Anderson Review Magnetic Exchange 1963

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 116

The0 ry of Magnetic Exchange Interactions:

Exchange in Insulators and Semiconductors

PHILIPW. ANDERSON
Bell Telephone Laboratories, Incorporated, Murray HiU, New Jersey

I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
11. Historical Introduction. . . . . . . . . . . . . . . ........... 101
111. Theories of Ferro-, Ferri-, and Antiferromagnetism. . . . . . . . . . . . . . . .
1. The Heisenberg Hamiltonian and Localizable Spins. . . . . . . . . . . . .
2. Statistical Theories of Magnetism.. . ............ 116
IV. Spin Waves .........................
V. Empirical Values of Exchange Constants in Insulators.. . . . . . . . . . . . . . . . . . . . 142
VI. Theory of Superexchange.. ............................
3. Formal Description of Magnetic Insulators: The Magnetic State.. . . . . . . . 146
4. Origin of Superexchange: Kinetic Exchange. . .
5. Potential Exchange and Other Exchange Effects.. ..................... 169
6. The Isolated Magnetic Ion: Ligand Fields lear
Resonance.......................... ... 181
7. Semiempirical Approaches and Theoretical ................ 197
VII. Double Exchange.. .. ........................................ 209
Acknowledgments. ........................... 214

1. Introduction

The subject of exchange in magnetic materials can be divided into two


parts, referring to insulators and to metals. This distinction is useful from
the magnetic point of view because in insulators (and, of course, in semi-
conductors) the spins and magnetic moments whose alignments lead
to magnetic effects are certainly localizable so that phenomenologically,
at least, they can be described by a spin Hamiltonian which contains
spin operators, and exchange terms of Heisenberg type. Thus are avoided,
from the start, some of the most serious questions of principle which
bedevil the theory of metallic ferro- and antiferromagnetism. Also, there
appears to be little relationship between the mechanisms important in
metals, such as conduction electron polarization, and those in insulators.
99
100 PHILIP W. ANDERSON

Finally, it can be fairly said that most of the mechanisms in insulators are
at least qualitatively, and sometimes quantitatively, understood, while the
exchange question in metals is still almost completely open. The present
review will concern itself entirely with the former, better-understood divi-
sion of the subject.
Except for a few relatively minor questions (e.g., validity of the spin-
wave theory of antiferromagnetism) , we can hope that the statistical
mechanics and dynamics of ferro- and antiferromagnetism in insulators
involves merely continuing improvement of our approximation methods.
I t is only recently that the theory of the exchange parameters has begun
to reach a comparable state; thus while there are a number of reviews
of the statistical and dynamical aspects of magnetism, fewer cover ade-
quately the latter question-that of the exchange integral J, if you like.
Two useful reviews are that of Nagamiya,’ which covers a small part of
the material rather fully, and that of Marshall,2 which is a good brief
survey from a special point of view. For this reason, and because the
question of the calculation of J is, physically, the more fundamental, we
shall concentrate on it. On the other aspects of magnetism in insulators,
we shall emphasize in all case%what helps relate experimental facts to the
exchange phenomenon, and otherwise treat them relatively briefly.
In Part I1 will be given a short historical discussion of the subjects
of antiferromagnetism and of exchange in insulators. For a more com-
plete coverage of most of the earlier material, the review article of Naga-
miya, et aL3 is recommended. Part I11 will concentrate on the Heisenberg
Hamiltonian, with a brief derivation and a discussion of some of the
statistical theories of magnetism based upon it, primarily molecular
field theory, which is by far the most generally useful in the experimental
measurement of exchange. Part IV includes a discussion of some of the
basic facts about spin waves, and a review of some of the most recent
developments in that field. Part V gives a table of measured exchange
integrals in insulators, based in part on a comparable c~mpilation.~
Parts I through V can be thought of as preliminaries to the central
portion of the paper, Part VI. In Part VI the theory of superexchange in
magnetic insulators will be derived and compared with experimental
results. After a discussion of the physical state of magnetic insulators, we
1 T. Nagamiya, “States of atoms in magnetic crystals.’’Lecture at Welsh Foundation,
Dallas, Texas, 1958.
* W. Marshall, in “Perspectives in Materials Research,’’chapter on Magnetism. Panels
on Materials Research of the National Science Foundation, Washington, D. C., 1961.
a T. Nagamiya, R. Kubo, and K. Yosida, Phil. Mag. Suppl. 14, 1 (1955).
J. S. Smart, in “Magnetism” (H. Suhl and G . Rado, eda.) Academic Press,New York
(to be published).
THEORY OF EXCHANGE I N INSULATORS 101

shall review older theories and present ideas about superexchange, and
give a discussion and a diagrammatic classification of all the possible
higher-order processes. We shall then survey the question of computing
exchange integrals, which involves first a review of the basic facts of ligand
field theory.
Finally, in Part VII we shall review the theory of Zener double exchange.

II. Historical Introduction

“he three central themes of the theory of magnetism in insulators were


stated early in the 1930’s: the Heisenberg-Dirac Hami1tonian;- the con-
cept of antiferromagnetism,’ll and the idea of superexchange.12To these
one might add the ideas, of basic importance in understanding the field,
of the crystal field theory.13
The Heisenberg paper pointed out that ordinary Coulomb interactions
among electrons, together with the Pauli principle, could lead to an ex-
change effect strongly coupling their spins, which might give rise to ferro-
magnetism. Dirac gave an explicit proof for a particularly simple case-
electrons, each confined to a different specified orthogonal orbital. He
showed t.hat the exchange interaction coupling the spins of electrons
localized in an assembly of orbitals, of which a typical pair are $4~) and
# i ( ~ ) , may be written:

Vm = -C JijSi‘Sj, (11.1)
i.i

where the “exchange integral” Jij is

Jij = drl jdrze2 I rl - rz I - ’ ~ ~ * ( r i ) ~ j ( r i ) ~ j * ( r z ) ~ i ( r z )

W. Heisenberg, Z.Physik 38, 411 (1926).


6 W.Heisenberg, Z.Physik 49, 619 (1928).
P. A. M. Dirac, Proc. Roy. Soc. Al23, 714 (1929).
8 J. H. Van Vleck, “Theory of Electric and Magnetic Susceptibilitiea,” Chapt. XII.
Oxford Univ. Press, London and New York, 1932.
L. NBel, Ann. phys. [lo] 18, 5 (1932).
l o L. D. Landau, Physik. Z. Sowjetunion 4, 675 (1933).
11 J. H. Van Vleck, J . Chem. Phys. 9, 85 (1941).
la H. A. Kramem, Physiaz 1, 182 (1934).
1s J. H. Van Vleck, “Theory of Electric and Magnetic Susceptibilities,” Chapt XI.
Oxford Univ. Press, London and New York, 1932.
102 PHILIP W. ANDERSON

The relevance of the “Heisenberg Hamiltonian” (11.1) expressed in the


si-sj form to the problems of magnetism seems first to have been realized
by Van VlecklBalthough more complicated equivalent arguments were of
cause employed by Heisenberg.6 It seems to be a valid, and certainly a
useful, approximation in almost all of the problems of magnetism in insula-
tors and semiconductors. (See, however, the section on “double exchange.”)
In these, to at least some approximation, because of the insulating property,
the electrons with spins are confined to localizable sites; thus we do not
need to include their translational degrees of freedom as part of the mag-
netic problem.
I t is with (11.2) that the great question”of the sign of exchange begins.
The first formulation of J in terms of e2/rI2,which is correct only for truly
orthogonal orbitals, is necessarily positive because it is the self-energy of the
(complex) charge distribution $i*$j. The second form is approximately
correct for slightly nonorthogonal 0rbita1s.I~That the second form is
often nearly correct for moderately large interatomic distances (the prob-
lem of “direct exchange”) has been shown by Herring.l6
If the two electrons are in an attractive potential, the second form
can become negative. Calling the attractive potential - V (r) ,

= J .11” - 2SljVij,
where J;j” represents the true exchange integral, S the overlap, and V;j
the cross matrix element of the attractive potential. It is this attractive
interaction between antiparallel electrons in nonorthogonal wave functions
that is the basis of the Heitler-London scheme of chemical binding.
The nearly universal prevalence of normal chemical bonds containing
electron pairs of opposite spin shows us that usually the second term in
(11.3) will prevail. There are two reasons for this: first, attractive po-
tentials of atom cores are generally quite a bit stronger than single-electron
repulsions, the latter being further weakened by correlation effects and
by the wider spatial distribution of the electronic charges; and second,
when #; and # j are made orthogonal, the J integral clearly must have
cancelling positive and negative parts, since #; *$i must change sign and
have zeros.l8
14 W. Heitler and F. London, 2.Physik 44,455 (1927).
16 C. Herring, Revs. Modern Phys. (to be published).
16 R. McWeeny, Proc. Roy. SOC. 8337, 288 (1955).
THEORY OF EXCHANGE IN INSULATORS 103

The hypothetically ferromagnetic interaction (11.1) formed the basis


for Heisenberg's theory of ferromagnetism.6 Clearly, the lowest state of
such an interaction is that in which all spins are parallel, or ferromagnetic.
While it is now generally accepted that magnetic interactions result from
some such combination of Coulomb effects with the Pauli principle, it is
highly probable-and well accepted in the field-that this specific mecha-
nism is very seldom the sole cause of ferromagnetism. This may be said
because in insulators, the only case in which the Heisenberg Hamiltonian
is correct, the interaction is almost universally antiferromagnetic.
N&19first expressed the possibility that the negative sign of interaction,
if present, could lead to a state in which different subsets or sublattices of
the spins in a crystal could align themselves antiparallel. This is the
state later" called antiferromagnetism. N&l, however, did not discover the
thermal transition point analogous to the Curie point, now called the N&l
point; Landaulo stressed the phenomenological similarity of ferro- and anti-
ferromagnetism and suggested the existence of the N&I point. N&l's
theory also suffered by being at first intended for metals, where antiferro-
magnetism is a much more complicated and subtle phenomenon than in
insulators.
HulthBn'' proposed a spin-wave theory of antiferromagnetism of the
semiclassical Heller-Kramers type, but in his thesisI8 he pointed out the
still serious questions of principle about such a theory, adapting earlier
work of Bethels to calculate the exact ground state for a one-dimensional
antiferromagnet. He suggested that it was not in fact antiferromagnetically
ordered. The dilemma was essentially that one can prove that the lowest
energy state of an antiferromagnet must be a singlet state, and thus have
no directional character. But the ordered state is not an S = 0 state-in
fact far from it; it does not even have time-reversal symmetry. (Recently,
Dzialoshinsky20-2ahas pointed out some interesting experimental effects
arising from this paradox.) While the very existence of antiferromagnets
demonstrates that Anderson's suggestion%as to how this dilemma is to be
resolved is correct, there has as yet been no final resolution from the stand-
point of exact theory.
17 L. HulthBn, PTOC.Amsterdam Acad. Sci. 39, 190 (1936).
18 L. HulthBn, Arkiu Mat. A S ~ T WFys.
L . 268 (11) (1938).
19 H. A. Bethe, 2.Physik 71, 205 (1931).
20 I. E. Dzialoshinsky, Z h w . Eksptl. i Teoret. Fiz. 32, 1547 (1957); Soviet Phys. JETP

6, 1259 (1957).
2 1 I. E. Dzialoshinsky, Zhur. Eksptl. i Teoret. Fiz. 33, 807 (1957); Soviet Phys. JETP
6, 621 (1957).
22 I. E. Dzialoshinsky, Phys. and Chem. Solids 4, 241 (1958).
2s I. E. Dzialoshinsky, Zhur. Eksptl. i Teoret. Fiz. 33, 1454 (1958); Soviet Phys. JETP
6, 1120 (1958).
P. W. Anderson, Phys. Rev. 86, 694 (1952).
104 PHILIP W. ANDERSON

Meanwhile N&l,24aBit ter,26and especially Van Vleck” were developing


the simpler molecular field type of theory for antiferromagnetism. These
theories were compared with early experiments (x of MnO and CrzOa,26.n;
specific heat of MnOB; x of MnOB).
The idea of superexchange first appeared when Kramers12 tried to
understand the early results of adiabatic demagnetization, which indicated
that small exchange couplings existed even between ions separated by one
or several diamagnetic groups (halide ions, waters of crystallization, etc.) .
He pointed out that the magnetic ions could cause spin-dependent pertur-
bations in the wave functions of intervening ions, thereby transmitting the
exchange effect over large distances; but ‘no specific mechanisms were
discussed.
The subject was stimulated enormously after 1945 by two experimental
advances. The first of these was the development of technically important
insulating ferromagnetic materials: first the f e r r i t e ~ then ~ ~ ~hard
; ma-
t e r i a l ~ ,and
~ ~ finally the iron All of these materials have in
common the property of being ferrimagnetic, a name invented by NBel to
describe substances with antiparallel sublattices of spins that have un-
equal total magnetic moments, leaving therefore a net remanent mo-
ment. Thus, one is able to exploit the predominantly antiferromagnetic
exchange effect in insulators to obtain a ferromagnetic moment.
The second of these major developments was the appearance of neutron
diffraction as a technique for observing patterns of magnetic order in
ferro-, ferri-, and, particularly, ant if error nag net^.^^^^^^^ Neutrons of thermal
energy have wavelengths of the order of lattice constants, and interact
more strongly through their magnetic moments with spin and orbital
l a L. N6e1, Ann. phys. [ll] 6, 232 (1936).
a F. Bitter, Phys. Rev. 64, 79 (1938).
26 K. Honda and T. Ishiwara, Sci. Repts. TBhoku Imp. Univ. 4, 215 (1915).
~7 K. Honda and T. Sone, Sci. Repts. TBhoku Imp. Univ. 3, 223 (1914).
470 T. Ishiwara, Sci. Repts. Tahoku Imp. Univ. 3, 303 (1914).
28 R. W. Millar, J . Am. Chem. Soe. 60, 1875 (1928).
29 H. Bizette, C. Squire, and B. Tsd,Compt. rend. acad. sci. (Paris) 207, 449 (1939).

ao J. L. Snoek, Philips Tech. Rev. 8, 353 (1946).


31 J. L. Snoek, “New Developments in Ferromagnetic Materials.” Elsevier, Amsterdam,
Holland, 1947.
zz J. J. Went and E. W. Gorter, Philips Tech. Rev. 13, 181 (1952).
s3 J. J. Went, G. W. Rathenau, E. W. Gorter, and G. W. Van Oosterhout, Philips Tech.
Rev. 13, 194 (1952).
F. Berthaut and F. Forrat, Compt. rend. a d . sci. (Paris) 242, 382 (1956).
35 S. Geller and M. A. Gilleo, Acta Cryst. 10, 239 (1957).
86 C. G. Shull and J. S. Smart, Phys. Rev. 76, 1256 (1949).
57 C. G. Shull, W. A. Strawer, and E. 0. Wollan, Phys. Rev. 83, 333 (1951).

a7a C. G. Shull, E. 0. Wollan, and W. C. Koehler, Phys. Rev. 84, 912 (1951).
THEORY OF EXCHANGE IN INSULATORS 105

magnetic moments than directly with electron charges. Thus, the diffraction
of neutron beams reflects the symmetry of the spin moment arrangement
as well as of the atomic arrangements, which affect the diffraction through
specifically nuclear forces. In an antiferromagnet the spin arrangement
has lower symmetry than the lattice alone and leads to extra diffraction
peaks. An example is given in Fig. l,as showing the diffraction pattern
of MnF2 and the spin structure deduced from it. In the ferrites, it is often
possible to combine chemical reasoning with observed saturation moments
to determine sublattice patterns; in the antiferrornagnets this is not
possible and the neutron diffraction tool is indispensible.
These few concepts have provided the basis for an expanding science
of insulating magnetic materials during the postwar years. Some of the
major developments are the following.
A general concept which developed outside of this field but has had
wide application in it is radio frequency resonance spectroscopy. Para-
magnetic resonance has been used to measure exchange parameters, both
d i r e ~ t l yand
~ ~by
~ ~means of linewidth results and exchange broadening
and narrowing the0ry.~l-42Nuclear resonance has been applied in the
measurement of wave functions of the electrons responsible for ionic
spins4a-46and in the antiferromagnetic state to measure sublattice mag-
netization~."~ Ferromagnetic
~ resonance was applied very early in its
development to ferrimagnetic substances@' and is capable of yielding funda-
mental information about magnetic properties (e.g., direct measurement
of spin-wave-like or spin-wave modes61-") . Finally, antiferromagnetic
resonance was first understood theoretically by Nagamiyas4and KitteP
38 R. A. Erickson, Phys. Rev. 90, 779 (1953).
39 B. Bleaney and K. D. Bowers, Phil. Mag. 43, 372 (1952).
40 J. H. E. Griffiths, J. Owen, J. G. Park, and M. F. Partridge, Proc. Roy. SOC.
A260,
84 (1959).
41 E. Ishiguro, K. Kambe, and T. Usui, Physicu 17, 310 (1951).
42 P. W. Anderson and P. WeiSs, Revs. Mode?n Phys. 26, 269 (1953).
43 M. Tinkham, Proc. Roy. Soc. 8336, 535, 549 (1956).

I 4 R. G. Shulman and V. Jaccarino, Phys. Rev. 103, 1126 (1956).


46 V. Jaccarino and R. G. Shulman, Phys. Rev. 107, 1196 (1957).
R. G. Shulman and V. Jaccarino, Phys. Rev. 108, 1219 (1957).
47 N. J. Poulis and G. E. C. Hardeman, Physica 18, 201 (1952).
48 N. J. Poulk and G. E. C. Hardeman, Physicu 18, 315 (1952).

49 V. Jaccarino and L. R. Walker, J . phys. r d i u m 20, 341 (1959).

60 W. A. Yager, J. K. Galt, F. R. Merritt, and E. A. Wood, Phys. Rev. 80, 744 (1950).

6 1 R. L. White and I. H. Solt, Phys. Rev. 104, 56 (1956).

62 J. F. Dillon, Phys. Rev. 112, 59 (1958).

68 M. H. Seavey, Jr. and P. E. Tannenwald, Phys. Rev. Letters 1, 168 (1958).

64 T. Nagamiya, Progr. Theoret. Phys. (Kyoto) 4, 350 (1951).


C. Kittel, Phys. Rev. 82, 565 (1951).
106 PHILIP W. ANDERSON

01 I I I I I 1 I
I0 I5 20 25 30 35 40
COUNTER ANGLE IN DE6REES

FIQ. 1. Diffraction pattern and magnetic structure of MnF2.m


THEORY OF EXCHANGE IN‘ INSULATORS 107

who pointed out that exchange enters the expression for the frequency,
then observed and interpreted by the Leyden groupBJ7 in the low N&l
temperature salt CuC12.2H20, and finally observed in the “classical)’
antiferromagnet MnF2.s*+sg
The developments in basic theory have been of three kinds. First is the
quantum mechanics and statistics of the Heisenberg Hamiltonian. Various
consequences of the spin-wave idea have been worked O U ~ but ~ ~not, ~ ~
justified further. At the same time, it has been found important and useful
to correct the molecular field theory for next nearest neighbor and geo-
metric effects.61-6aOther types of quantum theories and of statistical
theories have been applied to antiferromagnetism, but have probably not
yet justified their difficulty with usefully more accurate results, at least
as far as the study of real magnets is concerned. (See, however, M. E.
Fisher64for an important application to antiferromagnetic Curie points,
and J. S. Smart4 for applications to exact determination of exchange
parameters.)
A second field in which important developments have appeared is the
crystal field theory, now generalized to “ligand” field theory,BSwhich has
been applied to many important properties of these magnetic materials,
such as lattice distortions, atomic distributions, optical spectra, and now
is influencing the understanding of the patterns of order and directionality
in antiferromagnets and even the calculation of exchange.
The final advance has been in the understanding of the exchange inter-
action itself, begun by the suggestion of explicit interpretations of Kramers’
idea,66-67going on to new mechanisms,@*69 and finally arriving at a synthesis
qualitatively different from the Kramers concept.7o
66 N. J. Podis, J. van den Handel, J. Ubbmk, J. A. Poulis, and C. J. Gorter, Phys. Rev.
82, 552 (1951).
67 J. Ubbink, J. A. Poulis, H. J. Gerritaen, and C. J. Gorter, Physica 19, 928 (1953).
u S.Foner, Phys. Rev. 107, 683 (1957).
F. M. Johnson and A. H. Nethercot, Jr., Phys. Rev. 114, 705 (19.58).
Oo R. Kubo, Phys. Rev. 87, 568 (1952).
61 L.NBel, Ann. phys. [12] 3, 137 (1948).

P. W. Anderson, Phys. Rev. 79, 705 (1950).


e3 J. S. Smart, Phys. Rev 86, 968 (1952).
a4 M.E.Fisher, Phys. Rev. Letters 1, 321 (1958).
66 K.W.H. Stevens, Proc. Roy. SOC. 8219, 542 (1953).
66 P. W.Anderson, Phys. Rev. 79, 350 (1950).

liaaJ. H.Van Vleck, J . phys. Tadium 12, 262 (1951).

67 G.W.Pratt, Jr., Phys. Rev. 97, 926 (1955).

68 C. Zener, Phys. Rev. 82, 403 (1951).

o9 N. F. Ramsey, Phys. Rev. 91, 303 (1953).


