ) ~ 1 (E —eq)_1 as / 0 can also be seen directly from (54).
In the limit of zero bandwidth, eq -> T0, the E^J, curves become flat and go
into E (^J = T0, E$= T0+1, giving the two levels c
states respectively discussed in § 5, the expression (59) going over into that of (40)
after q-summation.
252 J . H ubbard
Next, a more explicit expression than (59) for will be derived. From (55 )
and (58) one can obtain after a little manipulation the formula
P<J(E) = \ E - T 0- I { l - n ^ ) \ N - ' X M E - e t ) { E - T o - I ) + I* -e(To-<k)'}
q
l*00
J -C O q
f*cc E ( E - T , - l ) + In ^ T „
d tP
J -00 . E - T 0- I ( l - n _ „ )
P{g(E, (61)

Brillouin xone
b o u n d a ry
F ig u r e 1. A sk e tch o f a ty p ic a l p se u d o -p a rticle b a n d s tru c tu re 9J, E ^ . A P Y gives
th e u n p e rtu rb e d b a n d s tru c tu re w hile B P X is th e line E = T 0 + I .

P n ^ l-n ^ )
where g(E,n_a) = (62)
E-T
and P(E) = N - 1'Ei 8 (E -e <l) (63)

is the density of states corresponding to the band structure ek.

Thus pJyE) is obtained from P(E) by the simple transformation (61 ), (62 ). This
transformation is illustrated graphically in figure 2 which shows a typical g(E)
curve and the projection of P(E) into P(r(E). In the limit 0 the curve g(E) goes
over into the straight line AOB. The splitting of the band into two parts is seen to
be due to the infinity of g(E) at E—
In order to obtain some feel for the properties of the solution, it is, perhaps,
useful to consider the simple example given by the ‘square’ density of states
formula P{E) = 1/A if T0- ± A < E < T 0+ ±A
= 0 otherwise, (64)
 Electron correlations in narrow energy bands 253
for which one easily finds
/V W = 1 /A if E LU_1 < E < E ° _ 1A
or if E l _x < E < E 1>lr
=0 otherwise, (65)
where (a,/3 — ±1)
E°/j — + -|7 + J/?A + a ^ '{(■£/ —|-/?A)2+ ( 66 )

One may note th at since {E_11- E _ 1_1) + (E1-l - E h _1) = A, the pa{E) band
contains jnst one state per atom as it must.

/ / f profile of P(E)
fcj.... / / g ( E ) _ e

p rofile o f p(E )

F ig u r e 2. A sk e tc h o f a ty p ic a l g (E ,n ) curve. T h e p ro jec tio n o f th e u n p e rtu rb e d d e n sity o f

s ta te s fu n ctio n P (E ) in to th e p se u d o -p a rticle d e n sity o f s ta te s p a(E) is in d icated .

Turning attention now to non-magnetic solutions for which =

(ferromagnetic solutions will be considered in § 9), the Fermi energy ji will be deter­
mined by the equation
(67)

Thus y increases as electrons are added until the lower band is just full. I t will
then jump discontinuously as further electrons are added and then continue to
increase smoothly until the whole band is full. If E max , I£min are the highest and
lowest energies of the band structure one can see at once from figure 2 that this jump
occurs from a value y < to a value y >at a density nc deter
max.
■^max. 9(P<i 2^c)’ ( 68 )

Emin. =9(P>,^c), (69)

K = f P{g(E,in

where in (68) the smaller and in (69) the larger root must be taken.
 254 J . H ubbard
In general nc 4= 1, but in the case of the density of states curve (64 ), the sym­
metry between electrons and holes or, to be more precise, the fact th at in this case
P(E) has the property P(2T0—E) — P(E), requires th at nc = 1,
tinuity in ft of „ _ u7
-2_l iA2\ _
P>- P< = V(/2 + i A2) - i A>
which goes to 0 as / 0 and tends to / as A -> 0.

8. A TW O-ELECTRO N EXAM PLE

In order to obtain a better physical understanding of the solution obtained in the

preceding sections, and in particular how the band splits into two parts, it is in­
structive to consider the problem of two electrons moving and interacting in the
manner described by the Hamiltonian (10). Essentially the same problem has been
considered by Slater et al. (1953) who were mainly interested in the effect of corre­
lations on the condition for ferromagnetism.
Denote by j) the spatial wave function of an eigenstate of the tw
system, |^(b j ) |2 measuring the probability of finding one electron on atom i and
the other on atom j.Of the 4 N p2ossible state
for which j) = —rjr(j, i), and the other are singlet states for which
= i).If j) is an eigenstate with energy E of the Hamiltonian (10),
Ef ri hj ) =j) + 2 )Tjkfr(i,Jc) +(71)
kk