70 P.W.Anderson, Phys. Rev. 116, 2 (1959).
108 PHILIP W. ANDERSON

111. Theories of Ferro-, Ferri-, and Antiferrornognetisrn

1. THEHEISENBERG
HAMILTONIAN
AND LOCALIZABLE
SPINS

The Dirac derivation of the Heisenberg Hamiltonian (11.1) applies


only to the following highly oversimplified model: consider exactly N
electrons occupying exactly N different, orthogonal orbital states fn(r).
These are visualized as being, say, d orbitals on each transition metal ion
in a crystal. It is assumed that for some ypecified reason each orbital
contains always one and only one electron; that is, the electrons may be
localized and thus we may label the electrons by means of their orbitals
(sites) n rather than by referring to the particular electron j . We shall
discuss the physics behind this restriction rather extensively later.
The Hamiltonian is, of course,

the summations being over the N electrons i. If the orbitals fn are really
orthogonal, then the kinetic energy and the potential energy V(rJ that
results from ion cores, etc., do not depend on the spin orientations, and
we have only to study the Coulomb interaction e2r;;.
Dirac’s derivation involved a complicated argument with permutation
operators. The result can be arrived at in a few lines by means of second
quantized operators, which will be very useful also in the later develop-
ments. As an introduction to this algebra, therefore, we shall transcribe
Dirac’s derivation into it.
For the benefit of those somewhat unfamiliar with second quantization,
we shall give a brief discussion of it. Second quantization in a problem
such as this is simply a convenient way to handle the transformation to a
representation of quantum mechanics in terms of “occupation numbers”
rather than of explicit determinantal wave functions. In an occupation-
number representation, one describes a many-electron wave function in
terms of the number of electrons in each of a set of orthogonal one-electron
orbitals &. For example, the determinant
THEORY OF EXCHANGE IN INSULATORS 109

where the 4 ( n l . . . ) is the occupation-number representation of #. Of


course, the most general possible wave function is a linear combination of
determinants so that the representation of the most general possible
wave function would give the coefficients x(n1, 722, 123, in such an
..
- a * )

expansion in terms of the +(nl, : a )

# = c
n1-OO.1;n2-0,1
***x(n17 , w - . .)
722, n3, . * - ) 4 ( n l 122, (1.3)

One possible set of orthogonal orbitals is the &functions 6(r - ro). The
wave function in which one electron is certainly at ro is then 4(nr0= 1,
nrl =,0, nr2= 0- -). A one-electron function # ( r ) could then be expanded
as
# = &(ro)4(n,, = 1, n, = o - . - ) . (1.4)
The coefficient x, now only a function of the position of the one electron
(ro), is just the original numerical representation of the wave function
# ( T o ) . (This becomes clear when we realize that the probability density
at ro is just I x(ro) 12.)
Second quantization comes in because it is convenient, instead of writing
--
4(nl, n2,n3, .), to write the wave function containing electrons in states
n17nz, etc., in terms of creation operators. The creation operator cl* creates,
out of the vacuum, one electron in state &. For instance, our original
two-electron state # in (1.2) could be written as
J. = CI*CZ**O,
where #o is the vacuum state, containing no electrons. The general state
(1.3) may be written
# = C ( x ( j 1 ,j2 jn)cXcX ... cX)#o,
where the summation is over the various combinations of one filled orbital
j l , two filled orbitals jl,
jz,etc. Usually, of course, the number of electrons
m is fixed and this sum will be over the possible combinations of m filled
orbitals.
Electrons are fermions; thus no two can occupy the same state and
cl*cl*\k = 0. Also, we must, in the two-electron state (1.2), associate
cz*, the first electron created, with the second column of the determinant,
and cl* with the first column. Thus the opposite order gives the negative
determinant
Cl*CZ*\k= -c2*c1*.
These relations may be summarized in the anticommutation relation
ci*cr + cj*ci* = 0. (1.5)
110 PHILIP W. ANDERSON

The creation operator which creates an electron which is certainly a t r


is called the electron field operator $*(r) ; any ci* may be expanded in
terms of it as

as may easily be seen by considering the oneelectron states (1.4). The


index i and the variable r are here considered to include the spin variable u.
Conversely, the relation (1.6) may be inverted to get the electron field
in terms of the ci*:

a very important result.


We can also introduce the Hermitian conjugate ci of the creation
operator ci*) which annihilates electron i. That is, acting on a state ci*\k,
it has the effect of removing the electron i: cici*\k = 9 if \k did not already
contain electron i. Working out all possible combinations, we have
ni = 0 ni = 1

ci*ci9 = 0 ci*ci9 = \k

c,ci*\k = 9 CCi*\k= 0.
Clearly, the effect of these combinations is related to the effect of the
number operator ni by the identities
ci*ci = ni c,ci* = 1 - ni
, c,.
c.* + cici* = 1.
(1.8)

Without proof we give the generalization of (1.8) :


ci*cj + c,ci* = 6ij. (1.9)
Now we have only to discuss how to manipulate with these wave func-
tions, which means merely that we must express the Hamiltonian properly.
In doing this it will be best to consider a simple one-electron Hamiltonian
V(r) first.
In the occupation-number representation, if V(r) has a matrix element
THEORY OF EXCHANGE IN INSULATOR8 111

it means that the effect of V12on a wave function in which & is occupied
will be to replace it by 41 in the manner
V (rJ +Z (rd = V d 1(rJ ;
thus, in general,
VC#J(n1
= 0,122 = 1, . . a ) = V,,C#J(n, = 1, nz = O...)

+ C Vj24(nl
i
= 0,122 = 0 ... nj = 1) + ..a. (1.11)

The complicated Eq. (1.11) may be replaced by the simple concept that
VIZmultiplies an operator which annihilates the electron in state 2 and
replaces it in state 1, namely, c1*c2.Generally,

(1.12)

Inserting the expansion (1.6) for ck* and CI leads, on using the orthogonality
and normalization of the +’s, immediately to the even simpler result

(Here we have reintroduced the spin index u explicitly.) There is no need


here to go into the very similar argument by means of which we write
down the rest of the Hamiltonian,

+3 u,u‘
/ dr dr’$*(r’, d)$*(r, a)e2 I r - r’ I-llL.(r, u)lL.(r’, u’). (1.14)

(The order of operators in the Coulomb interaction is dictated by leaving


out the exchange self-energy.)
Equation (1.14), together with the anticommutation rules (1.9) and
(1.5), tells us how the Hamiltonian acts on any wave function defined in
terms of creation operators; thus they are a complete expression of the
quantum mechanics of the many-electron system.
Each orbital fn(r) in the Dirac model has two sets of creation and
destruction operators, c k and cn+, c*, and c,, which refer to the states of
up and down spin, respectively. The requirement that one and only one
112 PHILIP W. ANDERSON

electron may occupy orbital n means that only one or the other state (n+)
or ( n - ) may be full; i.e.,
n,,+ + n- = c%c,,+ + c*,c,- = 1. (1.15)
The usual Pauli spin operators u may be defined in terms of the fermion
operators as
un+ = snz + is,u = *
Cn+Cn-

,a = snz - is,” = C*n-Cn+ (1.16)


unz = 2 ~ n ’= n,,+ - nn-.
Note that all of these are pairs of fermions and thus, properly, commute
rather than anticommute with the spin operators for other orbitals. Equa-
tion (1.16) may be verified by seeing that the fermion combinations have
exactly the correct effect on wave functions satisfying (1.15).
We get the Heisenberg Hamiltonian immediately by considering the
expansion of the Hamiltonian (1.14) in terms of the operators cn..
The one-electron terms are independent of spin, as we can easily see
by observing that the first summation can contain no operators like a,,+
or un- and only combinations n,+ +
n,+, not n,,+ - nn-. The second term,
however, can depend on the relative spins. On inserting (1.7) into the
Coulomb interaction, we obtain

X 1dr 1dr’e2 I r - r’ I-lfn*,(r’)fn*,(r)fn,(r)fn,(r’). (1.17)

Dirac’s model directs us to ignore the possibility of transfer of electrons


from one orbital to another, so the pair (nl, nz) must be the same as the
pair (na,nr). This leaves us with only two terms

where

and

Jnnn= 1 1dr dr’j*nt(r‘)fn(r’)fn*(r)fnp(r)e2 1 r - r’ 1-l. (1.20)


THEORY OF EXCHANGE IN INSULATORS 113

The first term is simply the ordinary Coulomb repulsion between the
electrons in the two orbitals, and is also not spin-dependent. We use the
identities (1.16) on the second term:
*
C*~’+C+C~’--C~+ + Cnf-Cn+Cnl+Cn-
* * = -(stsn- + sigsn+) ;
c*,~+c*, , + C ~ ~ + C ~ + + c*,~-c n-Cn -cn-
* I
= (nnf+nn+ + nnf-nn-)
= --a - 2snzsn#’;
thus it becomes
V = -C
n,d
JnnJ(f + Sn’SnS), (1.21)

which is the Heisenberg Hamiltonian.


The coefficient Jnn#is the Coulomb self-energy of repulsion of the
“exchange” or “overlap” charge density f $ (r)fn (r) . This repulsion,
since it comes from the two electrons being at the same place, in a sense
does not exist for parallel electrons, so that ferromagnetism is favored.
That is, K is too large because it does not take into account the fact that
the electrons are prevented from overlapping; thus the correction J is
necessary for parallel spins.
Now let us consider whether the physical conditions for this derivation
ever occur in practice. The fact is that one of three cases occurs: (1) there
is no paramagnetism; (2) the orbitals are orthogonal but are degenerate,
or nearly so, so that it is important to discuss the effects of orbital
transfer and multiple occupation; or (3) the orbitals are spatially sepa-
rated and occupied singly, but are not orthogonal. The basic difficulty
is that one-electron energies, the energies resulting from the first term of
(1.14), and in particular from the ion-core potentials, are considerably
larger than Coulomb interactions between single pairs of electrons. This
is particularly true for exchange energies because

as a result of orthogonality. Hence, the exchange charge has nulls which


usually occur where the separate functions are large. Thus in case (l),
the most usual one, the various electron orbitals are nondegenerate (such
as the molecular orbitals in most molecules, the running waves in most
metals and valence crystals, s states in atoms). Then the electrons will
simply fill the various orbitals in antiparallel pairs because the one-electron
energy to be gained by doing so is much greater than the exchange energy
lost.
114 PHILIP W. ANDERSON

Case (2) occurs in many free atoms, a few molecules (02)) and the
ferromagnetic metals. In free atoms the orbitals are, through symmetry,
degenerate; when the number of electrons is not enough to fill a degenerate
shell exactly, there may be several orbitals that are only singly occupied.
The ferromagnetic sign of the exchange interaction (1.21) causes the
spins to align themselves parallel (Hund’s rule); but at the same time
the orbital moments are free to move and the orbital state must be deter-
mined simultaneously. In ferromagnetic metals, much the same kinds of
difficultiesarise. There is every reason to believe that the change in occup-
tion of orbital states caused by spin effects is an important phenomenon
in ferromagnetic metals.
Finally, there are the cases of relatively separated magnetic ions (or
free radicals or other groups) . These include the classical antiferromagnets
such as oxides and fluorides, hydrated salts, and free radical crystals.
The Heisenberg theory should apply here, but there is one difficulty: the
interaction is usually found to be antiferromagnetic. The reason for this
has been known for a long time: the best wave functions f,, are in some
basic sense not orthogonal to each other. Slater” has pointed out in great
detail that in the typical problem of H z , the use of truly orthogonal orbitals
in the Heitler-London picture leads to very bad results, because it discards
all the binding energy without improving the correlation energy ap-
preciably. That is, the one-electron energy, which is large, tends to pull
the electrons into identical orbitals. This is the binding effect. The Coulomb
repulsion, on the other hand, keeps the electrons on separate atoms. The
resulting compromise leads to the antiferromagnetic interaction we shall
discuss in Part VI. However, a fairly good approximation is given by as-
suming that the electrons stay in atomic orbitals which are not quite
orthogonal to each other. As a result, the foregoing calculation must be
modified.
It can be shown that to the lowest order it is correct to work with
nonorthogonal orbitals as though they were orthogonal, except that the
matrix elements which multiply a given combination of creation and
annihilation operators must now be those of the full Hamiltonian taken
with the given nonorthogonal functions. The reasoning is that one wants
the effect, in the n-representation, of operating with the Hamiltonian,
expanded in terms of the functions used. Thus, J contains a contribution
from the attractive core potential V , as shown in (11.3) of the introduction,
and can be antiferromagnetic.

J. C. Hater, J . Chem. Phys. 19, 220 (1951).


THEORY OF EXCHANGE IN INSULATORS 115

In Herring16the question of the validity of the Heisenberg Hamiltonian


will be discussed in general terms under a very broad class of circumstances.
It will be shown that the various kinds of objections which have been
raised to its use (e.g., SlaterT2) are not valid in most interesting cases.
We shall thus leave this question, as far as theory is concerned, with
this general assurance and with the comment that in Part VI of this
paper we shall give a general formalism for expressing exchange as the
result of a perturbation procedure which should usually converge rather
rapidly. Wherever this is valid, the si. sj Heisenberg Hamiltonian should
be very accurately correct.
There is some experimental evidence concerning the validity of the
Heisenberg Hamiltonian. Since the Coulomb interaction is scalar in spin
space, the only possible form for a two-spin interaction (neglecting spin-
orbit effects) is

V a = f(Si*Sj) = J(Si.Sj) + J1(Si.Sj)2+ a * . . (1.22)

One type of experiment for deciding whether there are higher terms such
as Jl in the series (1.22) is based on the comparison of the exchange

-
parameters deduced from high-temperature data (Curie-Weiss constant e) ,
where Si. Si 0 with those deduced from spin waves or other low-tempera-
ture data, or even from the Curie point. In MnF2 the agreement is fairly
good on all three counts.4-58Ji9As yet there are no discrepancies in any sub-
stance between 0 and T , or other measures of exchange integrals4 which
cannot be ascribed to another cause such as next-nearest neighbor inter-
action. Experiments on spin pairs in r ~ b y have ~ ~ sdemonstrated
~ ~ the
existence of very small terms of the J1 type, of about the order to be ex-
pected from Anderson’s perturbation theory (Ander~on;~O see also Part
VI of this paper).
We shall assume then that it is a fair approximation to take the
Heisenberg Hamiltonian at least as a phenomenological starting point,,
since the fact that the materials we study are insulating means that the
spins are localized. For the rest of these sections, then, we shall start from
the Heisenberg Hamiltonian and discuss the phenomena of magnetism
to which it leads, particularly with regard to determining the parameters
Jd .
n J. C. Slater, Revs. Modern Phys. 26, 199 (1953).
78 H. Statz, L. Rimai, M. J. Weber, G. A. deMars, and G. F. Koster, Phys. Rev. Letters
4, 125 (1960).
74 H. Statz, L. Rimai, M. J. Weber, G. A. deMars, and G. F. Koster, J . Appl. Phys. 33.
Suppl., 2185 (1961).
116 PHILIP W. ANDERSON

THEORIES
2. STATISTICAL OF MAGNETISM

A considerably body of theory relates to the following question: Given


the Heisenberg Hamiltonian, what is the statistical behavior of the system,
and in particular, what is the Curie point and the thermal variation of
various properties? This has also been a primary subject of several review
papers; thus we shall not go into it very thoroughly. The great difficulty
of the subject has led to much work with simplified models that are only of
qualitative help to us here. In particular, the extensive study of the Ising
is not of much practical help.
Recently, some serious attempts have been made to apply really
sophisticated statistical methods to simple cases of antiferromagnetism in
order both to test the theories and to determine some exchange parameters
to higher accuracy than is possible with the simple molecular field theory.
An excellent review article covering this work is that of Smart4; we shall
draw heavily on his results in Part V of this paper.
Nevertheless, from the point of view of the physics of exchange effects,
most of the work on accurate higher-temperature statistical theories
remains pretty much a mathematical exercise which does not yield many
important new physical results. Almost all of the physics of the phenomena
may be adequately understood in the simple molecular field approximation,
which we shall treat in some detail in the following. Most of the physical
results that are not accessible to the molecular field theory are contained
in the spin-wave method, which is accurate at low temperatures and is
important in the dynamics of magnetism and for understanding the
antiferromagnetic ground state. Part IV contains a discussion of some
results on spin waves.
The molecular field technique was developed, primarily by Heisen-
berg,4 Van Vleck,ll and NCel,B1 out of the very early ideas of P. Weiss,”
The simplicity of this method allows it to be applied in the complicated
situations which actually arise in practice; therefore, it is the one to which
we must look for most of our estimates of exchange parameters.
The Hamiltonian for general spins S, is
X = - C Jnn*Sn*Sni-Cgp~H*Sn. (2.1)
n,d n

We have included an external field, for purposes of calculating suscepti-


bility. The basic idea of this method is like that of the self-consistent field
method of quantum mechanics. We replace all the spins except the one
under consideration by their averages and treat the statistics of this one
76-G. R. Newell and E. Montroll, Revs. Modern Phya. 26, 353 (1953).
76 P.Weiae, J . phgS. 6, 667 (1907).
THEORY OF EXCHANGE IN INSULATORS 117

spin alone. This requires that we solve the problem of the effective Hamil-
tonian for spin n:

X, = - - S n * [ 2 C Jnn*(Sn#)+ ~PBH].
nl
(2.2)

The limit in which this becomes exactly valid is that of long-range


forces in which the number of neighbors Z goes to 00, because then the
fluctuations of the sum on the right will become negligible. (Note the
obvious generalizations to the anisotropic case, where H may contain the
anisotropy field and g may be a tensor, as well as to cases of zero-field
splitting, etc.)
It is convenient for much of the following to measure fields in energy
units, and thus to write
X, = -SS,.H,; (2.3)
where
Hn = HEn Ho, +
HE, = 2 C
nf
Jnnc (Sn?>,

The correct average value of the spin Snis well known to be


(Sn ) = SBS( SHn/k T) (2.5)
where S is the total spin quantum number. We shall later generalize to the
case in which S varies from one spin to another. Vectorially, S n is parallel
to Hn. Again there is a necessary generalization to anisotropic g-factors
in many cases. BS is the Brillouin function for spin S:
2s+1 2s+1 1 Y
Bs(y) = -coth -y -- coth -
2s 2s 2s 2s

= tanhy (S = 4) (2.6)

The number of quite complicated situations which this simple formalism


can describe is very large. We shall work out completely two simple cases
which illustrate some of the effects and then discuss several general theorems
about the more complicated cases.
118 PHILIP W. ANDERSON

One case we shall develop is the simplest antiferromagnet under the


assumption of two sublattices. According to this assumption, the most
important interaction is that between nearest neighbors; moreover, the
neighbors of an atom on sublattice 1 always lie on sublattice 2. We shall
also include a smaller intrasublattice term for illustration. Then the
atoms on a given sublattice all act the same; thus if we call the average
spins on sublattices (1) and (2) S1 and S2, then

Hi = -2zJ(SZ) - 2z’J’(Sl) + gpH


(2.7)
H, = --2zJ(&) - 2 ~ ’ j ’ ( S z+) gpH.
Here z and J are the neighbor number and exchange between sublattices,
whereas z’ and J’ are the same quantities within the sublattice. We have
taken the exchange integrals to have the antiferromagnetic sign.
At high temperatures no magnetization will occur spontaneously and
we may expand the Brillouin function about H = 0,

This gives

(S1)=
S(S + 1 ) ( - 2 z J ( S z ) - 2z’J’(S1) + gpH)
3kT

On adding these two equations and solving, we get

W(Stot) = (2.10)

where

C=2
g2fi2S(S + 1)
3k ’
2
e = -(Jz
3k
+ J ’ z ’ ) S ( S + 1). (2.11)
THEORY OF EXCHANGE I N INSULATORS 119

For this solution (S1)


= ( S z ) .On the other hand, a new solution appears
at H = 0 when (2.9) has a homogeneous solution; i.e., when

l + -
2
3k TN
z’J’S(S +
= 0,

or I

- ( J z - J’z‘)S(S + 1 ) .
2
TN = (2.12)
3k
This is the antiferromagnetic NBel point.
The effect of an interaction within the sublattice in making 8, the
Curie-Weiss temperature, different from T N appears here. Note, however,
that if J’ is too large and negative, other more complicated sublattice
structures will appear, so that e and T Ncannot, in general, become different
by orders of magnitude.
For an accurate solution below the NQ1 point, we must refer to the
original papers. Some aspects of this solution, however, are clear. Below
this point, H will have relatively little effect and the task is to solve self-
consistently the two equations

(81) = SBs[S/kT(-22J(Sz) - 22’J’(Si))]


(2.13)
(Sz) = S B s [ S / k T ( - 2 z J ( & ) - 2 d J ’ ( S z ) ) ] .
By symmetry, (81) = - ( S z )= ( 8 )in the antiferromagnetic case, so that

( S ) = S B s [ ( S / k T ) 2 ( z J - z’J’) ( S ) ] . (2.14)

This gives the same magnetization curve as ferromagnetism for the two
sublattices individually.
A certain number of the susceptibilities can be obtained rather easily.
Apply H perpendicular to the antiparallel spins. Then, to a f i s t approxi-
p r o b a t i o n , the molecular field parallel to the spins is constant at
2 ( z J - z’J’) (So) while the perpendicular molecular field involves both
H and the exchange field resulting ftom a small rotation of the spins. Expres-
120 PHILIP W. ANDERSON

sing the condition that the spin be parallel to the molecular field we obtain
HI
-=--
6s
- g p H - 2(2J + 2‘J’)GS
HI1 (S)O 2(2J - z’J’) (So) ’
which gives

(2.15)
g2P2
XI = - = const.
22 J

Note that (2.10) gives the same value at T = T N .


On the other hand, the “parallel” susceptibility, for H parallel to the
spin and internal field, must vanish a t T = 0 because then the applied field
simply adds to the internal field, while a t T = 0 (S,) depends on the
direction of H , only. Thus, assuming that the direction of (S) is fixed by
anisotropy fields, the average susceptibility at T = 0 should be 4 that at
TN. In this simple model there is a sharp break in the “parallel” suscepti-
bility at T N .
A second simple case is that in which we neglect J‘ but make the spin
quantum number Sl different from Sz. This provides a model of a ferri-
magnet. Now the high-temperature equations for susceptibility equivalent
to (2.9) are

(2.16)

The solution is much more complicated than (2.10).Let us introduce the


abbreviations

c
1 =
9l2P2SI(Sl + 1); cz = ...
3k
(2.17)
THEORY OF EXCHANGE I N INSULATORS 121

Then

and

This expression goes t,o infinity a t the temperature


T, = d&&, (2.19)
which is the ferrimagnetic Curie point, and below which there is spon-
taneous magnetization.
- On the other hand, at very high temperatures,
T >> T,, the susceptibility appears antiferromagnetic :
--+-(T+
1 1 ( c1cdlez)
112).

x c1 + cz c
1 + cz (2.20)

TEMPERATURE IN DEGREES KELVIN


FIQ.2. Typical susceptibility curves of l/x va. T for a ferromagnet, a ferrimagnet,
and an antiferromagnet.
122 PHILIP W. ANDERSON

Figure 2 shows the three possibilities for the behavior of the suscepti-
bility above the Curie point : ferro-, antiferro-, and ferrimagnetic; the latter
is unmistakable. Relations like (2.20) are particularly useful because the
molecular field assumption becomes exact at high temperatures.
Before going on to a few of the interesting effects arising in more
complicated cases, let us discuss some general aspects of the molecular
field theory. The basic equations in the region above the Curie-NBel point
are the susceptibility equations :

If the abbreviations

(2.22)

and
+ +
2J,,*S,1‘2S,~1’2(Sn 1)1‘2(Sn~1 ) l I 2
en,. =
3k
, (2.23)

are introduced, and the variable


2, = M,/Cn”2 (2.24)
is used, (2.21) may be written

Tx, = C ennlxnt+ C,‘l2H. (2.25)


nt

Equation (2.25) is a matrix equation for the vector in “sublattice space”


x = (xn) involving the matrix 8 = (en,,) and the vect.or C = (C,) :

x = (T1 - 8)-1C1/2H. (2.26)

The magnetization is given by the rather elegant formula

x =

The N&l or Curie point occurs at the highest eigenvalue of the matrix
8. As in the MnO structure, there may be many possible antiferromagnetic
patterns and each eigenvector x represents such a structure; structure
transitions occur at ratios of the various J’s where eigenvalues of 8 cross
each other.
THEORY OF EXCHANGE I N INSULATORS 123

The high-temperature solution comes from expanding (2.27) in powers


of B/T:

X“
(C112,C1’2) (Cl’Z,8C1’2)
+ ...
T T2

The “paramagnetic Curie point” is thus given by

2 C n n J Jnn,Sn(sn + 1) ( S n J ) (sn, + 1)
C n S n ( S n + 1)
0 - (2.28)

There is often, but not necessarily, a relationship between the order


at absolute zero and these eigenvalue equations. The order at absolute
zero, in the molecular field approximation, is given by minimizing
(X) = 2 C J,,*S,.S,*,
nnf

under the condition that I S, I = S,, and that the S,’S be taken as classical
vectors. If we define a new 8 matrix 8‘ by
= Jnnl SnSn#/Sk, (2.29)
we can find eigenvectors x‘ of the matrix 8’. As pointed out by Luttinger
and Tiszan and by Luttinger,l* if the largest of these satisfies the ‘(strong
condition” 1 x, l2 = 1, it is also the minimum-energy solution and thus
gives the absolute zero order.
I n general, 8 # 8’, although they will be similar for large S, even in a
ferrimagnet. It is also unusual in a ferrimagnet for the strong condition
to be satisfied. Nonetheless, the ground state order and the largest eigen-
values of 8 and 8’ often resemble each other closely.
Kaplan et ~ 1 . 7 9 3 8 0 have extended the validity of this type of theory in
such cases by means of a simple generalization. Essentially, the method
of Lagrangian multipliers is used. The main difficulty is that when the differ-
ent sites are crystallographically inequivalent, as they must be in a ferri-
magnet, the magnitudes of the eigenvector components for the different
kinds of sites are almost automatically unequal and thus the components
77 J. M. Luttinger and L. Tisza, Phys. Rev. 70, 954 (1946).
78 J. M. Luttinger, Phys. Rev. 81, 1015 (1951).
79 T.A. Kaplan, D. H. Lyons, K. Dwight, and N. Menyuk, J . Appl. Phys. 32, Suppl.,
135 (1961).
00 D. H. Lyona and T. A. Kaplan, Phys. Rev. 120, 1580 (1960).
124 PHILIP W. ANDERSON

are unable to satisfy the strong condition. Obviously, then, we should


append to the eigenvalue equations a series of conditions like

c xn2 = No
(2.30)
c xn2
aites b
= Nt,, etc.

These conditions may be introduced into the eigenvalue equations by


means of Lagrangian multipliers,
C O'nnlxn* +
nf
c
o , b .* *
Mn,fin = Enxn-

We see that this is equivalent to introducing a multiplier into the defini-


tion of x, which is Kaplan's method. By means of this procedure, he can
bring most ground state energy problems in ferrimagnets into the form of
eigenvalue problems. For the details of special applications, see the paper
of Lyons and Kaplan.
Many of these results simplify greatly in the case of pure antiferro-
or ferromagnets. In the first place, gn and Snthen do not vary from one
spin to another so that (a) Cn is constant, and (b) 8' = const X 8. These
simplifications are not very useful unless, at the same time, (c) all sites
are equivalent so that

(2.31)

If this is the case, one eigenvector of 8 is always the constant, and the
corresponding eigenvalue is

(2.32)
8. (const) = Oo(const).

Then Eq. (2.27) degenerates to

(2.33)

The Curie-Weiss constant is the sum of the exchange integrals, and the
l/x vs. T curve is linear in the molecular field approximation. All of the
other solutions of the eigenvalue equation are then purely antiferromag-
netic, since they must be orthogonal to the constant one.
THEORY OF EXCHANQE IN INSULATORS 125

If the lattice involved is also a simple Bravais lattice-all points being


translationally equivalent-all solutions also satisfy the strong conditions,
since they are simply z,,a eik.r.This is true of many antiferromagnets as
far aa the exchange Hamiltonian is concerned, even the rutile lattices
such as MnF2, MnOz. Then the highest N&l point agrees with the T = 0
stable solution. This will usually be the case in practice even when the
lattice is not of the Bravais type. Thus, the simple exchange molecular
field theory cannot explain such complications as ferromagnetic-anti-
ferromagnetic transitions, etc. in simple Bravais lattices (some such
complications are mentioned in Pratt”) .
‘One expects that most often the eigenvalue T N will be numerically
smaller than 80, since in the eigenvector some Z ~ ) Swill have to be negative,
leading to negative-sign terms in T N where they do not appear in
80 (assuming all signs of J’s the same, as is often but not always the case).
MnO is the classical case of this. In the fcc structure the nearest neighbors
are too numerous to be all antiparallel; theoretical O / T N ratios as large as
five occur.
Some “anomalous” situations arise which may help to illustrate the
variety of possible effects and the limitations of these theorems. One
might ask “Is the structure of lowest energy always a sublattice one?”
Clearly, it is always periodic; however, A. YoshimonS2has shown that in
the rutile lattice under certain circumstances a very long period, or even a
period not commensurate with any lattice translation, may occur when the
interaction with the nearest neighbor along the c axis is large. He argues
that MnOz actually has such a pattern, with a repeat distance of seven
lattice constants. Later, work by Kaplans and VillainMhas suggested that
this type of structure is a fairly general one, especially in metals. Kaplan
et ~ l . have,
7 ~ in particular, shown that the spiral structure is usually stabler
than the “angles” postulated by Yafet and Kitte186 in certain complex
circumstances for ferrite lattices. The conclusion to which one is driven
by their study of that not very complicated case is that there are no really
general characteristics of the spin arrangements except in the very simplest
situations. They find spirals, as well as Yafet-Kittel angles, in appropriate
regions. In some regions the patterns are as yet unknown and, no doubt,
even more complicated.
Another effect, calculable in molecular field theory, for which the
lattice of the fluorides provides a useful example, was discovered recently
81 G. W. Pratt, Jr., Phys. Rev. 108, 1233 (1957).
81 A. Yoshimori, J . Phys. SOC.Japan 14, 807 (1959).
a T. H. Kaplan, Phys. Rev. 116, 888 (1959).
S4 J. Viain, Phys. and C h .Solids 11, 303 (1959).

86 Y. Yafet and C. Kittel, Phys. Rev. 87, 290 (1952).


126 PHILIP W. ANDERSON

0
\

O 1 0

lo
0
0
FIG.3. Canting of sublattice spins in NiF2. Note that the axes of the F octahedra
are at 90"for the center relative to the corner sites. Ni: @; F: 0 .

by Dzialoshinsky.20,22(It appears that Oguchi*Bactually may have been the


first to suggest a phenomenon of this type.) I n cases in which the various
magnetic sites and ions are crystallographically equivalent, their sur-
roundings may be equivalent only after a space rotation. Then the presence
of anisotropy forces may make the sites effectively inequivalent and lead
to a ferrimagnetic moment. The classic example is Fez03;a simpler one by
far is NiF2, which has been analyzed by Moriya.868 (The first comment on
the application of this theory to the rutile structure was made by Dzialo-
shinsky.22)I n this lattice, the two inequivalent Ni sites per unit cell are
carried into each other by a 90" rotation about the c axis (see Fig. 3).
Each Ni site is surrounded by an irregular octahedron which has planes of
symmetry in the c, x, and y directions (x and y being at 45" to the a and b
axes) , so that the anisotropy is describable by the term
XA = D ( 2 S g 2- S,' - 5,') +E(S=' - S,'), (2.34)
for the body centers, but with x and y axes interchanged for the unit cell
corners. I n NiF2 it now appears that the antiparallel spins are in the (ab)
(ie., xy) plane ( D > 0). If then we were to add together the body centers
and cube corners and set S1 = - Sz, as for a normal antiferromagnet, we
would get XA constant in this plane. Thus the antiferromagnetic spins
are free to rotate in the plane.
86 T. Oguchi, private communication (1956).
*Q T. Moriya, Phys. Rar. 117, 635 (1960).
THEORY OF EXCHANGE IN INSULATORS 127

The basic point of Dzialoshinsky’s argument is that Sl= - S z is not


the real energy minimum because S1and - Sp need not be exactly parallel.
Introducing anisotropy fields in place of (2.34),we may write the total
fields on the two sets of spins as
HI = - l 6 J ( S z ) + 2 E ( 2 ( S z 1 ) - 9(S,1))
(2.35)
Hz = - l 6 J ( S 1 ) + 2E(f(S,z) - ~ ( A S ~ Z ) ) .
The equilibrium condition is that (S) be parallel to H,which, if O1 is the
direction of S1 relative to x and Oz that of Sz, gives
sin O2 - ( E / 8 J ) sin O1
tan O1 =
+ ( E / ~ Jcos
cos e2 ) el
sin el + ( E / 8 J ) sin e2
tan Oz =
cos el - ( E / SJ ) cos ep
These have two solutions. One corresponds to positions of unstable
equilibrium at O1 = 0, O2 = T and O1 = ~ / 2 O2, = - ~ / 2 ,with one spin
parallel and one antiparallel to its anisotropy field. The other corresponds to
stable positions near the 45” line, where
E
cot 20 - - = - cot 2e2
- 8J

(2.36)

5~ E
e2 N - + -;
4 16J
thus there is a magnetization of ( S ) E / S J , because the two sublattices are
canted to each other, each lying at slightly less than 90” to its own aniso-
tropy field. It is this small degree of accommodation to the anisotropy
which makes this position more stable.
This “Dzialoshinsky phenomenon” was explained by him on the basis
of rather complicated group theory, but the essence of it is, as in the above
example, as follows. Exchange alone makes the sublattice spins want to
set themselves antiparallel. However, the sites of one sublattice, while
equivalent to those on the other sublattice, may be equivalent only after
a rotation, reflection, or often a screw rotation or glide reflection, rather
than after a pure translation. I n such a case, if the anisotropy is such that
the magnetization has a component perpendicular to the reflection plane or
128 PHILIP W. ANDERSON

rotation axis, the anisotropy field will normally leave the sublattices canted
slightly to each other because this rotation or reflection will usually turn
the anisotropy figures of the two sublattices relative to each other.
This has been one of the most important applications of the general
group theory of magnetic crystals developed by Schubnikov87and others.*@I
This theory, no doubt, will eventually take on a basic role in the theory
of antiferromagnetic and ferrimagnetic patterns comparable to that of
the Bouckaert et U Z . ~ ’ paper on the space groups in crystallography-i.e., not
widely used in its full generality, but of great utility in the classification
and organization of background knowledge.
The basic concept of this theory is that one can expand the space
groups to include operations in which nonparallel spin vectors are carried
into each other. For instance, a true rotation by 180” does not reverse the
sign of the spin and so does not carry the + into the - spins in an anti-
ferromagnetic structure. If it connects antiparallel spins it must be supple-
mented by the time-reversal operation, a new operation of the magnetic
groups. A space inversion, on the other hand, or a reflection or glide plane,
does reverse a spin and must be multiplied by time reversal to connect
parallel spins. The number of possible space groups is thus much increased.
The Dzialoshinsky phenomenon appears when a truly antiparallel structure
turns out to be equivalent in symmetry to a canted structure-i.e., there
are no reasons based on symmetry for maintaining the antiparallel struc-
ture, except in the approximation in which time-reversal factors out of the
magnetic group entirely because exchange is the only force on the spins.
Many interesting properties of these “Dzialoshinsky ferrimagnets”
have been worked out by MoriyasBaand by Borovik-Romanov. Both
authors have also pointed out the possibility of piezomagnetism in a num-
ber of cases. This is a phenomenon deducible from general group theory
as well as from special models. CoF2, in particular, has been shown to have
the e f f e ~ t . ~ ~ * ~ ~
Piezomagnetism has some interest from the point of view of general
theory. By symmetry the true quantum ground state of the antiferro-
WA. V. Schubnikov, “Symmetry and Anti-Symmetry of Finite Figures.” Acad. Sci.
U.S.S.R.Press, Moscow, 1951.
88 L. D. Landau and E. M. Lifahitz, “Statistical Phyaica,” p. 129.State Tech. Libpress,
MOBCOW, 1951.
89 B. A. Tavger and V. M. Zaitaev, Zhur. Eksptl. i Teoret. Fiz. 30, 564 (1956);Soviet

Phys. JETP 3, 430 (1956).


90 N. V. Belov, N. N. Neronova, and T. S. Smirnova, Akad. Nauk S.S.S.R., Tr. Imt.

Krist. 2, 33-67 (1955).


9 1 L. P. Bouckaert, R. Smoluchomki, and E. Wigner, Phys. Rev. 60, 58 (1936).

T. Moriya, Phys. and Chem. Solids 11, 73 (1959).


98 A. S. Borovik-Romanov, Zhur. Eksptl. i Teoret. Fiz. 38, 1088 (1960).
THEORY OF EXCHANGE IN INSULATORS 129

magnet cannot be the polarized state normally assumed. Therefore one


supposes that the actual state is a very narrow degenerate multiplet, with
a total splitting of order times a characteristic energy. This implies
that the antiferromagnetic state is a wave packet which would break up
and “turn over” by quantum tunneling in some very long time. Neutron
diffraction sets a limit of -10-8 sec, and nuclear resonance perhaps 10”
sec., on this time. But piezomagnetism demonstrates that the antiferro-
magnetic state does not reverse during macroscopic times.
One point which should be mentioned in connection with the weak
ferromagnets is the mechanism of anisotropic exchange which was postu-
lated, by Dzialoshinsky empirically20022to explain some cases, particularly
Fez03 itself. Group theoretically, an energy of the form
Banma = D*(SiX S2>, (2.37)
is perfectly allowable in many asymmetric crystal structures. D is a con-
stant vector related to the local symmetry of the sites involved. The
essential requirement for such an anisotropic energy term is the absence
of a center of symmetry between the magnetic sites housing SIand Sp.
If such a center existed, SI and S2 would interchange in the operation of
inversion and S1 X SZ would reverse its sign. This anisotropic exchange
-
term D ( S1X S2) was explained by Moriyaea*on the basis of the effect of
spin-orbit coupling in the present theory of superexchange. It will be dis-
cussed later, in the section on superexchange.

IV. Spin Waves

The theory of spin waves and of Bloch walls and other dynamic effects
is, on the whole, less valuable in giving us information about exchange per
se than the molecular field theory, especially in the cases of ferri- and anti-
ferromagnetism. On the other hand, this subject is of basic importance
for several reasons. A review of exchange in insulators should cover the
fundamental points briefly.
It is often emphasized that the spin-wave theory is exact at low tem-
peratures in the Heisenberg model of a pure ferromagnetic, which pro-
vides an exact limit for comparing the various statistical theories at higher
temperatures. Unfortunately, this exactness is physically irrelevant, since
almost no substance in nature is known to be a pure Heisenberg ferro-
magnet. In the cases of antiferromagnetism and ferrimagnetism, no exact
theory exists even at T = 0 (the ground state). The spin wave theory is,
however, one of the most useful schemes for approximating the properties
T.Moriys, Phys. Rev. 120, 91 (1960).
130 PHILIP W. ANDERSON

of the ground and higher states. In the case of ferrimagnetism a theorem


of Herring and Kittelg4shows that the thermal properties are almost exact
at the lowest temperatures.
Fortunately, the review article of Nagamiya et aL3 covered spin waves
fairly extensively. More recently there has appeared the lucid and carefully
selective review of Van Kranendonk and Van V l e ~ k The . ~ ~latter review,
in particular, discusses rather carefully the questions of principle involved
in the theory of the antiferromagnetic ground state, so that we can bypass
all but the most recent developments in that field. Our primary purpose
here will be to derive briefly, by the simplest possible method, the most
useful results of spin wave theory, and to bring the reader up to date on
some of the more important aspects of spin wave theory developed since
the article of Van Kranendonk and Van Vleck was written.
In order to furnish a contrast with Van Kranendonk and Van Vleck,
the method used here to derive the main results of spin wave theory will be
the method of approximate equations of motion. (Other similar methods
are the random-phase approximation, the generalized Hartree-Fock method,
the new Tamm-Dancoff method.) A list of author-references in which
equations of motion have been applied to spin problems is Heller and
Kramerslg6HulthBn,17 Herring and KittellMBogolyubov and Tyablikovlg7
and many others.
As we discussed earlier, in the case of insulating materials it is ade-
quate to work entirely with the Heisenberg Hamiltonian,
X = -3 C J i j S i - S j - C gipSj*H + XAN, (IV.1)
i.i i
XANbeing a contribution representing the anisotropy energy, which can
be essential in the antiferromagnetic case.
Now we simply write down the equations of motion of the spin opera-
tors. The reason classical theory is so nearly satisfactory for spin waves is
that the quantum and classical equations of motion are identical, as one
can prove by means of the commutation rules; the equations of motion are

_ -- -i CX,Sil
dSi
dt A (IV.2)
= Hi X S i ;
Hi = C JiiSi + gjpBH + HAN. (IV.3)
6

94 C. Herring and C. Kittel, Phys. Rev. 81, 869 (1951).


96 J. Van Kranendonk and J. H. Van Vleck, Revs. Modern Phys. 30, 1 (1958).
G. Heller and H. A. Kramers, Proc. Roy. A d . Amsterdam 37, 378 (1934).
97 N. N. Bogolyubov and S.V. Tyablikov, Dokludy Akad. Nauk S.S.S.R. 126, 53 (1959);
Soviet Research in Phys. (English Translation) 4, 604 (1959).
THEORY OF EXCHANGE IN INSULATORS 131

It should be re-emphasized that Eqs. (IV.2) and (IV.3) are entirely


exact quantuni-mechanical equations. All too often the methods based on
them are stigmatized as “semiclassical” and as if they did not possess
the full rigor of the Slater,’Oo or harmonic-oscillator*O1techniques.
This is not so.
It is convenient now to specialize by considering the simple ferro-
magnetic case first, and then the antiferromagnetic and ferrimagnetic
cases. I n the ferromagnetic case we need no anisotropy field for most im-
portant results. For the two components perpendicular to H , we write

or, defining
s.
l+ = s. - is&
1% sj- = Sj,*) (IV.4)

In the ferromagnetic ground state, Eq. (IV.5) can be solved exactly


because every spin Si is parallel to the external field. Thus, sks*g = #*,.
If we apply Eq. (IV.5) as an operator expression to the ferromagnetic
ground state qg,

we get a linear equation in the operators Si, which can be diagonalized by


setting

(IV.7)

Then, after the various possible distances 6 between neighbors are intro-
duced,

98 F. Bloch, Z . Phyaik 61,206 (1930).


F. Bloch, Z . Physik 74,295 (1932).
100 J. C.Slater, Phys. Rev. 62, 198 (1937).

l01 T.Holstein and H. Primakoff] Phys. Rev. 68, 1098 (1940).


132 PHILIP W. ANDERSON

Equation (IV.8), which can be written

[Hi

implies that
f$+l\ko = [s c J6[1
6
- exp ( i 0 ' S ) l + gPEH
I Iqoi

190 = *,, (IV.9)


and that *,,is an exact excited state with energy
x\E, = (EN + Eo)*p (IV.10)
E, is the ground state energy and E,, is ,the excitation energy of the spin
wave of wave number p,
E, = S c Ja[l
6
- exp (iS.p)] +g p d . (IV.11)

For very low energies, p 'v 0. This then reduces to the usual p2 approxi-
mation:
EN c
Ja[S* pI2 -I- g p B H .
I
(IV. 12)

Equation ( I V . l l ) is of course exact only for a single spin wave in an


otherwise unperturbed ferromagnet. Once two or more spin waves are
present, they will scatter and interact. On the other hand, it often turns
out that Eq. (IV.ll) is a very good approximation at finite temperatures
because the interactions of spin waves are not very strong except at tem-
peratures of the order of magnitude of the Curie point. We shall see the
reason for this shortly. Thus, as we shall see in Part V, experimental
values of exchange integrals can be deduced fairly accurately from experi-
ments using results from spin-wave theory, such as the low-temperature
specificheatlo2or the magnetization curve,lm and even direct or nearly direct
measurements of spin-wave frequencies by resonance techniques or neutron
scattering.lo4
The fact that spin waves interact very weakly at low temperatures
was first demonstrated rigorously by Dyson,lMJoB although Kittel (private
communication) and others had pointed out that the circumstance that
exchange (and therefore spin-wave interactions) did not alter the ferro-
magnetic resonance frequency at all implied something of the kind. In
fact, it had become customary in relaxation calculations to neglect the
corresponding higher-order exchange terms even before Dyson's paper.
lm J. S. Kouvel, Phys. Rev. 102, 1489 (1956).
103 A. C. Gomard, V. Jaccarino, and J. P. Remeika, Phys. Rev. Letters 7 , 122 (1961).
B. N. Brockhouse, Phys. Rev. 106, 859 (1957).
106 F. J. Dyson, Phys. Rev. 102, 1217 (1956).
106 F. J. Dyson, Phys. Rev. 102, 1230 (1956).
THEORY OF EXCHANGE IN INSULATORS 133

The implication is that since the ferromagnetic resonance formally is


simply the spin wave of longest wavelength, one expects, by continuity
arguments, other long wavelength spin waves to behave similarly and thus
to be scattered little by other spin waves.
This fact can be demonstrated very straightforwardly in our formalism
using Eq. (IV.5). In that equation, the operators Sj, and S k r cannot be
replaced by exactly the numerical value S except at absolute zero. On the
other hand, at moderately low temperatures they will on the average be
fairly close to S ; thus we may attempt an operator expansion in terms of
the spin-wave variables S,, and Sp-.
Starting from the identity
s:..+ s:.z + = S(S + l), (IV.13)
we observe that
Sj+Sj- = Sz! + S$u - i(SjuSjz - SjsSjU),
so that
S$q + S$g = Sj+Sj- + S,. (IV. 14)
(Again we use the commutation rules for spin operators S X S = is.)
Substituting in Eq. (IV.13), we get
SX + Sjs = 5 2 + S - (SfiSj-). (IV.15)
The two Hermitian conjugate operators Si+ and Sj- play the role of
creation and destruction operators for “spin deviations.” That is, the
operator Si- acting on the state in which Si. = S is exactly zero, while
from the commutation rules it may be shown that
~j++(Sjz = S) = m+(Sj. = s - 1)
(IV.16)
Sj-+(Sjz = S - 1) = m + ( S j z = s>.
Thus we may take as normalized “spin deviation operators,”
ai+ = (2S)-1/zSi+
(IV.17)
aj- = (2S)-’/2Sj-.
In states near the ground state, in which all spins are nearIy paralleI and
Sj, S, the “number of spin deviations on sitej” is
ni = aj+aj- = (1/2S)Sj+Sj-. (IV.18)
This may be verified from Eq. (IV.l5), since if we substitute
Sj, S - ni = S - (Si+Sj-/aS), (N.19)
that equation reduces to
-2Snj +
n? - ni = - Sf+Si-.
134 PHILIP W. ANDERSON

Now in the two states ni = 0 and 1 of small total deviation nr = n,;


thus we obtain Eq. (IV.18).
We may expand Eq. (IV.5) in terms of spin-wave operators S,, as
defined by Eq. (IV.7) and
s,- = N-112 e-"r.iS.1--' (IV.20)
i

It is more convenient in fact to deal with normalized operators [see


(IV.17)]
a& = (2s)-'12s&) a, = (ZS)-'12S,-. (IV.21)
Then we obtain
hUj+ = i
k
Jjk[(S - ak+ak-)aj+ - ak+(S - aj+aj-)] + ig/JBHaj+.
Note that the additional terms in this expression have the form
(IV.22)

which tells us immediately that these terms are small whenever only spin
waves of long wavelength are excited, because the difference of aj- and
ak- must be small in that case.
Now we can Fourier-analyze the extra terms (IV.22) and obtain

(IV.23)
To obtain a first approximation to the properties of the spin waves
in the presence of other thermally excited spin waves, we "linearize" Eq.
(IV.23). This means that we assume that it acts on a state with a fixed
small average number of spin waves present. We find

(IV.24)

the last equality being obtained by assuming that the ax's are at least
approximately a set of independent boson operators. A discussion of this
assumption is given later. We then obtain for the extra terms
THEORY OF EXCHANGE IN INSULATORS 135

where a lattice with the atoms at centers of symmetry (allowing us to


sum over &a) has been assumed. We see immediately that the spin-wave
energy varies exactly as it would if we replaced the exchange integral Ja by

(IV.26)

which in any simple structure means that the spin-wave frequencyvaries as

We shall see that cxnx( T ) varies as T3I2;the extra factor 1 - cos 2 - 6


adds an extra power of T,giving the Dyson factor of T4for the spin-wave
scattering corrections.
It is important to realize that no computation as precise as the foregoing
has been carried through for the more realistic cases of ferri- or antiferro-
magnets. As we remarked earlier, there are practically no pure Heisenberg
ferromagnets. In each case, we expect some such theorem because, again,
the spin wave of longest wavelength must remain unaffected by exchange.
Thus there remains a real problem here to justify the use of spin waves in
the ferri- and antiferromagnetic cases. Although this problem probably
lies within the reach of the techniques of modern many-body theory it has
not yet been solved. (Note, however, that suggestive, if not conclusive,
arguments indicating the same kind of result in the antiferromagnetic case
have been given by Oguchilo7and Keffer and Loudon.lo8Their result for
the sublattice magnetization is a small correction in T6 added to the spin-
wave result a T2.) As was evident from the early work of Van Kranen-
donklO*lll and Schafroth,llz in the absence of the Dyson cancellation factor
[the coefficient (1 - cos 3c.b) in Eq. (IV.26)] the corrections are really
appreciable and the use of standard spin-wave theory comes seriously
into question.
The history of the Dyson correction has been rather complicated.
Schafroth112and Van Kranendonk,lo*lll had treated spin waves as particles
interacting with hard core repulsions, an approximation in the basic
method of Holstein-Primakofflolwhich appeared at the time to be justi-
fiable, and which gave large interactions with a strong effect at low tem-
107 T. Oguchi, Phys. Rev. 117, 117 (1960).
108 F.Keffer and R. Loudon, J . Appl . Phys. 32, Suppl., 2s (1961).
109 J. Van Kranendonk, Physica 21, 81 (1955).