Since for the triplet states = 0, the last term of (71) vanishes for these
states, so the triplet states are quite undisturbed by the interaction. This is simply
because the Hamiltonian (10) only contains interactions between electrons of
opposite spin. Thus attention can be restricted to the singlet states.
In the singlet case we now write
j) = N - 12 2 {k, K) exp {iK. (Rt. + R,.) +
J>
< . (R
K k

Substituting this into (71) and using (7) one obtains

E(j)(k,lL) = {%+jk + % _|t}^(k, K) + IAr_12 ^ ( k ,?K). (73)

Thus solutions with different ‘total momentum’ K are not coupled to each other,
a consequence of translational symmetry. From (73 ) one has at once
J A - i2 ^ ( k ',K )
0 (k, K) = __ r (74)
E ~
e K + |k — % -4 k

1
whence sk E ~ eK+£k ~ eK -
(75 )

gives the energy levels for a given K.

The nature of the solutions of equations of the type (75) are well known. The
equation has N roots. The right-hand side has infinities at the N energies given by

E ~ %+Ik + eK-|k (76)

Electron correlations in narrow energy bands
255
for the N values of k, so there are (N 1) roots trapp
—

These N —1 roots lie in the unperturbed energy band given by (76). There is one
other root. For large enough I this root is quite separate from the band (76), forming
a ‘bound’ state. When I is large compared to the width of the band (76), this root
is given by T0+ I as may easily be seen from (75). For small I this ‘bound’ state
does not separate from the band (76).
Thus for large I there are N (N —1)singlet ‘scattering’ s
perturbed band and N ‘bound’ states (one for each of the N values K) with high
energy. In the limit I-> oo the latter states disappear altogether. Th
of an ‘excluded’ volume effect of the type familiar from van der Waals’s equation.
When I-> oo no two electrons can be on the same site. Thus if one electron is already
present (in any one of its N possible states) and another electron is added, then there
are only 2V—1 states available to this second electron, whence it follows th at there
are only N (N —1) possible states available to the two electron system rather than
the N2possible states for a pair of non-interacting electrons. When I is large but
finite the remaining N states reappear with high energy. One may now surmise
th at when m electrons are already present then only N —m states are available (in
the limit I ^oo) to any further electron added to the system, the remaining
states reappearing with high energy when I is finite but large. In this way one can
understand how the two bands of figure 1 arise. The lower band is essentially the
unperturbed band with some states excluded, these states reappearing in the upper
band.
This example reveals a weakness of the approximate solution of § 6. The discussion
given above only applies when 1 is sufficiently large for the ‘bound ’states to separate,
but the solution of §6 gives a splitting into two bands for all non-zero I. Obviously
the approximation is over-estimating the importance of correlation effects for small
I, presumably as a consequence of the drastic approximations of equations (45)
to (47).

9. T h e c o n d it io n f o r f e r r o m a g n e t is m

In §3 the condition for ferromagnetism predicted by Hartree-Fock theory

was considered. Here the way in which this condition is affected when corre­
lation effects are taken into account (in the approximation of §6) will be
examined.
One expects the condition for ferromagnetism to be more restrictive in a theory
which takes into account correlation effects than in Hartree-Fock theory. The
reason is simply th at ferromagnetism occurs when the (free) energy of the ferro-
magnetically alined state is less than th at of the non-magnetic state. Now, when
correlation effects are taken into account it is mainly the correlations between
electrons with anti-parallel spin which are being introduced since electrons with
parallel spin are already kept apart by the Fermi-Dirac statistics even in the
Hartree-Fock approximation. Thus the introduction of correlation effects will
lower the energy of non-magnetic states more than that of the ferromagnetic states,
and so make the condition for ferromagnetism more stringent. This is indeed found
to be the case.
256 J . H ubbard
Using the formula (61) for the density of states and the condition (17), na is
determined at the absolute zero of temperature by the condition

w ,= T P^E^(77)
J -oo
which is the analogue of (18); (isi determined by the conditio
take over the discussion of the condition for ferromagnetism in H artree-Fock
theory given in § 3 almost word for word. One finds th at the condition th at ferro­
magnetism just be possible is th at n^is a double solu
condition is just (67) together with

- i <78)
I t is difficult to picture the condition (78) without reference to some specific
density of states function P(E). Consider then the density of states function given
by (64). In this case H artree-Fock theory gives according to (21) the condition for
ferromagnetism / >A (79)
independently of n. To investigate the form taken by the condition (78) in this case
one may note th at the density of states formula (65) can also be written
p,(E) = ( l / A ) { « ( £ - C li_ , ) - S ( j ; - r _ u ) + 9 ( £ - ^ _ 1) - « ( E - f r , 1)}, (80)
which when substituted into (78) gives
1 p dE^n
-1 2 2 / ? - j ^ Kf)dE. (81)
A j - c — ±1 ft —± 1 an
If [i is in the lower band, E 'L ^ ^ < fi < E a_11 (one need only consider this case
because of the symmetry between electrons and holes) then this condition becomes