110 J. Van Kranendonk, Physicu 21, 749 (1955).

111 J. Van Kranendonk, Physicu 21, 925 (1955).

112 M.R. Schafroth, Proc. Phys. SOC. (London)A67, 33 (1954).


136 PHILIP W. ANDERSON

peratures. The somewhat different calculation of Kramers113 and


OpechewskP4 gave even greater corrections. Dyson's papers then ap-
peared.laJo6 They were perhaps some years ahead of their time when
regarded as applications of field-theoretic techniques to many-body
theory. They modified the Holstein-Primakoff theory in a direction whose
importance we will discuss shortly, and demonstrated the weakness of the
interactions for long spin waves. It was soon shown, however (Suhl and
Walker, private communication; Yosida, private communication;
Oguchi107)that the standard Holstein-Primakoff theory used until then
was also perfectly correct on the T4term-i.e., in understanding the weak-
ness of the interactions-and that the Van Kranendonk-Schafroth errors
were the result of approximations within a basically correct theory.
More recently, Brout et aZ.11b117and Bogolyubov and Tyablikovg7J1*
have reopened the Ta vs T4 problem, which led to the work by Keffer
and Loudon.lOBMoriya and Herring also demonstrated the cancellation of
T3 terms from a physical point of view. The treatment used here is similar
to that of Moriya.
It is not in the province of this review to discuss exhaustively the
calculation of the properties of magnetic materials in the spin-wave ap-
proximation. It has been treated well in the papers of Nagamiya et aL8
and Van Kranendonk and Van Vleck.g6It is sufficient to say that the
usual approximation simply involves the observation that the operators
aj+ and aj- are, as far at least as the two lowest states S, = S and S - 1
are concerned, very similar to Bose-Einstein operators; thus it is assumed
that the spin waves, which are derived from them by a unitary transforma-
tion, are also bosom. One then uses the boson number distribution Eq.
(IV.24) to give the number of excited spin waves of wave number 1.For
instance,

(IV.28)
at low temperatures. The average spin-wave energy is
cx
Ex(%.)a (IV.29)

118 H. A. Kramera, Communs. Kamerlingh Onnes Lab. Univ. Leiden 22, Suppl. 83 (1936).

u4W. Opechowski, Physica 4, 715 (1937).


116 R. Brout and H. Haken, BUU. Am. Phys. SOC. 6, 148 (1960).
116 R. Brout and F. Englert, Bull. Am. Phys. SOC.6 , 55 (1961).
117 F. Englert, Phys. Rev. Letters 6, 102 (1960).
118 S. V. Tyablikov, Ukrain. Mat. Zhur. 11, 287 (1959).
THEORY OF EXCHANGE IN INSULATORS 137

giving a specific heat proportional to TaI2.As we see, Eq. (IV.29) also


gives the modiflcation Eq. (IV.27) in the spin-wave frequencies, leading
to the T4correction to M ( T ) .
The basic advance on the Holstein-Primakoff lol method achieved in
Dyson’s work was not the absence of the Ta term-which as we saw above
may be understood quite simply-but the presentation of a new and far
more rigorous method for introducing boson operators for spinors. The
Holstein-Primakoff method uses the rather complicated relationship
ai+ = [l - (mi/2S)]1’2bj+, (IV.30)
where the ai+ is the operator defined in Eq. (IV.17) and bj+ may be shown
to be rigorously a boson operator. Equation (IV.30) must then be expanded
in power series in writing the Hamiltonian. It is extremely hard to estimate
the errors which come from its use. Whereas nj has of course only 2 s + 1
eigenstates, mi = bj+bi has an infinity, which are separated from the
real 2 s+ 1 states by the square-root factor, which is zero at a certain
point in the spectrum.
Dyson, on the other hand, using a somewhat different definition of the
bi, was able to prove rigorously that the resulting errors are at worst
exponential in -1/T. His rigorous proof that it is feasible to replace
spinors by bosons is unique in many-body theory. Unfortunately, it is
too complex to reproduce here. We can, however, demonstrate the Dyson
substitution in the case S = 3. It is
8z = l a- m j (IV.3la)
aj+ = b t ( 1 - mi) (IV.31b)

ai = (1 - mj)bj.v bi, (1V.31~)


the last equality being valid in physical (mi = 0 or 1) states only, which
are the only ones with which Eqs. (IV.3la) and (IV.3lb) have any con-
nection. Substitution of Eq. (IV.31) into the Hamiltonian and equations
of motion gives exactly Dyson’s theory. It is interesting that Eq. (IV.31)
may also be used in ferri- and antiferromagnetism, but that the resulting
Hamiltonian and equations of motion do not simplify to any great degree.
We will mention briefly a few of the latest developments in antiferro-
and ferrimagnetic spin-wave theory. The reader will remember that the
basic difference between ferromagnetism and the other forms is that in the
latter cases the “naive” ground state in which each spin is aligned exactly
along an appropriate molecular field is not exact in any sense. One way to
present this* in the case of an antiferromagnet, for example, is to rewrite
the Hamiltonian reversing the alternate spins by a canonical transformation
138 PHIJAP W. ANDERSON

to the direction into which one expects them to point. I n order to preserve
the standard commutation rules, the rotation must be a proper one:
sublattice A : unchanged
sublattice B : S, + S,
s,+ -s,
s, + - s,; (IV.32)
i.e.,

(IV.32’)

Now the Hamiltonian of a simple antiferromagnet becomes


J
x = -- C C [ S i z S j z
2 ionA jonB
+ +(Si+Sj+ + Si-Sj-)], (IV.33)

with terms SSi, which create pairs of spin deviations, rather than SiS-,
as in the ferromagnetic case. At the same time, the equations of motion
now effectively couple the creation and destruction operators for spin
deviations, whereas only creation operators were coupled together in (IV.5) :
dsj+
h- =iJ C (SkzSj, -I- S k - s j z ) . (IV.34)
dt neighbors

The standard approximate solution of Eq. (IV.34) is obtained by


substituting S for Sj, and S k z , which is, of course, only an approximation,
since the true ground state is not known and certainly does not obey the
relation S , 1 9Q)= S [ qg).Then one obtains for the frequency of the
spin wave of wave number k

fiwk = ( C
ne ighhors
J S ) W j (IV.35)

where
1
Yk = - C exp (z11-8). (IV.36)
neighbors

[S is defined as the neighbor distance as in (IVB).] This spectrum is


linear in k at low frequencies, like the phonon spectrum. Hence the thermal
properties of antiferromagnets have a certain resemblance to those of the
lattice vibrations of a solid.
THEORY OF EXCHANGE IN INSULATORS 139

The operators which make up the normal modes of spin-waves are


even more different from those in the ferromagnet than the spectrum of
energies (IV.35). If we define

A,+ = (g) 112

ion A
exp (ip-Rj)uj+

(IV.37)

in which the ai+-operators are defined as in Eq. (IV.17) but with the
reversal Eq. (IV.32) on sublattice B, the normal modes are given by
operators x, and y, defined by
A,+ = + y,, sinh 8,
x,,+ cosh 8,
(IV.38)
B,+ = x, sinh 8, + yr+ cosh 8,,
where
tanh 28, = -7,. (IV.39)
Note that as p + 0, 7, + 1, 8, 00 and the amplitudes of the spin
---f

waves diverge. Again, this phenomenon is closely analogous to the behavior


of lattice vibrations, and is a consequence of rotational symmetry in spin
space, just as the divergent amplitude of lattice vibrations is of the transla-
tional symmetry in real space. Since all real physical systems do not have
true rotational symmetry in spin space, but instead have a certain amount
of crystal anisotropy, this divergence, and the linear portion of the fre-
quency spectrum, are removed by the anisotropy. The result is the
“anisotropy gap” in the spectrum, of order liw,
change) ‘I2.
-
(Anisotropy X ex-

In terms of the x, and y, operators, one may write the energy as

x % -$NZJS(S -k 1) + k
fiwk
- (xk*xk
2
+ xkxk* + Yk*Yk + ykyk*)

(IV.40)
= -$NZJS(S + 1) + x b k ( n k
k
+ a).
The noteworthy thing about Eq. (IV.40), as well as the corresponding
expression for the magnetization of the sublattice in terms of spin waves,
is that each term contains a zero point contribution (the 3 in the last
term) which represents an approximate value of the energy and other
140 PHILIP W. ANDERSON

properties of the ground state itself. The importance of zero-point effects


for spin waves was pointed out by Klein and SmithllBand Anderson.24
Very similar results can be obtained for ferrimagnets. In that case the
two degenerate branches of the frequency spectrum of the antiferromag-
netic case split up, one becoming the normal k2 spin-wave spectrum, the
other an optical branch which is high or low in frequency depending on
how closely compensated the ferrimagnet is. Kasuya and LeCrawlZ0
have recently shown that this second branch of the ferrimagnetic spin-
wave spectrum can have important consequences for ferrimagnetic relaxa-
tion.
One substance on which the antiferromagnetic spin-wave theory has
recently been tested rather severely is MnF2. From the experimental side,
one can obtain an extremely accurate curve for the variation of sublattice
magnetization with temperature by measuring the resonance frequency of
the F nuclear spins in the hyperfine field caused by the Mn spins.4gExisting
measurements of antiferromagnetic resonanceK8 give one an accurate meas-
ure of the anisotropy gap, while the anisotropy energy itself, since it must
be almost entirely dipolar, is fairly well known, as are the exchange con-
stants as a result of pair measurements on Mn in ZnF2 (see the careful
analysis by Smart4). The predicted dependence obtained from spin-wave
theory is of the form
Msublattiae(0) - M(2') a !P2
exp ( - f i w , / k T ) . (IV.41)
The constants in Eq. (IV.41) are all known accurately. Under these
circumstances, it was disconcerting to find that in fact the observed curve
was not well fitted by Eq. (IV.41) even with empirically adjusted con-

Recently, however, Keffer (private communication) has made calcula-


tions taking into account the higher-order terms originating in short
wavelength spin waves, which he finds are anomalously large in the anti-
ferromagnetic case. The result is to give a very exact confirmation of the
results of spin-wave theory, at least in the lowest-temperature region
where its validity is most likely.
A second experimental confirmation of a result of spin-wave theory
has been obtained recently by Walker121and Geschwind.lZ2As we pointed
out, there is a contribution both to the energy and to the sublattice mag-
netization from zero-point fluctuations in spin direction, caused essentially
l19 M. J. Klein and R. S. Smith, Phys. Rev. 80, 1111 (1950).
uoT.Kasuya and R. C. LeCraw, Phys. Rev. Letters 6, 223 (1961).
lZ1 L. R. Walker, J . Appl. Phys. 32, Suppl., 2648 (1961).
Izz S. Geschwind, J. A w l . Phy8. 32, Suppl., 2633 (1961).
THEORY OF EXCHANQE IN INSULATORS 141

by the spin-deviation pair creation terms in the Hamiltonian Eq. (IV.33).


Such a zero-point term also affects the observed anisotropy energy of a
ferri- or antiferromagnet. In fact if, as in the cubic case, the relevant
anisotropy comes from an energy term depending very sharply on angles,
the zero-point fluctuation effect can be large.
The cubic anisotropy of a single Fe3+ ion in the surroundings it has in
the garnet lattice is well known from measurements of Geschwind.122On a
naive theory, this one-ion anisotropy should lead to about twice the
ferromagnetic crystal anisotropy actually observed in Y - Fe garnets.
Walker121 has shown that this discrepancy can be reasonably well ac-
counted for on the basis of zero-point fluctuations of the sort discussed
above.
Another, and apparently more direct, way to evaluate the ground-
state energy and sublattice magnetization of an antiferromagnet is by
straightforward perturbation theory, based on the “naive” molecular
field ground state as a starting point. That is, one can consider the first term
of Eq. (IV.33) as a zeroth-order Hamiltonian XO, diagonalized by product
states of the various spin functions, and calculate with perturbation
theory, using the second two terms aa X1. Different versions of such a
perturbation theory have been given by Mills et aZ.,12‘ and by
Walker (private communication). The series thus obtained appear, from
an evaluation of their earlier terms, to be converging rather well. They
give what appears to be a fairly satisfactory answer for the energy of the
ground state. On the other hand, the agreement with spin-wave theory
on sublattice magnetization is rather poor.
It was clear, however, that there is really a basic difference between the
two methods, and in fact, that perturbation theory, even though it may
give satisfactory answers, cannot be convergent, especially for the isotropic
antiferromagnet. The reason is that the actual analytic behavior of the
result of spin-wave theory and of perturbation theory, expressed say as a
function of a parameter X multiplying XI, are entirely different. Spin-
wave theory gives nonanalytic results in X as X + 1. But this nonanalytic
behavior is an essential consequence of the spin-rotation invariance and
must be contained in a true theory.
Recently, Mills et al.124 have demonstrated that spin-wave theory can
actually be recovered by summing a restricted (but divergent) subclass of
diagrams in the perturbation theory. This, of course, demonstrates that
spin-wave theory is indeed the more accurate, but does not prove anything
about its essential validity.
118 H. L. Davis, Phys. Rev. 120, 789 (1960).
114 R.L. Mills, R.P.Kenan, and J. Korringa, Physics 26 (Suppl.), 5204 (1960).
142 PHILIP W. ANDERSON

We may conclude, then, by remarking that in the antiferromagnetic


and ferrimagnetic cases, spin-wave theory has had a number of experi-
mental successes, and seems by all odds to be the most promising theoretical
approach to an understanding of the ground state and low-temperature
properties. However, the method needs support from the fundamental
theoretical point of view, particularly in showing that the effects of inter-
action of long-wavelength antiferromagnetic spin waves are negligible as
in the ferromagnetic case.

V. Empirical Values of Exchange Constants in Insulators

This chapter will be a compilation of some of the experimental data


on exchange constants in insulators. Fortunately for us, J. S. Smart has
recently made a very complete and careful compilation of such data. The
greater part of this chapter will be taken almost verbatim from his work.4
We are very grateful to him for his kind permission to use his article in
this way prior to publication.
In Smart’s analysis, the Heisenberg Hamiltonian is given in the same
form as our Eq. (11.1) ; thus the definition of J is the same as in that
article and in Anderson.70 The Nee1 or Curie temperature with a Hamil-
tonian like Eq. (11.1) having a single interaction is given by

and the molecular field ground state energy by

N-’E, = -2 I J 1 S2. (V.2)


In view of the appearance in Eqs. (V.l) and (V.2) of the quantity J S 2
or J S ( S+ l ) , the former being essentially half the energy involved in
reversing one spin relative to another, it seems physically more meaning-
ful to tabulate that energy,
EJ = I J I S2, (V.3)
a
which is the quantity called Jeff in Anderson.” We will tabulate both
numbers, J and EJ. EJ is more strictly comparable from one ion to another
if, as is not unreasonable in many cases, we assume that most of the ex-
change comes from one or perhaps a few of the 25 possible overlapping
pairs of the five d-functions on each ion.
THEORY OF EXCHANGE IN INSULATORS 143

We now give what is essentially Smart’s list of the experimental effects


he uses to obtain the data; most have been discussed fully in previous
chapters so we will content ourselves with a brief discussion of relevant
types of possible errors in each measurement.
(1) Weiss temperature e [Eq. (2.29)]. The formulae are exact, but
unfortunately true only in the high-temperature limit which is seldom
attained experimentally, and occasionally obscured by temperature-inde-
pendent susceptibility terms. Thus, this is often the least rather than the
most reliable method.
(2) Molecular field analysis. ( a ) N b l point TN [Eq. (2.12) etc.]:
Experimentally, this can be and sometimes is determined with great ac-
curacy, by measuring almost any property but the susceptibility, which
however is, the quantity measured most often. Except for questions of
statistical theory, which usually lead to only l0-20% errors, this is
theoretically sound except in a few isolated cases in which the antiferro-
magnetic structure is in a state near a basic instability. Unfortunately, two
such cases are very important: the chloroiridates, which are close to the
unstable case of the fcc structure with nearest neighbor interaction
(we do not quote TN values for this reason) ;and MnO, which is near the
transition from the second to the third kind of order for the fcc, a point
where the disorder actually has finite e n t r ~ p y ’ ~and~ J where
~ ~ therefore
errors of factors of two or more are plausible. A few cases, such as COO,
contain large crystal-field effects, which K a n a m ~ r i ’ ~has ~ J ~discussed
~
on the basis of molecular field theory.
( b ) More complete molecular jield analyses: Here one usually studies
the full temperature range of the susceptibiIity. In the case of ferrimagnets
in particular, this often gives almost complete data on the three assumed
parameters ( JAA,JBB,J A B ) However,
. Smit and WijnI3Opoint out that the
results for the internal exchange on the sublattice of smaller magnetization
are quite unreliable in the molecular field theory. We will mark as question-
able results for which this is true. There is also a strong question, when
(2.12) alone is used, whether the smaller exchange integral is not usually
greatly overestimated.
125 Y.-Y. Li, Phys. Rev. 84, 721 (1951).
126 F. Stern, Antiferromagnetism in face-centered cubic lattices. Unpublished memo-
randum, 1953.
127 F. Stern, Phys. Rev. 94, 1412 (1954).
128 J. Kanamori, Progr. Theoret. Phys. (Kyoto) 17, 177 (1957).

129 J. Kanamori, Progr. Theoret. Phys. (Kyoto) 17, 190 (1957).


130 J. Smit and H. P. J. Wijn, “Ferrites,” pp. 160-162. Wiley, New York, 1956.
144 PHILIP W. ANDERSON

(3) More refined statistical analyses. ( a ) x( TN): Susceptibility at or


near the N&l point. Here Smart takes advantage of an extension of the
statistical theories by Danielian and Stevens.131 These authors have ex-
tended the high-temperature series expansion for the susceptibility to near
the NCel point. In some cases Smart also uses his own applicationlaaof the
Bethe-Peierls-Weiss method to an analysis of x near the N&I point in
order to estimate exchange for a number of substances. Both methods
require a nearest-neighbor-only assumption which apparently is the only
severe limitation.
( b ) Magnetic energy and entropy near the Niel point: This is beset more
by possible statistical theory errors than’ ( a ) .
(4) “Exchange field” experiments. ( a ) Antiferromagnetic resonance :
Here the difficulty is that one needs a good value of the anisotropy energy
to get J ; this has led to reliable results only,in the case of MnF2. The
Faraday effect falls in much the same class.
( b ) High-field magnetization: This would be useful if more measurements
on really interesting compounds were available. In actual fact, rather a
large mass of old data is available for quite dilute materials because of the
early interest in using such materials for adiabatic cooling, but there seems
little point in compiling this work because the main interest in such com-
pounds does not lie in their exchange effects.
( 5 ) Spin-wave efects. ( a ) Low-temperature magnetization and specific
heat: Since the connection (Part IV) between these quantities and exchange
is exceedingly direct, this kind of measurement is ideally suited for the
determination of exchange constants. Unfortunately, it is reasonably
difficult to make accurate enough measurements.
( b ) Inelastic neutron scattering: Where available, this gives the spin-
wave spectrum and thus quite a rigorous value for exchange constants.
( c ) Spin-wave magnetic resoname measurements: These are rare but,
when available, very sound and direct.
(6) Exchange clusters. Probably the most unequivocal measurement
which is available in any large number of cases is the magnetic resonance
or susceptibility study of exchange clusters. When two or more ions are
effectively isolated, by dilution or by the natural crystal structure, from
other magnetic atoms in the crystal, their susceptibility and magnetic
resonance behavior can be calculated in full detail. Comparison with
experiment yields values not only for the main isotropic exchange integral
but also for small anisotropic components and other fine details. Notable
1*1 A. Danielian and K. W. H. Stevens, Proc. Phys. SOC.(London)A77, 124 (1961).
la J. S. Smart, Phys. and C h . Solids 11, 97 (1959).
THEORY OF EXCHANGE IN INSULATORS 145

examples of such studies are the series on the chloroiridateb and meas-
urements on Mn in ZnO and ZnFzfor comparison with MnO and MnFz.lSJS4
(7) Linebreadths. A final type of measurement-which is not, in fact,
listed by Smart-is the measurement of exchange broadening or narrowing
in paramagnetic resonance.
This has been applied to a rather limited number of compounds of
magnetic interest, (see particularly Anderson and we is^^^ for some com-
parison with Method 1)) but in no case has the crystal structure and
arrangement been studied accurately enough to determine more than an
average of some unspecified number of exchange parameters, so we cannot
include any results from linebreath measurements in the tables.
In commenting in general on the following tables, we should say that
a very large amount of data relevant to exchange interactions has been
omitted. The reason is that it was felt important to concentrate on a few
cases in which one could obtain rather definite, accurate numbers, and in
addition on those cases in which there was very general interest for his-
torical or technical reasons, such as MnF2, NiO, MnO, Fea04,YIG, etc.
For instance, there is a large literature on molecular field analyses of
ferrites, which while of great value is not terribly accurate for evaluating
all exchange parameters for reasons explained by Smit and Wijn.IaoWe
have therefore concentrated on a few typical cases, and in fact, it appears
that the important tetrahedral-octahedral F e F e interaction JT--Odoes
not really vary enough to justify any extensive lists.
In a number of other cases we have left out inaccurate data such as
rough estimates from exchange narrowing or 0 values, in cases in which
the crystal structure does not appear to require the presence of only one
or a few interactions. Moreover in most cases we have not attempted to
evaluate exchange parameters ourselves from papers in which published
values are not given (except see Marshall134*).
Even with such limitations there remains an adequate selection of
exchange parameter values which are sufliciently well authenticated to
satisfy the theorist attempting to compute exchange from &-st principles
for many years. On the whole the results show a number of regularities
which will be discussed later in this paper, as well as occasional interesting
irregularities.
(Note: Since the experimental sources are not all relevant to the general
subject of this paper, they are given separately in a bibliography at the
end of Tables I-VI.)
1s B. A. Cola, J. W. Orton, and J. Owen, Phys. Rev. L.etters 4, 116 (1960).
11 J. Owen, J. Appl. Phys. 32, Suppl., 2135 (1961).
W. Marshall, Phys. and Chem. Solids 7 , 159 (1958).
TABLE
I. COMPLEX ACETATES
AND OTHER CLUSTER hfATERIALS

Compound Ion pair J (OK) EJ (“I0 Method References“

Fe$+-Fe,+ -39 -244 6 (susc.) \


Fea+-Fes+
Crs+-Cr3+
-41
- 14
-256
-32
6 (susc.)
6 (susc.) I 1s
Cr3+-Cr3+ - 17 -43 6 (susc.)
cu+-cu+ -205 -51 6 (susc., resonance) 4-8

-
-200 to 240 -50 to 60 6 (susc., resonance)
-
cu+-cu+ 9-1 1
cu+-cu+ -385 -96 6 (susc., resonance) 18
TCNQ -200 -50 6 (susc., resonance) 1% 14

a L = various.
b TCNQ = tetracyanoquinodimethane.
c References for Tables I-VI will be found on pages 152 and 153.

TABLE
11. EXCHANGE (fcc STRUCTURE)
IN CHLOROIRIDATES

J resonance (6b) EJ J9(l) EJ Anisotropy of


Compound Ion pair (OK) (“K) (“I0 (“K) J(Jz/J) References

KdrCb (IrCI,) ++ -5.8 -1.7 -5.4 -1.6 0.11 16


(NHdJrCls (IrCb)++ -3.8 -0.95 -3.4 -0.85 0.16 16
TABLE
111. EXCHANQE ANTIFERROMAQNETS
IN SIMPLE WITH NEAREST-NEIQHBOR
INTERACTIONS
ONLY

Compound Structure Ion Pair J ("K) EJ ("K) Method References Notes

pMnS Cubic sincblende Mn++ -12.4


-77 13.
&MnS Hex. sincblende Mn++ -10.5 -67
LaFe03 Perovskite Fe -26.3
-28.8
- 164
-180
3a(TN)
3a(x)
} 18, 19

NdFe03 Perovskite Fe3+ -27.0 -168 3a(T") 18,80

HoFe03 Perovskite Fez+ -24.8 -155 3a(T~) 18, 80

ErFe03 Perovskite Fe$+ -22.0 -137 3a(T~) 18, 80

LaCr01 Perovskite Cr3+ -26.9 -61 3a(T~)


-30.8 -69 3a(x)

CaMn03 Perovskite Mn4+ -9.2 -20.8 3a(T~) 18,81

KMnFa Perovskite Mn++ -3.1 -19 34TN) ), 18,62

'
-3.6 -22 3a(x(~N)) i
KFeF3 Perovskite Fe++ -6.0 -24 3a(T~)
18, 23,24
-6.0 -24 3a(x(T~)) 1
KC OF^ Perovskite Co++ -9.6 -22 3a(~N) 1 18,82

'
-11.8 -27 3a(x(T~)) f
KNiF, Perovskite Ni++ -44.5
x

'
-44.5 3a(TN) 18, 82, 83 Corrected for const.
-45.7 -45.7 3a(x(TN)) J
CrF3 wo1 Cr3+ -6.7 -15 3a(T~)
18,86,86
-7.0 -16 3a(x(7")) J
c
TABLE I N SIMPLE ANTIFERROMAGNETS
111. EXCHANGE WITH NEARESFNEIGEBOR
INTERACTIONS omY-&ntinued b P
00

Compound Structure Ion Pair J (OK) E J (OK) Method References Notes

FeFa WOa Fea+ -14 -88 342”) 18,26


CoFi WOa coa+ -23 18,26
MoFa wo3 Moa+ -15.6 18,bY
CrFe Rutile Cr++ -1.93 18,98
MnFe Rutile Mn++ -1.70 18,99 Assuming other J’s (“Ji’) ~d
-1.85 18,$9 negligible, proved by !
r
CI
consistency cd
-1.73 -10.8 3b(C~) 30 Method 2b(b9),
-1.72 -10.7 i(e)
.. 99 Method 6(32); ?
-1.74
-1.78
-lo*’
-11.1
31 Probably Jt <~ J I

FeFe Rutile Fe++ -2.88 -11.5 3a(TN) 33 g


-3.03 -12.1 3a(x) 33 \Consistency plus method
-2.75 -11.0 3a(x) 34
-3.01 -12.0 3b(Csp) 30
NiFI Rutile Ni++ -8.1 -8.1 3a(T~) 36
-8.7 -8.7 3b(C~p) 30
TABLE (J1)AND NEX-NEAEESTNEIGHBOR
IV. NEAREB-NEIGHBOR (Js) EXCHANGE
INTEGRALS
IN FACE-CENTERED
CUBIC
MAGNETIC
hmcm
iM;l
Compound Ion J1 ("K) E J I (OK) J, ("K) E J (~OK) Method References Notes 0
a
4
MnO Mn++ -7.2?
- 14
-45
-87
-3.5?
- 14
-22
-87
(2b)
(6b)
36-58
39
m f t a not reliable s
d
FeO
COO
NiO
Fe++
co++
Ni++
-7.8
-6.9
-50
-31
-15.5
-50
-8.2
-21.6
-85
-33
-49
-85
(2b)

gk! g:$:
56,40
corr. for crystal field); corr. for con-
{stant X ; mft reliable?
E
M

aMnS Nn++ -4.4 -27 -4.5 -28 (2b) 4 4 46 2


pMhS Mn++ -10.5 -66 -7.2 -45 (2b) 46 \? mft probably not reliable
MnSI Mn++ -5.6 -35 -5.9 -37 (2b) 47,@ I
H
MnTa Mn* -7.4 -46 -1.2 -7.5 (2b) 47,48 0
c!i
Molecular field theory.
c
cn
0

V. FERRIMAGNETS
TABLE

JTair.-TLtr. Joci.-oet. JT<L~.-O~~. EJ(TO)


Compound Structure Ion Pair (OK) (OK) ("W (OK) Method References
a
L ~ o . ~ ~ F ~ z . ~ s O ~Spinel Fe3+-Fe3+ -19.5 -10.6 -24.0 -150 2 49 s
FeaO4 Spinel Fe3+-2+-Fe3+-2, -17.7 4-0.5 -23.4 - 146 2 60 a*
-23 - 144 (5b) 61, 62 2
- d

:
COA1204 Spinel Co++-Co++ -3.5 - .. - 1 63 M

-
5FeZ033Y203 Garnet Fe3+-Fe3+ -15.1 -8.3 -35.0 -219 2 54
1:
JTT+JOOreduced 5a 66,67
-
5Fe& 3Lu?03 Garnet Fe 3+-Fe3+ -15.1 -8.3 -35.0 -219 2 55
VI. MISCELLANEOUS
TABLE CASESOF INTEREST

Compound Structure Ion pairs Exchange interactions Method References

azo3 NzOa Cr3+ Cr3+-Cr3+nearest neighbors: J = -180", Pairs: 6b 68


EJ = -405"
0
Mn02 Rutile Mn4+ Ji = -9.5', Eji = -21 2 (spiral structure) 69 id
4
Jz (along c ) = -15", E J ~= -34
J3(1C) = -So, E J ~= -11

CrOz Rutile Cr 4+ 4J1 + JZ + 2J3 = -I-150" (ferromagnetic!)


6
MnF3 Layer Mn3+ JI(layer) = 1.7', E J I = 6.8" 2 26 M
!2
4
J2 (out of layer) = -2.6", Ejz = -10.4"

CuC12.2H20 Rhombic complicated Cu++ J along c (nearest neighbor) N -20"; EJ = -5" 3 (special) 61
E5
J l , (next nearest) = +3", EJ = +0.75" %0
%
CrBs Hexagonal layer Cr3+ J t (inlayer) = +5.44", E j t = +12.24"
JI (across layers) = +0.85", E J Z= +1.98" 5a 6.2
ferromagnetic!
Jt + ( u 1 / 3 )= 6.26" 1 65
152 PHILIP W. ANDERSON

FOR TABLES
REFERENCES I-VI
1. L. A. Welo, Phil. Mag. 6, 481 (1928).
2. L. A. Welo, Phys. Rev. 32, 320 (1928).
3. K. Kambe, J. Phys. SOC.Japan 6, 48 (1950).
4. B. C. Guha, Proc. Roy. SOC.6206, 353 (1951).
5. B. Bleaney and K. D. Bowers, Phil. Mag. 43, 372 (1952).
6. M. Foex, T. Karantassis, and N. Per&, Compt. rend. mud. sci. (Paris)237, 982
(1953).
7. N. Perakis, A. Serres, and T. Karantassis, J. phys. radium 17, 134 (1956).
8. B. N. Figgia and R. L. Martin, J . C h a . Soc. p. 3837 (1956).
9. A. Gilmour and R. C. Pink, J . Chem. SOC.p. 2198 (1953).
10. H. Abe, Phys. Rev. Sa, 1572 (1953).
11. R.L. Martin and H. Waterman, J. Chem. Soc. p. 2545 (1957).
13. R. G. Kepler, P. E. Bierstedt, and R. E. Merrifield, Phys. Rev. Letters 6, 503 (1960).
14. D. B. Chesnut, H. Foster, and W. D. Phillips, J. Chem. Phys. 34, 684 (1961).
15. J. H. E. Griffiths, J. Owen, J. G. Park, and M. F. Partridge, Proc. Roy. SOC.A260,
84 (1959).
16. A. H. Cooke, R. Lazenby, F. R. McKim, J. Owen, and W. P. Wolf, Proc. Roy. Soc.
A260, 97 (1959).
17. A. Danielian and K. W. H. Stevens, Proc. Phys. SOC.(London)A77, 124 (1961).
18. J. S. Smart, Phys. and Chem. Solids 11, 97 (1959).
19. G. H. Jonker, Physicu 22, 707 (1956).
20. W. C. Koehler, E. 0. Wollan, and M. K. Wilkinson, Phys. Rev. 118, 58 (1960).
21. E. 0. Wollan and W. C. Koehler, Phys. Rev. 100, 545 (1955).
22. K. Hirakawa, K. Hirakawa, and T. Hashimoto, J. Phys. SOC.Japan 16,2063 (1960).
23. R. L. Martin, R. S. Nyholm, and N. C. Stephenson, Chem. & Ind. (London) p. 83
(1956).
24. N. Elliott, private communication.
25. E. 0. Wollan, H. R. Child, W. C. Koehler and M. K. Wilkinson, Phys. Rev. 112,
1132 (1958).
26. W. Hansen and M. Griffel, J. Chem. Phys. 30, 913 (1959).
27. M. K. Wilkimon, E. 0. Wollan, H. R. Child, and J. W. Cable, Phys. Rev. 121, 74
(1961).
28. J. W. Cable, M. K. Wilkinson, and E. 0. Wollan, Phys. Rev. 118,950 (1960).
29. S. Foner, J. phys. radium 20, 336 (1959).
30. J. A. Hofmann, A. Paskin, K. J. Tauer, and R. J. Weiss, Phys. and Chem. Solids 1,
45 (1956).
31. F. M. Johnson and A. H. Nethercot, Jr., Phys. Rev. 114, 705 (1958).
32. J. Owen, J. Appl. Phys. 32, 2135 (1961).
33. H. Bizette and B. Tsai, Compt. rend. ucud. sn'. (Paris) 212, 119 (1941).
34. J. W. Stout and L. M. Matarreae, Revs. Modern Phys. 26, 339 (1953).
35. R. G. Shulman, Phys. Reu. 121, 125 (1961).
36. M. Foex, Compt. rend ucud. sn'. (Paris) 227, 193 (1948).
37. R. W. Miller, J. Am. C h a . SOC.60, 1875 (1928).
38. H. Bmette, C. F. Squire, and B. Tsai,Compt. rend. mud. en'. (Paris) 207,449 (1939).
39. B. A. Colea, J. W. Orton, and J. Owen, Phys. Rev. Letters 4, 116 (1960).
40. H. Bizette and B. Tsai,Compt. rend. mud. sci. (Paris) 217, 390 (1943).
41. J. R. Singer, Phys. Rev. 104, 929 (1956).
42. J. Kanamori, Prop. Themet. Phys. (Kyolo) 17, 177, 197 (1957).
THEORY OF EXCHANGE I N INSULATORS 153

FOB TABLESI-VI
REFERENCES
43. J. R. Tomlinson, L. Domssh, R. G. Hay, and C. W. Montgomery, J . Am. Chem. SOC.
77, 909 (1955).
44. N. Perakis, J. Wucher, A. Serres, and G. Parravano, Colloq. Natl. de Magnetisme
Commemoratif de l’oeuvre de Pierre Weiss, Paris, 1967, p. 159 (1958).
45. R. Lindsay and J. J. Banewicz, Phys. Rev. 110, 634 (1958).
46. L. M. Corliss, N. Elliott, and J. M. Hasting, Phys. Rev. 104, 924 (1956).
47. L.M. Corliss, N. Elliott, and J. M. Hastings, private communication.
48. J. M. Bastings, N. Elliott, and L. M. Corliss, Phys. Rev. 116, 13 (1959).
49. G. T. Rado and V. J. Folen, J . Appl. Phys. 31, 62 (1960).
50. L. NBel, Ann. Phys. 3, 137 (1948).
51. B. N.Brockhouse, Phys. Rev. 106, 859 (1957).
52. T. Risk, K. Blinowski, and J. Janik, Phys. and Chem. Solids 9, 153 (1959).
53. P. Cossee and A. E. Van Arkel, Phys. and Chem. Solids 16, 1 (1960).
54. R. AlBonard, J. C. Barbier, and R. Pauthenet, Compt. rend. &. sn’. 242, 2531
(1956).
55. R. AlBonard and J. C. Barbier, J . phys. radium 20, 378 (1959).
56. J. E. Kunzler, L. R. Walker, and J. K. Galt, Phys. Rev. 119, 1609 (1960).
57. S. S. Shinozaki, Phys. Rev. 122, 388 (1961).
58. L.Rimai, H. Statz, M. J. Weber, G. A. deMars, and G. F. Koster, Phys. Rev. Letters
4, 125 (1960).
59. A. Yoshimon, J . Phys. SOC.Japan 14, 807 (1959).
60. C. Guillaud, A. Michel, J. Benard, and P. Fallot, Compt. rend. acad. sci. (Paris)
219,58 (1944).
61. W.Marshall, Phys. and Chem. Solids 7 , 159 (1958).
62. A. C.Gossard, V. Jaccarino, and J. P. Remeika, Phys. Rev. Letters 7,122 (1961).
63. A. C. Gosaard and V. Jaccarino, private communication.

VI. Theory of Superexchange

3. FORMAL
DESCRIPTION
OF MAGNETIC : THEMAGNETIC
INSULATORS STATE

The “textbook” approach to exchange on which such formulae as (11.1)


are based starts by calculating, in the Heitler-London picture in which the
electrons stay localized in their respective orbitals 4 i ( r ) , the energies of
the spin states of two neighboring atoms. The result, which we have already
discussed, is to demonstrate a spin-dependent Coulomb interaction related
to the exclusion principle.
On the other hand, a whole solid built of such atoms is not necessarily
a magnet, para, ferro- or antiferro-, but may be a dismagnetic molecular
crystal, a semiconductor, or a metal. This points up the fact that the
first task of a theory of magnetism is to understand the magnetic stute
itself: what is it that makes a material a magnetic insulator rather than
a metal, a semiconductor, or a superconductor? We could not expect
to work with semiconductors without understanding the concept of the
154 PHILIP W. ANDERSON

forbidden band gap, or with metals without the Fermi surface; what corre-
sponds to this type of concept in magnets?
The existence of such a problem is rather disguised from view by the
fact that many atoms in isolation, or as solutes in very dilute solution, are
magnetic: the magnetic state is clearly a characteristic of the low density
state of electrons on atoms far apart in some sense. Thus the usual approach
to magnetic interactions, of bringing magnetic atoms together and as-
suming that their electronic wave functions are unchanged as they ap-
proach, is not completely unfounded. Where such an approach would be
meaningless in dealing with semiconductors or metals (one atom cannot
be a metal), we shall soon learn that, although it must be handled with
caution, it is here the correct type of starting approximation. Nevertheless
it is valuable, not only as a help to our understanding but as a prerequisite
to knowing how to calculate exchange accurately, to understand the es-
sential character of this magnetic state and how it is related to other
electronic states.
A second approach has been attempted13SJ36in the case of the insulating
antiferromagnets. This is much more closely related to the band theory
which has been fruitful in all other branches of solid-state theory. As we
will discuss shortly, these substances do not fit into the band theory
because, being magnetic, they clearly do not have filled bands; on the
other hand, they are undoubtedly insulators. One may suppose then that
in the antiferromagnetic state there is a filled band with upward spin on
one set of atoms, a filled band with downward spin on the other. This is not
plausible for that great majority of magnetic substances which are highly
diluted, (e.g., CuC12.2H20) because such substances remain insulators
even above the NBel point, where such an alternating band theory is
meaningless. The model might be maintained for only a few highly con-
centrated salts such as V203,which do show changes in conductivity a t what
may be the NBel point. Nonetheless, in the great body of antiferromagnets
which are more or less ionic salts, there is little evidence for any fundamental
differences from the really dilute cases.
In fact, just as the best physical way to define the other states of solids
is in terms of their low-lying electronic excitations-for example, those of a
metal have a momentum near the “Fermi surface’’ in momentum space,
a t which surface their energies go to zero, and carry current; those of
a semiconductor are near one or a few points in momentum space, have a
finite gap and carry current, etc.-one can define the magnetic insulating
state as that state for which the low-lying excitations are spin reorienta-
1% J. C. Slater, Phys. Rev. 82, 538 (1951).
136 T. Matsubara and T . Yokota, Proc. Intern. C m f . Theoret. Phys., Kyoto, 1963 p. 693
(1954).
THEORY OF EXCHANGE I N INSULATORS 155

tions, usually reasonably well described by a Heisenberg Hamiltonian.


The current-carrying excitations resemble those of a semiconductor, and
are usually higher in energy. This definition in terms of elementary excita-
tions picks out what we believe are the fundamental properties of the
magnetic state.
The reason why such a state might come about was clearly set out by
N. F. Mott in 1949137 in physical terms. Following Mott, let us discuss some
simple dilute magnet, say CuCl2-2Hz0. On each Cu ion there is, one
supposes, one hole in the d shell which has a spin 3 and approximately
1 p moment;
~ i.e., on each Cu atom there seem to be precisely nine d elec-
trons, not eight, and not ten.
If the d holes were holes in a d .band rather than in an atomic, localized
d shell, we would, on the other hand, expect metallic conductivity-which
is not observed in such crystals even at elevated temperatures-and
magnetic behavior more like a Pauli paramagnet. The electronic structure
of a partially filled band does involve ionized states like d8 and d10 as well
as 8,as is easily verified. It is also noteworthy that the symmetry properties,
such as the crystal field effects and the anisotropy of paramagnetic reso-
nance, and even such facts about structure as that Cu++ magnetic ions
always prefer noncubic point symmetry according to the Jahn-Teller
effect, are readily explained on the localized d electron model. A band
model deals far less directly with the local point-group symmetry.
It is thus unquestionable that the localized model is experimentally
correct. Why then is it valid? We examine the first step in creating the
band model from the localized one: namely, we take one electron from a
Cu++ ion making d8, and put this electron far away on another ion, making
d’o. This electron now is repelled by nine other d electrons on the same ion,
rather than by the eight it saw previously; thus in this approximation the
system as a whole has increased its Coulomb energy by an amount U. U
may be estimated from the ionization potentials of the free ions, and is of
the order 5-15 e ~ . ~ ~ J ~ ~We J ~ show
8 J 3 the
8 ~ original state in Fig. 4a, the new
one in Fig. 4b.
On the other hand, we can refine the strictly localized approximation
somewhat, allowing the electron to regain some energy because it now can
wander throughout the crystal. Its localized state has the energy of the
average band state, whereas one may place the free electron in the very
137 N. F. Mott, Proc. Phys. SOC.(London) 862, 416 (1949).
1% J. H.Van Vleck, Revs. Modern Phys. 26, 220 (1953).
1880 One of the factors contributing to the tendency of d and f electrons to be magnetic
is the fact that this repulsive energy is greater for d and f than for s or p states be-
cause of their “inner shell” nature: the d electron wave functions are concentrated
near to each other.
156 PHILIP W. ANDERSON

4
-
a

-
a

-
a -
a
-
a
-
a
-
a

(a) MAGNETIC STATE i .(b) ONE FREE PAIR

FIG.4. Localized electrons on a lattice. (a) Pure magnetic, localized state; @) state
with a mobile electron and hole pair.

bottom state of the band, gaining +ae:half the bandwidth for the electron.
Similarly the extra hole may be allowed to gain &bh by being placed in a
freely wandering band state. At high electronic densities and for s and p
electrons, this kinetic energy gain may outweigh the Coulomb correlation
energy loss; in that case, the resulting state is metallic, since now a very
large number of electrons may detach themselves from their ions and run
freely. Actually, in the metallic state the energy loss U may be decreased
by screening, so that one expects that the energy as a function of number
of ionized electrons is as in Fig. 5 (taken from Matt'").
In the dilute case, on the other hand, U outweighs t b . For example,
we shall see that in the oxides 8 is probably -1-2 ev, and much smaller
in more dilute systems. Then it is energetically favorable for the electrons
to stay localized, and all the properties of the usual magnetic insulator
follow. It is interesting to notice that this is an effect of correlation energy
overcoming the kinetic energy which plays the important role in the usual
band theory, i.e., in the theory of the “high density” electron gas. U is a
Coulomb integral between two d-functions on the same atom; the exchange
integral between the two orbitals would be an order of magnitude smaller,
as we have discussed before. This is the simplest illustration of an im-
portant argument which may be carried over to the metallic magnet case.
Whenever exchange plays a large role, Coulomb correlation of the foregoing
type must play an even larger one, since it is a larger effect fundamentally.
Where exchange dominates kinetic energy, correlation must be by far the
largest element in the problem.
THEORY OF EXCHANGE I N INSULATORS 157

t
E
0

I I I I I I I I I
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
P/ N
FIG.5. Electronic energy of a dilute solid aa a function of the relative concentration
r / N of free electrons (after Mott1*7). (a) Magnetic case; (b) metallic case.