1 i ^ - 1 ,- 1 (82)
A dn ’
or, using the formula (66) with n_a =\n

1< + <83)
Since fi is in the lower band one must have 1. But for 1 the condition (83)
cannot be satisfied for any I and A. Thus the approximate correlation theory of
§ 6 predicts th at ferromagnetism is not possible for the density of states function
of (64 ) even though Hartree-Fock theory gives the condition (79 ).
I t might now be inquired whether the impossibility of ferromagnetism in a
general consequence which can be deduced from the approximate solution of § 6.
That the answer to this question is no, can be demonstrated at once by giving an
example of a density of states function (Efor
P ) which
Consider the density of states function
P(E) = l/8 if T0—|-A <-|A + f £
or if T0-b -|A — < < T0+ ^A
= 0 otherwise, (84)
which represents two square bands of width •§ symmetrically disposed about T0.
This density of states might be thought of as an approximation to a more general
 Electron correlations in narrow energy
257 bands
density of states function which has two high peaks at each end of the band and
a low density of states in between. By a discussion similar to th at given above for
the density of states function (64) one may show th a t for small the condition for
ferromagnetism is I-A/
8< (85)
I n A iy
which can always be satisfied by making 8 small enough.
From this example some impression can be gained of what conditions are favour­
able to ferromagnetism. I t is clear th a t (85) can only be satisfied if 8 is somewhat
smaller than A from which one may infer th a t it is necessary th at the Fermi energy
be in a part of the band in which the density of states is rather greater than the mean
density of states throughout the band. Further, one may note th at since the right-
hand side of (85) is a monotonic increasing function of A, for fixed S the condition
(85) is more easily satisfied for large A. This suggests th at the most favourable
condition for ferromagnetism is when the Fermi energy lies in a high density of
states peak which is well away from the centre of gravity of the band, and th at a
high density of states peak in the middle of the band would be ineffective in
producing ferromagnetism.
Finally, one may note th at since for the band structure (85) p(/i) = 1/8, the
condition (85) can formally be written
1< Ex.(86)
analogous to (21) provided one defines an ‘effective intra-atomic exchange energy’
4x. by 1A
“ 7 V{(i/ + iA)2-i» A /} (87)
Jex is always less than I, the reduction being due to the weakening of exchange
interactions by correlation effects. From (87) one sees that even when I becomes
very large Iex never becomes much greater than the bandwidth A.

The author thanks Dr P. W. Anderson for a very helpful discussion on certain

aspects of this work.
R eferen ces
B ohm , D. & P ines, D. 1953 P h ys. R e v . 92, 609, 625.
Condon, E . U . & S hortley, G. H . 1935 Theory of atomic spectra. C am bridge U n iv e rsity P ress.
E d w ard s, D. M. 1962 Proc. R o y . Soc. A, 269 , 338.
G ell-M ann, M. & B ru eck n er, K . A. 1957 P hys. Rev. 106 , 364.
H errin g , C. 1952 P h y s . Rev. 85 , 1003; 87 , 60.
H u b b a rd , J . 1957 Proc. R oy. Soc. A, 240 , 539.
H u b b a rd , J . 1958 P roc. R o y . Soc. A, 243 , 336.
K u b o , R ., Iz u y a m a , T. & K im , D. 1963 J . P hys. Soc. J a p . 18 , 1025.
L a n d a u , L. D. 1957 J . E xp t. Theor. P hys. ( U .S .S .R .), 32, 59. T ra n sla tio n : Soviet P hys. J E T P ,
5 , 101 .
P in es, D. & N ozieres, P . 1958 P hys. Rev. Ill, 442.
S aw ada, K ., B rueckner, K . A., F u k u d a , N . & B ro u t, R . 1957 P hys. Rev. 108 , 507.
S later, J . C. 1937 P hys. Rev. 52, 198.
S later, J . C., S ta tz , H . & K o ste r, G. F . 1953 P hys. R ev. 91, 1323.
T hom pson, E . D. i 960 T hesis (M .I.T.). To be p ublish ed in A n n . P hys.
W a tso n , R . E . i 960 P h y s. Rev. 118 , 1036.
Z u b arev, D. N . i 960 Usp. F iz. N a u k , 71 , 71. T ra n sla tio n : Soviet P hys. Usp. 3, 320.