In the next sections of this part of the article we will give a qualitative
discussion of the physical mechanisms underlying exchange effects when
looked at from the foregoing point of view, following Ander~on.~O We will
compare this discussion with other treatments of the problem, and mention
as briefly as possible the formal expression of these physical results.
It will become clear that the most essential information required for a
theory of exchange effects is a knowledge of the actual wave functions of
magnetic ions in crystals, which is the central problem of ligand field
theory, so that the following section will review the relevant part of that
subject. Finally, in the last sections of this part the attempts which have
been made at calculations of exchange will be reviewed. In addition we will
describe the semiempirical qualitative regularities which the theories do
explain satisfactorily.

4. ORIGINOF SUPEREXCHANGE
: KINETIC
EXCHANGE
Originally, superexchange acquired its name because of the relatively
large distances, occupied by normally diamagnetic ions, radicals or mole-
cules, over which the exchange effect often was found to act. In the best-
known example, the monoxides of the iron group ions Mn+ +, Ni+ +, etc.,
the ions which interact are separated by more than 4 A, and the overlap
of their actual atomic d shell wave functions is negligible.a7From the first
it was clear that the wave function at these “ligands” was modified by the
presence of the magnetic ions and that this modification of the wave func-
tions had a magnetic character which then could give exchange interactions
with other ions.12,ss
158 PHILIP W. ANDERSON

Andersonasconsidered the simplest such modification : the transfer of


one of the electrons of the ligand ion into the d shell of the magnetic ion.
Because this transfer could only take place into an empty d orbital, the
electron left behind has its spin parallel to the full d orbits and by assump-
tion has an antiferromagnetic interaction with the d-electrons on the op-
posite side, giving antiferromagnetism. Figure 6 presents a diagram illu-
strating this process. For a simple, clear exposition of this “transfer” point
of view, see PI.larshall*and Van Vleck.66” The discovery and measurement of
the hyperfine interaction of the ligand nuclear spin with the magnetic ion
electron spins demonstrated graphically that the ligand wave functions are
partially magnetic, and to the expected
On the other hand, with the passage of time and with further refine-
ment, this type of theory became involved in increasing uncertainties and
complexities. We shall list a few of these here. ,

1. In this theory the exchange effect appears in a rather high order of


perturbation theory, namely, to the third order in the case of Anderson’s

T THE IONIC CONFIGURATION

2 ELECTRONS ARE I N
P STATE ON o--

SPIN s 0 S

II THE “E x c I T ED” CONF I GURAT Io N


Mn’lON 0-ION Mnt+loN

1 ELECTRON IN
P STATE ON 0-

SPIN stf i z1 S

vToT(STOT) = a ~ l o N l c ( S ~ ~ ~‘EXCITED
)+b (STOT)

FIQ.6. “Ground” and “excited” configurations in the original superexchange process.


THEORY OF EXCHANGE I N INSULATORS 159

theory. Its magnitude leads one to the conclusions that the perturbation
theory is not rapidly convergent, and that the early terms which do not
lead to magnetic effects are rather large.
2. Related to the foregoing, several other apparently distinct perturba-
tion effects seem to require inclusion, e.g., double as well as single
transfer integrals, and so on. Yamashita and K ~ n d o and ' ~ ~ Keffer and
Oguchi140list a number of such effects. Nonetheless, all of these other effects
lead to the same conclusions concerning sign and order of magnitude.
Moreover, slight modifications of the wave functions often carry one of
these effects into another. Both facts suggest that the physical effect must
really be much simpler than these theories would indicate.
3. The most serious question which arises is that concerning the
orthogonality dilemma and the choice of wave functions. The obvious
thing to do if one really has to calculate is to start from free-ion wave
functions and put the solid together from them. At normal densities, the
wave functions then overlap appreciably; this is in fact necessary if one is to
obtain an antiferromagnetic sign for exchange, as we have explained earlier.
Overlap, on the other hand, not only complicates calculations enormously
but indicates that in some sense we are not working with the right set of
functions. For instance, correct solutions of Hartree-Fock equations never
overlap. Nonetheless, as SlateF has pointed out (he was anticipated by
Wigner (private communication to G. H. Wander)), overlap appears
essential for arriving at correct conclusions with the Heitler-London
method, if one uses localized orbitals. While Slater's a r g ~ m e n t ' that,
~
overlap cannot be treated in a convergent way in a solid is not valid15
this does represent a real problem.
A second difficulty, besides overlap, with free-ion functions is that there
is every reason to believe that they become considerably modified in the
solid.141 Such modifications, if inserted in the theory, introduce rather
large arbitrary parameters such as the modified overlap integrals.
It has turned out that all these objections can be met by looking at the
problem from a rather different point of view.70 In this new point of view
one tries to make an explicit separation between two very different aspects
of the problem: (a) the first is what one might call the ligand field theory
aspect. That is, one solves the problem, simple in principle but of course
immensely difficult in practice, of obtaining a wave function of a magnetic
ion which is surrounded by the various diamagnetic groups, and in fact
by the whole of the lattice, excluding the exchange efects of the other magnetic
J. Yamashita and J. Kondo, Phys. Rev. 109, 730 (1958).
140 F. Keffer and T. Oguchi, Phys. Reu. 116, 1428 (1959).
141 W. Marshall and R. Stuart, Phys. Rev. 123, 2048 (1961).
160 PHILIP W. ANDERSON

ions. That these exchange effects do not disturb the ligand field wave
function, and thus that such a separation is possible, is shown experi-
mentally by the agreement of hyperfine interactions with ligand-ion nuclei
for dilute and concentrated versions of the same as well as for para-
magnetic and antiferromagnetic temperature regions.4446 A second piece
of evidence is provided by the agreement of crystal field parameters and
exchange integrals in concentrated and dilute s y ~ t e m s . ~ ~Thus,
J * experi-
mentally at least, there is a well-defined wave function belonging to each
magnetic ion which is not changed substantially by its magnetic sur-
roundings in those cases which have been studied. Questions relating to
the energy levels and the wave functions on the individual magnetic ion are
the general province of the ligand field theory. (b) The second aspect of
the problem is, in principle, entirely distinct. It centers about the question
of the way in which two magnetic ions defined in the above way interact
when they come close together. I n principle, this is a more complicated
problem, involving all the headaches of the orthogonality question,
etc. ; mathematically and computationally, however, (b) becomes much
simpler once we have divested ourselves of the diamagnetic lattice by
assuming that aspect (a) is solved to start with. This allows us to see our
way through the difficulties of principle without any great effort.
The foregoing separation is analogous to that which occurs in effective
mass theory, and which has led to such great progress in understanding con-
ducting materials. I n that theory one lumps the whole problem of the
interaction of the electron with the rest of the lattice into certain param-
eters which can be obtained either experimentally or, with much less ac-
curacy and at the expense of long computations, from fundamental theory.
Then the problem of the interaction of the electron with an external field
or an extra charged center becomes no more difficult than the corresponding
problem for free electrons, although it may contain unknown parameters.
In our theory as in the effective mass theory, there are two levels of so-
phistication. The fist level is that of treating the extra electrons in terms
simple one-electron Hartree-Fock functions, while the second centers
about treating them as excitations of a many-body system. In principle a t
least, the use of experimental parameters allows us to work on the latter
l ~ ~ if we use experimental parameters, of course, the theory is
l e ~ e 1 . Even
useful and nontrivial in the sense that it leads to relationships among the
parameters which are not obvious.
142 A. M. Clogston, J. P. Gordon, V. Jaccarino, M. Peter, and L. R. Walker, Phys. Rev.
117, 1222 (1960).
143 J. W. Stout, J . C h a . Phys. 31, 709 (1959).
144 W. Kohn, Phys. Rev. 106, 509 (1957).
THEORY OF EXCHANGE IN INSULATORS 161

This separation also has two incidental advantages which nonetheless


are very real and are worth mentioning. The first is that the basic perturba-
tion theory entering in part (b) involves the interactions between the
actual magnetic ions and their electrons, the intermediate ions being
involved only in the solution of (a) ; this is a relatively weaker perturba-
tion, which means that the convergence is good in the usual cases. The
second advantage also has to do with the elimination of the intervening
ions. There is no distinction in principle between exchange caused by direct
overlap of the wave functions on magnetic ions without intermediate atoms
present, and exchange through nonmagnetic groups, that is, between what
is uqually called “direct” exchange and “superexchange.”
Starting with part (a) of this program, our problem in simplest terms
is that of the state of an electron on a one-electron magnetic ion in a
crystal. There are a number of possible ways of defining this state phys-
ically. First, for example, we could replace the magnetic neighbors of this
magnetic ion by nonmagnetic diluents, as is often done experimentally in
resonance studies. This has the disadvantage that, while the state on the
ligands seems to be independent of the concentration experimentally,
the value of the wave function at the magnetic neighbors-which must be
vital in studying exchange-clearly is dependent on what the neighbors
are. A second possibility would be to surround the spin with antiparallel
neighbors. It is then localized by the repulsion U,and in fact the Hartree-
Fock equations for such a situation must lead to a localized solution.
This is probably a perfectly satisfactory method, but it has not been
worked out, and does not seem to lend itself well to perturbation tech-
niques. A third alternative is to surround the added electron with magnetic
ions of the same variety but to leave their &orbitals empty, so that the
magnetic electron can run freely in the lattice in a Bloch wave. The Bloch
wave is the natural solution to a one-electron problem, and therefore lends
itself to a theory in which interaction effects are to be treated later.
Still another possibility is to calculate from the purely ferromagnetic
case, since the exchange terms then cancel the self-repulsion of magnetic
electrons on the same ion and again the Hartree-Fock solutions are running
Bloch waves. One could do the exchange problem, together with that of
conduction vs. magnetism, by reversing a single spin and studying the
resulting energy states. If we knew ahead of time that the correct Hamil-
tonian was the Heisenberg one, such a procedure would probably be the
most rigorous one. The perturbation procedure we use here is, however,
more transparent and is relatively rigorous. Whenever it fails, Heisenberg’s
Hamiltonian is dubious, among other things. It is well to keep in mind,
however, that a good working definition of the Bloch wave states of the
162 PHILIP W. ANDERSON

magnetic electrons would be the Hartree-Fock solutions for magnetic elec-


trons in the ferromagnetic case.
The use of the Bloch waves as our fundamental one-electron states
emphasizes nicely one of the important simplifying features of this ap-
proach. All of the nonmagnetic electrons in the system then can be treated
more or less as in the free-electron theory of metals and semiconductors,
namely, strictly as “core” electrons, whose wave functions may be taken
as given quantities. They are not expected to change very much as excita-
tions occur in the magnetic electron system. Such core electrons, including
those in 0-- or other ligand p functions, play only two important, if
passive, roles. They contribute to the .self-consistent field; and, most
important, the magnetic electrons’ wave functions must be made ortho-
gonal to all the cores. We will argue later that the true spin polarization of
cores, including the 0-- p electron band, represents a minor effect, in
contrast to the situation for the free (or “s”) electrons in the Zener-Ruder-
man-Kittel type of exchange via free electrons in metals.
We do not wish to imply that the wave functions of the magnetic
electrons are just tight-binding Bloch waves made up of atomic orbitals
orthogonalized to the p functions. We can expect that the “core” wave
functions, and particularly those of the more covalent type of ligand, such
as 0- -, must, in the best Hartree-Fock solution, make partially covalent
mixtures with the d functions. Thus if the magnetic electrons nominally in
cation d functions are to remain orthogonal to all core functions, they
must form antibonding admixtures with 0- - f u n c t i o n ~ . ~ ~ ~ J ~ ~
We will see that it is this admixture of p function into the magnetic
wave function which is the primary mechanism of “superexchange,” in the
sense of exchange via nominally nonmagnetic ions. After including this
admixture we can for practical purposes drop the nonmagnetic electrons
out of the problem. The reason superexchange can be felt over such large
distances, then, is the not very mysterious fact that the actual magnetic
wave functions themselves spread over rather long distances because
of covalent binding effects with the “core” electrons. The most important
physical question is how much the magnetic functions do mix with ligand
functions, and a t the time this is written the problem really remains
very much open. There is extensive resonance data in fluorides whose
complications we will discuss in a later section; but the more interesting
oxides have not been studied very much. We-will discuss this question in
the section on ligand field theory. R
As far *as formal mathematical machinery is concerned, neither parts
(a) nor (b) of our program require anything more complicated than the
IB J. H. Van Vleck, J . Chem. Phys. 3, 843 (1935).
146 J. Owen, Proc. Roy. Soc. A227, 183 (1955).
THEORY OF EXCHANQE I N INSULATORS 163

Hartree-Fock equations. While we may keep in mind that considerably


more generality and rigor might be attained by emulating effective mass
theory and assuming that part (a) is solved in a full many-body sense,
this introduces unnecessary complications. The reader is referred to Ander-
son’s paper70for the first attempts in that direction.
To proceed, the ligand field wave function +i(r) is assumed to be a
solution of the Hartree-Fock equations:

Here the integrals are also spin-sums (so that the exchange term vanishes
when j and i have opposite spins) and the j-sums may extend over all
occupied functions including i if we like.
In order to confine ourselves strictly to problem (a) we must specify
that in some sense we can neglect the relative arrangement of the magnetic
electrons. As we remarked, the most formally exact method of doing this
is to set all the spins parallel. Nonetheless, it seems clear that for this
purpose we can adopt with adequate accuracy the Wigner-Seitz scheme
of removing the charge of the magnetic electron from the cell in which it
finds itself, remembering that this is merely a scheme to find usable wave
functions. Then the equation for the ligand field wave function is

Only one other special provision need be mentioned: if there is more than
one “d electron” per atom, the others are to be included in the Hartree-
Fock problem. Since Eq. (4.2) has the periodicity of the lattice, its solu-
tions are Bloch waves +k with wave vectors k.
Aside from those interactions which have been omitted, the most im-
portant properties of these functions can be deduced from the quantity
e(k). This contains both the effects of crystal fields and the kinetic energy.
164 PHILIP W. ANDERSON

This can be seen with full rigor and simplicity only in the unrealistic case
that the various d-bands are nondegenerate and well separated from each
other. Then to each d-band ((m” there corresponds a particular energy
spectrum e,(k), Since e,(k) is periodic it may be expanded in a Fourier
series:
em@) = a, +c
b , ( ~ )exp ( - z k . ~ ) ,
1
(4.3)

where T are the fundamental (nonzero) translations of the lattice.


Now we construct Wannier functions from the Bloch function solutions
4 k of Eq. (4.2)

(4.4)

The effect of the self-consistent field on the localized functions 4, is given


by applying (4.3):
Xec4,(r - n) = urn+& - n) + c bm(v)4m[r - (n + TI].
T
(4.5)

Here the diagonal term is in the nature of a crystal field separating the
different d-levels m, and the b,( 1) are “transfer integrals” representing
the effect of the kinetic energy in causing the electrons to move through
the lattice.
A more rigorous discussion using the modern theory of oneelectron
functions can be carried out.147It shows that Eq. (4.5)actually must be
generalized to
Xso4,(r - n) = ud,(r - n) + c b,~()$,~[r
mt.7
- (n + TI]. (4.6)

The discussion of this point, and the formula, given in Anderson70 are over-
simplified. Equation (4.6) is to be considered the analog of the one-
electron Hamiltonian of the effective mass theory. It summarizes most of
part (a) of our program; that is, it is the result of a complete ligand field
theory treatment.
We now go on to consider the interactions between magnetic electrons,
that is, part (b) of the program. Magnetic electrons on the same ion core
interact very strongly, of course. These interactions have three forms.
1. Two electrons, both on the same ion, repel each other with an
average electrostatic energy which does not depend much on their relative
orbital or spin functions. This is the repulsive energy we have already
called U,and have estimated to be around 10 ev for d electrons (for details
see Anderson70). It is this energy which keeps the electrons localized; if
14’ E. I. Blount, Solid Stale Phys. 13, 305 (1962).
THEORY OF EXCHANGE IN INSULATORS 165

there are exactly m electrons on every core the total repulsive energy will be
m ( m - 1)
N 2 u,
while if m + 1 tire on one core and m - 1 on another it is
N
m(m + 1) + ( m - l ) ( m - 2 )
2 2

N m ( m - 1)
- u + u. (4.7)
2
Thus as a zeroth-order approximation we assume each electron stays on
its own core exactly; the excited states are elevated by an amount U above
this.
2. There is a true exchange energy
e2
Jmml =
1 &*(r - n)+mt*(r - n)
Ir - r’1
+m* (r’ - n)&(r’ - n) dr dr’,

(4.8)
which appears only for parallel spins and is attractive: i.e., because two
electrons in similar wave functions tend to avoid each other through the
“ F e d hole” which occurs when spins are parallel, they find it less re-
pulsive to be close to each other. Thus when possible the spins on individual
ions are parallel (first Hund’s rule).
3. There are also small differences in the repulsive energies which
depend on the relative orientation of the orbits; e.g.,

These diferences are “Slater integrals” which determine how the orbital
moments orient themselves in the free ion. In the real crystal, there will be
competition between the urn’s(the crystal field parameters) and integrals
like (4.9), which will determine the actual ionic state according to the
usual concepts of ligand field theory.
All of these interactions have the effect of coupling the spins and orbits
of the isolated ions together into a single ionic state. We shall show very
schematically in Section 6 how these states are determined for the transi-
tion metals.
166 PHILIP W. ANDERSON

The exchange effects we are really interested in are those which lead
to mutual coupling of the spins on different ions. The two major effects are
contained completely within the unrestricted Hartree-Fock theory. They
have been called70 Superexchange and Direct Exchange. This nomen-
clature is satisfactory within the confines of the present form of the theory
because the idea of superexchange as an interaction which depends im-
portantly on the presence of the intervening ions does not play any great
role in it, since the background diamagnetic lattice has been eliminated.
On the other hand, it may well be that the distinction between super-
exchange and direct exchange may have value in other contexts, if only
historical ones. Therefore it is necessary to invent a new terminology
to describe the two concepts. The terms we shall use are kinetic exchange
and potential exchange, chosen in order to emphasize that antiferromag-
netism is the result of a gain in kinetic energy, the ferromagnetic effect
of a gain in potential energy.
Kinetic exchange occurs because the best Hartree-Fock orbitals are not
the + m f ~ of Eq. (4.4) when some of the spins are antiparallel to others.
If two neighboring spins are parallel, their orbitals must of course be
orthogonal; but when they are antiparallel, the spin functions are auto-
matically orthogonal so the orbitals may overlap each other. The energy
which may be gained is easily calculated by perturbation theory. For two
t,wo antiparallel neighbors at a distance 7 using (4.7)and (4.6).

(4.10)

and, using just, a single pair of


2b2( T )
Ah’ (parallel-antiparallel) = - -. (4.11)
U
1470 It should be noted that essentially the expression (4.11) was arrived at by Kondo14*
in calculating the difference between the totally ferromagnetic and totally antiferro-
magnetic states. Kondo also noted the physical origin of the main antiferromagnetic
term, although his discussion was not aa complete aa Anderson’s later one. There is
also some resemblance to P. 0. Lowdin’s “alternant orbitals” scheme.149 Lowdin
did not, however, appreciate the relevance of his scheme to the superexchange problem
in general.
l4S J. Kondo, Progr.lTheore.4.Phye. (Kyoto) 18, 541 (1957).
140 P. 0.L6wdb.1, Phyu. Rev. 97, 1509 (1955).
THEORY OF EXCHANGE IN INSULATORS 167

Thus the effect is always antiferromagnetic. It depends on the presence of


half-filled orbitals with nonvanishing b’s associated with the two ions;
since b is of the order 0.5-2 ev, it runs of the order 100-1000°K in energy
in fluorides and oxides.
The demonstration of the equivalence to an s 1 - s ~coupling scheme is
simply carried out. In terms of second-quantized operators, the one-electron
energy (4.6) is
Xac = C amC*,,TCt,u + C bmmr(T)Cn*mCF!r.u, (4.12)
m,n.u n,m,ml,r
0

whereas the second-order perturbation is

On using the relationships between C’s and spin vectors:


c:c+ - crc- = 2s,
c:c- = + is,, sz

this may quickly be shown to be equivalent to an s1.s2 coupling term:

There are several interesting side remarks to make here before going
on to the other possible types of interaction.
1. The constant term of (4.14) has a negative sign; this corresponds
to the fact that the energies of parallel spins are not affected, but those of
antiparallel spins are lowered. This is opposite to the true exchange effect,
where antiparallel spins are not affected. This energy difference may well
be large enough to take into account in analyzing optical measurements on
pairs of solute magnetic ions in dilute paramagnetic solid solutions.150
It has been suggested151that certain transitions in V and Ti oxides
actually are thermal transitions between a low temperature, insulating,
antiferromagnetic state of the type we have discussed here, and a high
temperature, metallic form of d-band. Present experiment^'^*-'^^ indicate
that the low-temperature states may not be antiferromagnetic. Such an
antiferromagnet, however, might be expected to have very low sublattice
so A. L. Schawlow, D. L. Wood, and A. M. Clogston, Phys. Rev. Lelters 3, 271 (1959).
lS1 F. J. Morin, Phys. Rev. Letters 3, 34 (1959).
169 S. J. Pickart and G. Shirane, cited by Goodenough166 (1961).
16) S. J. Pickart and A. Paoletti, cited by Goodenought“ (1961).

164 P. H. Carr and S. Foner, J . Appl. Phys. 31, Suppl., 3445 (1960).

166 J. B. Goodenough, Phys. Rev. 117, 1442 (1960).


168 PHILIP W. ANDERSON

magnetization, which might be missed in a neutron experiment. Thus


the experimental situation is not yet conclusive. If such metal-antiferro-
magnet transitions could be verified they would be of the utmost im-
portance in clarifying our understanding of the theory of the dilute electron
gas. (Note added in proof: S. C. Abrahams has informed the author privately
that Ti203 does indeed show a weak antiferromagnetic neutron diffraction
pattern.)
A second possibility for such transitions, rather similar to the sug-
gestion of Goodenough,166is that they are associated with a lattice dis-
tortion large enough to create an energy gap at the Fenni surface. As his
paper emphasizes, there are often considerable atomic motions at the
transition point. This possibility would, however, usually require a sym-
metry change at the transition. Whether such symmetry changes always
occur has not been investigated.
2. The perturbation theory which leads to the preceding result is

since b/U -
fairly rapidly convergent, at least for most insulating antiferromagnets,
1/10. This is not the case for the older superexchange theories,
and is the result of starting from the exact localized functions. The fact
that we have a well-defined perturbation theory allows us to estimate
higher forms of exchange effects, in particular the validity of the Heisen-
berg Hamiltonian, to very good accuracy. The next purely two-atom
exchange term will be proportional to ( S1. S2)2 (the third-order terms
cannot be of scalar symmetry) and will be of order b4/Ua-i.e.,
times the S1.S2 terms.1s6aSuch high-order effects are now observable in
the microwave spectra of pairs of magnetic solute
3. The lowering of energy is also of great interest in the general theory
of dilute Fenni systems. A very dilute Fenni gas might be expected to

-
form an “electron lattice” as &st discussed by Wigner166J67 and elaborated
by others. At extreme dilutions (T, 2 20) no overlap of wave functions
would occur; but at the dilutions (T, perhaps 10) at which the transition
between metallic and insulating “lattice” behavior takes place, the energy
gain from kinetic exchange may play a large role in stabilizing the lattice
structure.
4. The uses which Moriyagaahas made of what is essentially the
equivalent of Eq. (4.11) in the theory of anisotropic exchange should
be mentioned. In the presence of spin-orbit coupling, the one-electron self-
Actually such t e r n only appear for ions with more than one electron or hole, and
+ +
will have spin coefficientslike: [ B ( n ) . B , ( n ~)l[S,,,~~(n).S,,,,~,(n
T)].In t e r m
of total spin, however, they will look like (S1aSt)’.
166 E. Wigner, Phys. Rev. 46, 1002 (1934).

lK7 E.Wigner, Trans. Faraduy Soc. 54, 678 (1938).


THEORY OF EXCHANGE I N INSULATORS 169

consistent field (4.6), or if one likes the one-electron band energy (4.3),
is no longer spin-independent but may contain spin-orbit terms of the usual
types familiar from modern band theory. Typically, such terms would be
of the order h/A, where h is a spin-orbit energy and A a band separation,
which in this case is of order b itself, -1 ev, and of the same order as
crystal field splittings. Moriya has shown how such terms can give two
effects when inserted in the perturbation formula (4.11). First, when there
is a center of symmetry between the two ions concerned, the well-known
pseudodipolar exchange important in the theory of crystal magnetic
anisotropy is obtained; second, in the absence of a center of symmetry,
one obtains a cross-term between the spin-orbit part of b and the ordinary
transfer giving the Dzialoshinski anisotropic exchange mentioned in
Section 2, which is linear in h/A and thus may be much larger on occasion
than the pseudodipolar term. also pointed out that this pos-
sibility existed.

5. POTENTIAL AND OTHEREXCHANGE


EXCHANGE EFFECTS

a. Potential or True “Direct” Exchange

The next term, which is still entirely within the Hartree-Fock ap-
proximation, is potential exchange. This is the true Heisenberg term, in-
volving only the interelectronic potential of interaction,
-Jmm*(T) (2 + SI*SZ), (5.1)

X 4mf*(r’- n - T)4m(r’ - n) dr dr’. (5.2)


Our definition ensures that (5.2) is always positive, and makes a clear
distinction between the physically quite distinct effects of the kinetic
energy and of the repulsive interelectronic potential energy. The exchange
integral may be written

(5.3)

where
pLmt(r) = -
z)4m(r - n).
&#(r - n (5.4)
This is the Coulomb self-energy of the “overlap charge distribution”
pLm)(r) .
K. W. H. Stevens, Revs. Modern Phys. 25, 166 (1953).
170 PHILIP W. ANDERSON

Suppose that the localized wave function +d has an admixture E of


ligand function:
4 m ‘V +d - E+lig ....+
-
b t2 X 1 ry, i.e., -
One may estimate, as we shall discuss later, that it is reasonable to set
(+lig 1 Xs,I +lig) .e2. Since U is also of the order 1 ry,
the kinetic exchange effect is altogether w e 4 ry.
Now J is also of the order €4 if we assume, as is reasonable, that most of
J also comes from near the ligand. At first, we might assume that in ( 5 . 2 )
one could take only the €+Iier terms for all four +’s, and thereby obtain e4
times a Coulomb repulsion like U. This result is nearly as large as - b2/U.
It is, however, an overestimate. The reason is rather an important one,
and not widely enough appreciated. In fact, the same reasoning may be
applied with wide generality to show why the Hartree method is usually at
least a rough approximation to the Hartree-Fock, as the latter is to the
true representation.’e
The essential point is as follows. Exchange or off-diagonal Coulomb
integrals involving orthogonal orbitals tend to be rather small, often as
much as an order of magnitude smaller than one would expect. We see
this effect in atomic spectroscopy again and again. There the direct integral
U is -15 ev whereas exchange between the same pair of orbitals is -1-2 ev,
as are the Slater F’s and G’s. Again, correlation energy runs -1 ev per
electron which, since the energy denominators in perturbation theory are
-1 ry for the processes involved in correlation, means that the scattering
matrix elements are rather small: namely, -2-3 ev. To repeat: average
fields, such as Coulomb repulsions, or term value differences, always seem
to outweigh comparable exchange or correlation effects.
The reason for this is seen by looking at (5.4). When orbitals +m and
4mt are orthogonal, we have the orthogonality relationship

1 dr pkmI(r) = 0, (5.5)

which means that the overlap charge distribution phmt(r) must be equally
often positive and negative; it must have nodal surfaces, at which it
vanishes, between the positive and negative regions. Usually the nodal
surface is found just where the overlap would otherwise be expected to
be largest; between the atoms, for instance, if 4m and refer to
neighboring atoms. Figure 7 shows schematically how 4 m and +ml look for
a pair of d cations separated by a typical ligand such as F- or 0--. We
see that here p is positive close to the atoms, and of order d, but that
between the atoms the charge density is negative and, of course, also of
order 2. One suspects that, because of the relatively long-range nature
THEORY OF EXCHANGE IN INSULATORS 171

FIG.7. Illustration of MacWeeny’s theorem for the exchange integral of two d func-
tions separated by a ligand: (a) and (b) are the two orthogonalized d functions, and (c)
is the net overlap charge distribution.

of Coulomb forces, the potential of the charge between the atoms succeeds
fairly well in compensating that on the atoms, so that J is perhaps some-
what less than an order of magnitude smaller than e4U. As we shall see,
there is good experimental evidence to show that potential exchange plays
a role which is minor, but by no means negligible, in determining Curie
points in oxides and fluorides.
Potential exchange is the major interaction term between two orbitals
which are orthogonal by symmetry, as, for example, between a d,, orbital
on one cation and a dzy orbital on the other, in the configuration M++-
0---M+ +. Because of this fact, that potential exchange occurs even be-
tween orbitals of differing symmetries (as pointed out by Moriya in a pri-
vate communication), there is a much greater multiplicity of potential
exchange terms than of kinetic exchanges ones in a case like Mn+ +-O- --
Mn++. This may help to overcome the intrinsic smallness of potential
exchange.
172 PHILIP W. ANDERSON

b. Indirect and Polarization Effects

Finally, we come to consideration of the true polarization effects.


Actually these are of several kinds, and are not quite as simple as sug-
gested in Anderson’s paper.1° Only two are important at all; we shall now
discuss these.
A first kind, which we might call true indirect exchange, is, as a matter
of fact, still contained within the unrestricted Hartree-Fock approximation.
As discussed in connection with antiferromagnetism by S l a t e P and in
connection with hyperfine interactions by HeinelS8and Wood and Pratt,lSs
the entire Hartree field for electrons of one spin direction is different from
that for electrons of opposite spin in the presence of magnetic electrons,
because of the exchange terms in the Hartree-Fock equations. In particular,
consider an electron associated with one of the nominally diamagnetic
ligand functions 41 in the presence of a magnetic electron of up spin in the
magnetic function dm. The Hartree-Fock equations for 41 t and 414 are

Here V’ is the potential due to all other electrons than &.


We may find the difference between rpzt and 414 by perturbation theory.
Let

where

The term (ml I lm) is the ordinary exchange integral, which we are
certain is small enough that it will shift the energy level I slightly without
leading to any visible effects on the wave function. The term (ml I mm)
168 V. Heine, Phgs. Rw. 107, 1002 (1957).
J. H. Wood and G. W. Pratt, Phys. Rw. 107,995 (1957).
THEORY OF EXCHANGE IN INSULATORS 173

is unimportant except for helping to maintain the formal orthogonality


of 1 and m.
The polarization effects come from the first-order off-diagonal terms,
which lead to

thus the modification in the density of upward spins is approximately

As noted by Heine and Slater, this expression has a form very similar
to a straightforward polarization of the j wave function. The only difference
is that, whereas in the case of dielectric polarization the matrix element
which causes the polarization is

we find, in (5.10)) the nonlocal operator

(5.12)

replacing the simple operator z. Thus (5.10) will lead to effects similar to
those associated with polarization by an ordinary field. The polarization,
for instance, is always in the direction of the effective field, that is parallel
spin polarization moves in the direction of the wave function. On the
other hand, the effect will be somewhat weakened both by the nonlocal
nature of the operator Fm(r, r’), and by the MacWeeny orthogonality
effect.
The order of magnitude of this effect can be expected to be rather
small-even though it appears as a second-order effect in the old-fashioned
type of perturbation theory,a in contrast to the true superexchange which
is apparently third order. First, we may compare the spin polarization at
the ligand nucleus. In the orthodox effect, this is given by the square of
the percentage of 4mwhich is mixed with 41 (squared because we measure
a density), that is, is proportional to c2. Experimentally, this amounts
to -0.05 for p electrons, -0.02 for s, in fluorides (KNiF6).l@JB1
180 R. G. Shulman and K. Knox, Phys. Rev. Letters 4, 603 (1960).
161 R. G. Shulman and K. Knox, Phys. Rev. 119, 94 (1960).
174 PHILIP W. ANDERSON

The exchange polarization effect which we have just discussed leads


to a polarization of the order
( m i I Im) ‘2 exchange integral
X slight reduction factor, (5.13)
‘j - ‘1 1 rY
or -10-3 for 2 p electrons in fluorides. This polarization may easily be
measured for two reasons: (1) it will be of antiparallel sign, since as we
noted the magnetic orbital attracts parallel spins towards itself, leaving
antiparallel ones behind; ( 2 ) it will occur even when +,, is quite orthogonal
to 41, as, for example, for an F nucleus actahedrally surrounding C1.3+.*@’
The predicted sign and magnitude are confirmed by Shulman’s experiment
on KzNaCrFa, although the magnitude is experimentally a bit small if
anything.
As for the resulting exchange effect, this will clearly be of the order

(5.14)

and may be ferromagnetic or ant.iferromagnetic depending on details of


the band structure, etc. The order of this is then
e4 X J2
N ~ N (potential exchange), (5.15)
rY
or -10°K in the case of the oxides. It seems likely that even where kinetic
exchange vanishes by orthogonality, potential exchange will swamp this
effect in substances involving d bands. It will play an important role for f
electrons, which do not tend to mix with ligand functions.
Each of the previous three effects came directly from a term in the
one-electron energy of an unrestricted Hartree-Fock theory. The first,
kinetic exchange, represents a lowering of this one-electron energy by
virtue of the lack of orthogonality of wave functions with antiparallel
spins. The second, potential exchange, is simply the mean value of thc
Fock exchange term between two magnetic functions. Finally, the third,
indirect exchange, is a term in the exchange polarization energy (within
the unrestricted H-F) of nonmagnetic orbitals which depends on the
magnetic arrangement because it is a cross-term between the polarizing
fields of two magnetic ions.
We must now take into account those exchange and correlation effects
which result from true many-body corrections to Hartree-Fock. I n such
corrections, we must always consider the simultaneous excitation, by their
Coulomb interaction, of two electrons from their one-electron wave func-
tions into excited states. We will give shortly a general bookkeeping method
THEORY OF EXCHANGE IN INSULATORS 175

whereby, by means of diagram techniques, all such complicated effects


can in principle be written down.
In fact, however, est,imates show that most of these effects are ex-
tremely small. Thus far only one such mechanism has been seriously pro-
posed to be of importance in superexchange. PrattlB7Goodenough and
Loeb,162and I C ’ e ~ b e t ’have
~ ~ J calculated
~~ the following possibility: a pair of
electrons in a single ligand orbital 41 interact with one another and are
excited simultaneously, one into the spin-up orbital on ion m and the other
into the spin-down orbital on m‘. The relevant matrix element is

This mechanism is antiferromagnetic in influence for the same reason


as simple k,inetic exchange, which it resembles closely; that is, it does not
occur if m and m‘ have parallel spin. The two electrons in the orbital I are
necessarily antiparallel. The magnitude is

(5.17)

This is quite small, even smaller than (5.15) because the energy de-
nominator is so large, involving a double excitation. The matrix element
(5.16) is of course also subject to the McWeeny reduction factor.
Nesbet ascribes much too large a magnitude to (5.17) in the 3d oxides,
partly because his estimate of the energy denominator is made on a false
basis. He computes the energy denominator directly from Madelung con-
stants and ionization energies and obtains a rather small value, about 5
volts. Keffer and Oguchi140 pointed out that this is far too small. One
way to see this is the following. One might argue that although Nesbet’s
procedure is inaccurate in a crystal as far from purely ionic as the 3d
oxides and that there could be considerable question concerning the correct
values of the various constants, it is still possible that he might obtain
thus a fair estimate of the difference in excitation energy of the first and
second ligand electron. It turns out, however, that the second ligand
electron requires at least 6 more volts of energy than the first; thus if
Nesbet’s total energy is correct, it would imply that zero or less energy
is needed to excite the first electron. This contradiction probably shows
that the pure ionic picture cannot be used to estimate energies in the
oxides. Nesbet’s empirical evidence for his mechanism is also dubious.
lae J. B. Goodenough and A. L. Loeb, Phys. Rev. 98, 391 (1955).
1°3 R. K. Nesbet, Ann. Phys. (N.Y.) 4, 87 (1958).
laaa R. K. Nesbet, Phys. Rev. 119, 658 (1960).
176 PHILIP W. ANDERSON

In all of the perturbation exchange effects, (5.14) and (5.17)) as well as


in the higher order effects which we will discuss shortly, a word of caution
about convergence is very important. A full discussion of this point will be
given by Herring.I6If one keeps a finite number of terms in these formulae,
but goes to the limit of large separation between atoms, that is weak inter-
actions, one will obtain wrong results because the wave functions of excited
states, for example free electron states, can spread over very long ranges.
Thus an error in summing the series can lead to a very bad result at large
distances. On the other hand, there is no reason not to expect convergence
if one takes a finite range or size of intertytion and at the fixed range sums
the perturbation theory as far as the terms remain large. This is the sense
in which all the succeeding terms we consider here may be expected to be
rather negligible.
For the sake of completeness, we will give here an outline, at least,
of the problem of classifying all of the higher-order effects. As with all
such complicated situations, the only conceivable way of keeping one's
bookkeeping straight is to resort to some type of diagram. In these diagrams,
which are not founded on any particularly erudite form of perturbation
theory, but represent a bookkeeping device, we use the following schemes
of notation (see Fig. 8).
I. Magnetic electrons are represented by solid horizontal lines. The ion
involved (usually m or m') will be shown next to the line; it is not neces-
sary to show the spin, as we shall see.
11. Nonmagnetic electrons are represented by dashed lines. We will
not use the Feynman hole convention because it would simply obscure the
meaning unnecessarily. Thus, there is no need for time arrows on our dia-
grams, since all interactions are between electrons in existence. The label
of the orbital ( 1 for ligand, or j for an arbitrary empty orbital) will be
shown next to the line.
111. Coulomb interactions are vertical wiggly lines. (a) shows a scatter-
ing of m' to 1 and j' to j by a Coulomb interaction. To simplify things, we
use a special notation when a Coulomb interaction involves exchange; i.e.,
one of the particles is scattered into the state originally occupied by the
other. This notation is shown in III(b) of Fig. 8; the lines are crossed in
such a way that the continuously occupied states remain at the same
horizontal level.
IV. A special notation is given the Coulomb interaction between two
magnetic electrons on the same ion. Here we show a succession of non-
scattering Coulomb interactions. This calls attention to the fact that when
one electron virtually occupies tthe magnetic state m with upward spin,
THEORY OF EXCHANGE I N INSULATORS 177

SYMBOL CONTRIBUTION

1 ELECTRON IN MAGNETIC 6, IN DENOM.


0R B ITAL

11 ELECTRON IN NON-
€j OR EL
MAGNETIC ORBITAL

m
(a) COULOMB INTERACTION

COULOMB EXCHANGE
INTERACTION

m T W O ELECTRONS IN SAME
ORBITAL u IN DENOM.
m

Y VIRTUAL JUMP BETWEEN


LOCALIZED ORBITALS
4
FIQ.8. Symbols used in diagrammatic treatment of exchange (explanation in text).

say, and m with downward spin is also occupied, one gets a spin coupling
not because of the Coulomb interaction, at least directly, but because
this state of affairs can occur only if the spins are antiparallel.
V. I n order to include the kinetic exchange effect itself, we must intro-
duce it somewhat arbitrarily, since it cannot be generalized to occurrences
in the intermediate states. This is a consequence of the fact that only the
electrons on the magnetic ions do not occupy solutions of the one-electron
Hamiltonian, so that only they have off-diagonal matrix elements b of the
kinetic energy. The notation for the action of such a matrix element is
given in Fig. 8V, and is drawn from the notation for an external field inter-
action in ordinary diagrams.
On the right in Fig. 8 are shown the contributions to the numerical
value from the various interaction parts of the diagrams; note that IV
contributes to the energy of the intermediate state and not to the matrix
178 PHILIP W. ANDERSON

elements. I n calculating the amount of interaction caused by a given


diagram, we of course simply divide the product of all interaction matrix
elements by the product of all intermediate state energy denominators, as
in the usual Brillouin-Wigner perturbation theory.
The rule for finding magnetic interactions, in contrast to the myriads
of irrelevant nonmagnetic correlation effects, is the following. At each
interaction it is essential that spin coupling be maintained, and spin cou-
pling cannot be transmitted through any ordinary Coulomb interaction. In
fact, there are only two ways spin coupling can be transmitted, corre-
sponding to our two basic exchange effects: (1) by actual exchange of the
electrons involved, which means that the electrons, the spins of which are
to be coupled, must actually exchange identity like III(b) ; (2) by the

FIQ.9. Some typical exchange diagrams: (a) type A (“direct” or ffpotentiJ’J);(b)


type B (kinetic); (c) nonmagnetic.
THEORY OF EXCHANGE IN INSULATORS 179

kinetic exchange effect, which occurs only for antiparallel electrons, and
results from virtual hopping into an empty antiparallel orbital. The
topological property, then, which any magnetic interaction diagram shows,
is that a horizontal line across the diagram either must cut an electron
line of one of the magnetic electrons, or a “superexchange interaction” in
which two magnetic electrons occupy the same orbital, one with spin up
and the other with spin down. In a nonmagnetic diagram, one can cut off
one magnetic electron from the other, passing through only non-U-type
Coulomb interactions.
One additional major simplification can be made. This is based on the
observation that all interactions in a diagram must be either pure potential
exchange or pure kinetic exchange type, so that we can classify our diagrams
into type A (potential) and type B (kinetic). The basis for this is the fact
that essentially two electrons occupy the same orbital in a kinetic exchange
spin interaction. Now if another magnetic electron wishes to interact via
exchange with these, it cannot tell which of the two electrons it is to
interact with, since there can be no exchange interaction with a singlet
state. Thus, there is no way to form a chain of mixed true exchange and
kinetic exchange interactions.
In Fig. 9 we illustrate (a) a general interaction of type A, (b) one of
general type B, and (c) a nonmagnetic case. To show how interactions are
computed, we observe that the exchange integral from (a) is

1 1
(ml I jl’) ( jm’ 1 m’l’’) 1 lm)
Ei + El’ - em - El El’ + € 1 -
~ ~E l - em
(1’‘l‘

(5.18)
(this may be either ferro- or antiferromagnetic, depending on the phases
of the matrix elements) whereas (b) is
1
(11 [ j’m’)
ej’ 4-Em’ - 2€1 - Ul + urn,(rnj’ I j’l”’)
1
X
2ep + up - 2El - Ul + U d ( j”m’ I 1‘1)
1
X (j”1’ I 11). (5.19)
€1. -l- e j l , - €1 - u1
In (c) we show the line which cuts the diagram crossing only normal
interaction lines.
180 PHILIP W. ANDERSON

=. ( 2 em - 2 e t + 2 urn-UJ’

FIQ.10. Diagrams for the exchange effects already discussed: (a) kinetic exchange;
(b) potential exchange; (c) exchange polarization; (d) PrateNesbet mechanism.

In Fig. 10, we give the diagrammatic representation of all four effects


discussed previously. The Nesbet effect is unique in that one spin coupling
is hidden in the initial state, namely, the joint occupation of I by the two
nonmagnetic electrons. It is now quite clear that the effects of next lowest
order beyond these will be represented by a number of third-order dia-
THEORY OF EXCHANGE IN INSULATORS 181

grams like Fig. 9a. It does not seem easy to get any diagrams of type B
of third order which are other than trivial self-energy insertions into Fig.
10d. Thus, under our conditions, the effects of higher order can be neglected
satisfactorily. That is, in fact, the only useful result of the foregoing
classification scheme.
With this we close the discussion of the origin of exchange effects.
It should be reiterated that only two effects appear to be of any quanti-
tative significance in practical cases involving d electrons:
1. kinetic exchange, the second-order effect of virtual electron transfer
between magnetic ions, always antiferromagnetic;
2. ferromagnetic “potential” exchange.
These effects may occasionally be supplemented by the exchange polariza-
tion effect.

6. THE ISOLATED
MAGNETIC
ION:LIGANDFIELDSAND INFORMATION
FROM RESONANCE
NUCLEAR

a. Phenomenological Ligand Field Theory

At this point we shall begin to descend from the heights of perturbation


theory, orders of magnitude, and matters of principle, and try to assess
as best we can the real numerical situation. We classified the problem into
two parts: the treating of the isolated magnetic ion and the interactions
between magnetic electrons. This section takes up the former.
There are a number of good reviews of the ligand field problem, par-
ticularly in the areas of chemical effects and optical spectra. See for ex-
ample, Ballhau~en,’~~ Bowers and 0wen,lffiGriffith and Orgel.lee Our in-
terests, however, are much more specialized, and the questions we ask are
much simpler, than those of these reviews. In addition, neither has the
impact of measurements by modern resonance techniques on our knowledge
of magnetic wave functions been reviewed, nor has there been a critical
discussion of the possible relevance of the semiempirical calcuations which
have been made. We will try to emphasize these areas. There exists a
fairly extensive literature on crystal fields in the lanthanide elements, but
practically nothing on exchange. We omit them for this reason.
As we have already pointed out, all the experimental parameters seem
to depend almost entirely on the nature of the central magnetic ion and
1w C. J. Ballhausen, Kgl. Damke Videnskab. Sekkab Mat. fys. Medd. 29(4) (1954).
16 K.D.Bowers and J. Owen, Repts. Progr. in Phys. 18, 304 (1955).
166 J. S. Griffith and L.E. Orgel, Quart Revs. (London) 11, 381 (1957).
182 PHILIP W. ANDERSON

the identity and shape of the first shell of ligands, so that we can confine
ourselves to the quantum mechanical discussion of this finite system
alone. The usual ligand field theory concerns itself primarily with the struc-
ture of the energy levels of this system, and occasionally with optical transi-
tion probabilities. The reason is, of course, that the optical spectra, the
ligand field contributions to chemical energies, and to some extent the
effects of the ligand field in the microwave spectra, depend primarily on
the energy levels and not on the wave functions.
Our main interest, on the other hand, is rather in the exchange integrals
and the quantities bmmt( T),the transfer integrals to neighboring magnetic
ions. Although the latter quantities are in principle directly measurable as
band parameters for mobile d electrons, self-trapping and polaron effects
will probably always conceal the true structure of the bands in practice.
We might expect much greater accuracy from direct computation using
wave functions, if we knew them. In any case, the J’s are not band param-
eters.
At the same time, a considerable body of new data is appearing on
hyperfine interactions with ligands, which gives us enormously detailed
and exact wave function information. This should stimulate theory by
providing both semiempirical data to insert into exchange theories and new
results to be explained from first principles.
We will summarize briefly the point of view of the orthodox “chemical”
crystal field theorists, exemplified by the review article by Ballhausen
and f i s t put forward by Van Vleck13 and Penney and Schlapp.166a Their
first observation is a phenomenological one: that most of the details of
the energy level schemes, in the iron group at least, can be summarized
by choosing a single parameter called “Dq” which would, in a purely
“isolated atom” theory, be related to the strength of the quartic component
24 + +
y“ 2‘ of the cubic electrostatic field at the magnetic ion. It seems
that whether or not the magnetic ion is really an isolated ion affected only
by electrostatic influences from the surroundings-and it is probably not-
enough of the scheme is determined by symmetry alone that this represents
B good starting approximation.
It is a useful exercise in understanding the phenomenological crystal
field scheme to see how the first few d electrons in the iron group work out.
The first and simplest case is d’: T9+ or V4+. Here the isolated ion has
1 = 2, s = 4, which is of course split by the spin-orbit coupling of -200
cm-l into a ’& and ‘D6,2 pair of states.
Now we introduce the cubic crystalline field Dq. The initial fivefold
degenerate d state splits into the two sets of states d, and d,, or more simply
1660 W. G. Penney and R. Schlapp, Phys. Rev. 41, 194 (1932).
THEORY OF EXCHANGE I N INSULATORS 183

d,* and dzy, where d,l has the symmetry of the pair 3z2 - r2, x2 - y2 (re-
presentation r3)and dzy that of the triplet xy, zx,yz (rS).
I n an octahedral environment, the electrons prefer to avoid the nega-
tively charged ligands and occupy the dzy states, which are lowered by
4 Dq, the d,2 states being raised by 6 Dq (center of gravity constant).
Since Dq is -lo3 cm-’ for many ligands, this is a splitting of lo4 cm’ or
-0.8 ev, which is very appreciable. We will see that this splitting of
d,s and dzyof the order of -lo4 cm-1 gives us a guidepost to the building-up
of many-electron cases. In purely cubic symmetry, the introduction of
spin-orbit coupling splits the bottom sextet further into a doublet and a
quartet, the triply degenerate orbital state acting in a sense like a p state
and giving (‘J” = # and 4 states. The latter lies lowest.
I n the solid, the orbital degeneracy should be, and often is, removed
by deformation of the surroundings of the ion. This is the “Jahn-Teller
effect”’67J@ (for effect on magneti~m’~~). The effect finds its origin in the
fact that the orbital states are degenerate and that the lattice can deform
in such a way as to lower the energy of one of the electronic states. The
Jahn-Teller theorem states, in simple language, that there is a parameter x
of lattice displacement for which the energies of the three states behave as
in Fig. 11. Since the energy change is linear in x,the ordinary forces causing
crystal equilibrium cannot compete because they are quadratic in x about
the symmetry point.
The fact is, however, that in many cases the Jahn-Teller matrix ele-
ments seem to be sufficiently small that the distortions do not necessarily
occur. The resulting splittings are small compared to spin-orbit and to
lattice zero-point motion effects. It is common for an octahedral complex
d4 or de (such as Cu++.6H20) to undergo a tetragonal Jahn-Teller dis-
tortion in order to lower d,l relative to dzz-u*, say. However, the correspond-
ing distortion seems to play less of a role in d z U ( r 6configurations.
)
The most prominent feature of the optical spectrum of d’ configurations
is the transition which carries one d electron from the r6to the orbital.
According to the foregoing reasoning, this occurs at 10 Dq, which in TiS+.
6H20 is found by measurement to be about 2 X lo4 cm-’. In general,
the Dq values are rather easily obtained from absorption spectra in a
similar way even in more complicated cases; these absorption bands give
the d complexes their characteristic colors. The resulting values of Dq
are presented in Table VII, for the ions surrounded by water octahedra.170
IE7H. A. Jahn, PTOC.Roy. SOC.A164, 117 (1937).
10 H. A. Jahn and E. Teller, PTOC.Roy. Soc. A161, 220 (1937).
J. H. Van Vleck, J . Chem. Phys. 7 , 72 (1939).
170 0. G. Holmea and D. S. McClure, J . C h . Phys. 26, 1686 (1957).
184 PHILIP W. ANDERSON

t
E

DEFORMATION OF OCTAHEDRON U-V


FIG. 11. Illustration of the Jahn-Teller theorem for a particular case: d,,, d,,., and d,.
in an octahedral environment.

Dq VALUESFOR IRONGROUPIONS
TABLEVII. OCTAHEDRAL

No. of
d electrons Ion 10 Dq (cm-1)

1 TiS+ 20,300
2 Va+ 18,000
3 w+ 17,000
V++ 11,800
4 Cr++ 13,900
Mna+ 21,000
5 Mn++ 7800
Fe3+ 13,800
6 Fe++ 10,Ooo
C03+ -18, OOO
7 co++ 9800
8 Ni++ 8500
9 cu++ 12,500 (uncertain: strong tetragonality)
THEORY OF EXCHANGE IN INSULATORS 185

As mentioned in the various reviews, the splitting varies from ligand to


ligand, but 0-- and HzO have values similar in magnit~de.”~
When we go on to d2, as for example in the case of V3+,we are presented
with a somewhat more complicated question as to the nature of the ground
state. There are two quite distinct approaches which can be thought of as
representing different limits for starting a perturbation theory: the weak
and strong crystal field limits.
If the crystal field is weak, we think first in terms of the isolated ion.
In the isolated ion, Hund’s first rule tells us that the spins of the two d
electrons prefer to be parallel because of the direct exchange integral
between the two orbitals, which are automatically orthogonal because they
belong to the same spherical potential. This integral is of the order -lo4
cm-l. The orbits then place themselves as nearly parallel as possible. Since
+
the exclusipn principle prohibits them from being parallel ( L = 2 2 = 4),
+
they enter ml = 2 and 1 states, giving L = 2 1 = 3. We have a 3F state.
Now in the presumably weak crystal field, the sevenfold F state splits
into 1, 3, and 3-component degenerate states, the triplet lying lowest.
This triplet has the symmetry Fa. The state is 31’4,which is ninefold de-
generate. Spin-orbit splitting and residual noncubic crystal fields will
resolve this degeneracy further in an actual case.
A 3P state lies next above the 3F state in the free ion, if we believe the
general rule that Hund’s first rule (enforced by the direct exchange inte-
gral J) is stronger than the second. The integrals which are relevant for
the second rule depend on the angular dependence of the exchange and
correlation integrals of the two orbitals, and are called the Slater “F”
and “G” integrals. These represent one of the important parameters of the
crystal field problem. Of course, the splitting they cause is seldom observed
directly, but it enters M a parameter in the secular problem for the level
scheme.
In the opposite limit of rather strong crystal fields, we suppose that
the largest parameter is lo&, so that the d,, and d.* orbitals are split by
a quantity rather large compared to the free ion splittings. Then it is clear
that the first step is to put the two electrons in two of the d,, orbitals, the
lowest single-electron orbitals. If the spins are parallel, this can be regarded
a,a simply leaving one hole in the dzu spin-up shell, so that the symmetry
is just that of a d,, hole. That is, we have a threefold orbitally degenerate
triplet, 31‘4, just as in the weak-field case. It turns out that the symmetry of
the lowest state can always be discovered by this trivial type of considera-
tion, in which we fill the r3and r6 shells in the obviously logical order.
17
’1 F. J. Morin, Bell System Tech. J . 37, 1047 (1958).
186 PHILIP W. ANDERSON

In this case, however (d2and d7 are the only examples), the actual wave
functions in the two limits differ. The (d,,)2 triplet level contains about
20% 3P and 80% 3F, and vice versa: 3F r4 contains some dzv - d.2 ad-
mixture. However, the secular problem of determining the correct state,
given Dq, is an easy one.
In Cr3+, the dzy shell is filled, and we obtain the first orbitally nonde-
generate situation. This is a “good” case for magnetic resonance. The
ground state is a quartet which is split into two Kramers’ doublets by
rather high-order effects of spin-orbit splitting, which mix the cubically
symmetric ground state with higher states.
With d4 we first encounter a third altetnative situation, the so-called
very strong field ca~e~7~-insome cases named, a bit misleadingly, the
“covalent” case. Here, we must first ask whether the intra-atomic exchange
integral which enforces parallel spin is strong enough to overcome the energy
10 Dq. The Pauli principle tells us that the fourth eIectron cannot go into the
dzy shell with spin parallel, so it has its choice of the d,2 shell with parallel
spin, the obvious choice for quasi-free paramagnetic ion complexes such as
Cr++ or Mn3+.6H20. These complexes have spin 2 and L = 2, and an
orbital wave function having the symmetry of the d.2 type hole. As men-
tioned earlier, they usually destroy the d.2 - dza-y2 degeneracy by a tetra-
gonal distortion. The other possibility is that in which the extra electron
occupies an antiparallel dzy orbital; this is favorable when 10 Dq > J .
The resulting state is very similar to the (d,,)2 case. This latter case is
the usual one in the “covalent” complexes like manganicyanides
( M ~ I ( C N ) ~ )d6
~ - also
. has the two choices, the familiar Fe3+ ion of the
ferrites, Fez03,etc., with a half-filled d shell and spin 8, or the ferrocyanide
ion, with spin 3. In the cases of the da cobalti-cyanide and ferricyanide ions,
the dzyshell is full and we have diamagnetism.
With this as a broad outline, we will rely on the wealth of review
literature in the field to supply the reader with greater detail. One in-
teresting new development is the improved group theoretical treatment of
crystal field theory developed by Koster and Statz.173J74 They find that
many of the results of crystal field theory are entirely consequences of the
most elementary aspects of the model and symmetry, so that they imply
nothing specific about the wave functions or the degree of covalency in the
complexes. Other aspects are not uniquely determined by symmetry;
for instance, in principle the spin-orbit coupling as well as the Slater inte-
grals could vary between d,, and d,2 orbitals. The fact that they do not vary
much is a fair argument for maintaining the isolated ion picture to some
extent.
172 M. Kotani, J . Phys. Soc. Japan 4, 293 (1949).
G. F. Koster and H. Statz, Phys. Rev. 113, 445 (1959).
G. F. Koster and H. Statz, Phys. Rev. 116, 1568 (1959).
THEORY OF EXCHANGE IN INSULATORS 187

b. Theory of Ligand Field Parameters and Wave Functions: Ionic

The subjects of primary interest to our main development are the origin
of the crystal field and the electronic theory of the magnetic wave functions.
The most important parameter is, of course, Dp, and, to a lesser extent,
the noncubic components of the crystal field, at least when they are as
large as in Cu+ +. In addition, the spin-orbit parameter, the intr%atomic
exchange and the Slater F and G integrals can be of importance, especially
since the deviations from free-ion values can tell us to what extent the wave
function in the solid is modified relative to the ion. Equally important are
the :interaction constants of the nuclear hyperfine structure.
Two qualitatively different approaches have been followed in this
problem. The first is the very straightforward point of view that one
should simply take the unmodified or only slightly modified free-ion func-
tion as given, and calculate directly the effect of the electrostatic potential
of the other ions on it.
The initial efforts in this direction were made by Van Vle~kl6~ and
Polder176 using free-ion d functions and assuming that the effect of the other
ions was just that of negative point charges at the positions of neighbors
such as F- or 0- -. One then calculates the mean energy shift

10 Dq = (dzzI c
neighbra
Ze2
(r - ri) neighbors (r
ze2
- ri)
Id,).
(6.1)

The result was in fair agreement with experiment.


As often happens, the first attempt to improve upon this agreement
completely destroyed it. This occurred in the calculation of Kleinerl?‘ on
the C?+O6- - complex, in which he used a reasonably good self-consistent
potential for the 0- - ions rather than the long-range approximation of a
pure Coulomb repulsion, as in (6.1). The Dp calculated by Kleiner had
the wrong sign. Shulman and Sugano’” have, as a matter of interest, ap-
plied Kleiner’s method to the NiFa octahedron, and have again obtained a
negative Dq. It appears that all calculations made using Kleiner’s principle
contain a large error in the negative direction.
The first great improvement was made by Tanabe and S ~ g a n 0 . l ~ ~
Their method was subsequently developed and interpreted physically by
Phillipsl79 and 0thers.l79aJ7~~J~
The difficulty inherent in Kleiner’s calcula-
D. Polder, Physics 9, 709 (1942).
176 W. H. Kleiner, J. Chem. Phys. 20, 1784 (1952).
177 R. G. Shulman and S. Sugano, Phys. Rev. (to be published).
17s Y. Tanabe and S. Sugano, J. Phys. Soc. Japan 11, 864 (1956).
J. C. Phillips, Phys. and Chem. Solids 11, 226 (1959).
179a J. C. Phillips and L. Kleinman, Phys. Rev. 116, 287 (1959).
179) J. C. Phillip and L. Kleinman, Phys. Rev. 116, 880 (1959).

Do M. H. Cohen and V. Heine, Phys. Rev. 122, 1821 (1961).


188 PHILIP W. ANDERSON

tion lay in the fact that the ionic cores of the ligand ions basically are
strongly attractive potential wells for the electrons. The effect on the un-
modified wave function, which has appreciable amplitude in the potential
wells, can easily overbalance the weaker repulsive potential caused by the
negative total charge of the ligands.
Tanabe and Sugano, however, calculated the crystal field assuming that
the d functions had been orthogonalized to the core functions, and ob-
tained practically the point-ion value, which, for the wave function avail-
able at that time, agreed with experiment.
In explanation of this fact, Phillips179 pointed out that the strongly
attractive potential well is not effective in attracting the d electrons from
the neighboring magnetic ions. This is a consequence of the fact that the
Cr d functions t#Jcr must be orthogonal to all of the filled wave functions
on the core of the 0-- ion. This effectively prevents them from taking
advantage of the attractive part of the potential. One can derive an effec-
tive potential that would be seen by the “smooth” (nonorthogonalized)
which contains, for practical purposes, only the point-ion term
part of t#JcrJ
retained by Van Vleck. This result is very similar to that obtained from
the calculation of effective potentials in the orthogonalized plane wave
method of energy band theory.
Cohen and Heine have given a very complete and satisfying treatment
of “Phillips’ theorem”.’@’ Suppose we desire to calculate the energy of a d
orbital@in a complex like Ni++(F-)s. If we assume that the 2s and 2 p orbitals
on the F- ion are much deeper in energy than the d orbital, all these
orbitals will be occupied from the start. As our wave function we must use
not @ but the orthogonalized combination

where Q I are the “core” functions 2s and 2 p on the F- ion. The wave equa-
tion for #,

(6.3)

can be shown to be equivalent to a wave equation for the “smooth” part


@ of *,
(X +
V R ) Q= EQ, (6.4)
where V Ris a nonlocal integral operator given by
VR(r, r’) = C ( E - Et)aI*(r’)QI(r).
t
(6.5)

This is an effective repulsive potential, for (6.5) is essentially positive if


THEORY OF EXCHANGE IN INSULATORS 189

Et is an energy low relative to the d function. That it nearly cancels the


deep attractive part Vt of V which binds the s and p electrons to the I?
ion can be seen from an approximate expression obtained by Cohen and
Heine, using the condition that the cancellation of V and VB be as good
as possible:

V9) at]

+
Analyzing the two terms of (6.6), we see that in calculating the mean
energy of 9, the second term represents a correction by only a small per-
centage, while the first is equivalent to

[v + VB](~’, r) N V(r> [6(r, r’) - C 9t*(r)9t(rt)].


t
(6.7)

If were a complete set of functions, (6.7) would be exactly zero,


because the sum would be a &function. We can expect that the at are very
close to complete in the region of the core potential Vt, so that the
bracket in (6.7) is small. This leaves only small residual terms in addition
to the long-range Coulomb part of the potential, which the at are unable
to represent.
Physically, one can see that the effective repulsive potential is necessary
in order that the I? ion, or any other “atom core”, not be able to bind
more electrons to itself. The fact that the core is stable implies that the
attractive potential seen by any additional electron in an orthogonal orbital
must be too weak to cause binding.
With the appearance of the result of Tanabe and Sugano, and Phillips’
explanation of it, it appeared that the questions of the source of the crys-
tal field and the form of the ligand field wave function had been solved.
The wave function was a pure orthogonalized d orbital, and the crystal
field came from strictly long-range Coulomb point ion effects as calculated
by Van Vleck.
Unfortunately, Freeman and Watson’*’ recently pointed out that this
apparently satisfactory state of affairs is illusory. The wave function for the
d-shell ion used by Kleiner and Tanabe-Sugano was not the correct Hartree
function for the CrS+ ion for which the calculations were made, but was
much too extended. Using wave functions of Watson,lszJmthe calculated
181 A. J. Freeman and R. E. Watson, Phye. Rev. 120, 1254 (1960).
I@ R. E. Watson, Thesis, M.I.T., Cambridge, Massachusetts, 1959.
1811 R. E. Watson, Phys. Rev. 118, 1036 (1960).
190 PHILIP W. ANDERSON

point ion value is found to be nearly an order of magnitude smaller than


observed Dq's. Experimental results from nuclear hyperfine splittings, 8s
well as our own superexchange theory, also argue strongly against a pure
point ion theory, as we shall see.
Before we go on to the covalent theory, the other general type, we
should mention a last attempt to salvage the free-ion theory made by
Marshall and Stuart.141 They suppose that in the solid the general form of
the atomic wave function for the d-shell ions is not correct, but must be
expanded by something of the order of 10% or so in the radial scale. This
conclusion is based to some extent on neutron diffraction data which are
not completely conclusive, but primarily on' the idea that the 0-- or F-
electrons actually will spread out onto the positive ions and help screen
out a fraction of the nuclear charge on the cation. It is hard to believe that
this effect can be either as large or as uniform as Marshall suggests, par-
ticularly because a major influence of the anion electrons probably comes
from the orthogonality effect, which has precisely the opposite result since it
multiplies the central amplitude of the wave function by a factor

and decreases the amplitude in the outer regions near the anion.
On the other hand, Marshall makes some valid points which we will
return to later. In particular, he points out that the reductions in param-
eters ascribed by O ~ e n ' ~and
6 others to covalency probably cannot be
explained in that way. They seem to require some screening effects, as we
shall see later. The experimentally measured expansion of the wave func-
tion'" is indeed a puzzling effect, but Marshall's use of this expansion seems
to represent a rather ad hoc procedure without strong theoretical justifica-
tion.

c. Theory of Ligand Field Parameters and Wave Functions: Covalent

In contrast to the ionic school of crystal field theory is the "partial


covalency" school. It was founded on a very early suggestion of Van
V l e ~ k ,was
' ~ ~expanded upon first in basic articles of Stevensa5and Owen,146
and then carried further by Orgel.165JB6
The idea underlying this school is that the greater part of the crystal
field splitting comes, not from a difference in mean Coulomb energy be-
tween point ion wave functions, but from actual modification of the wave
functions of the magnetic ion through admixture of s and p orbitals from
J. M. Hastinge, N. Elliott, and L. M. C o r k , Phys. Rev. 116, 13 (1959).
THEORY OF EXCHANGE IN INSULATORS 191

the ligands, in a way that would be natural if a partially covalent bond


existed between ligand and cation.
The basic idea is very simple. Consider an Ni+ +(F-) octahedron, and
assume the z axis lies along one Nif +-F- bond. Forgetting about the F-
2s electrons for a while, the F- 2p’s can be divided into the 2pa orbital, of
z symmetry, and the 2p7r’s, x,and y. It is clear that a covalent bond between
the 2pa orbital and the d.2 orbital I $ ( T ) a ( 2 9 - z2 - y2) along the same
axis could easily be formed. It is then supposed in the covalent theory
that the appropriate matrix element which causes this bond does exist,
but that the d function is considerably higher than the p a function in
energy. Therefore the best bonding function contains only a relatively
small admixture y of d,2 function
pu -I- ‘Yu d z 2 .
b o n d i n g ‘v (6.8)
This bonding function is then expected to contain two electrons, those
which originally occupied p , on the F-.
Now the only wave function which is left for the magnetic electrons to
occupy is the corresponding antibonding function
#antibonding ‘v dzz - (7, + 8a)pu) (6.9)
orthogonal to (6.8). Here S, is the overlap integral

8, = 1 I$d*I$p dr. (6.10)

There will be a corresponding energy shift in this orbital relative to d.2,


which one would expect to be given in perturbation theory by something
like
AEu Z rU2(& - Ep), (6.11)
and which will be such as to raise the energy of the antibonding state
(6.9) ; this is what is meant by its being antibonding. Of course, there will
be no corresponding admixture between d,, and p , because of symmetry,
but there will be a bond of d,, with p,. We can expect, however, that the
amount of covalency y, will be considerably smaller; therefore AE, will be
greater than AE,, and lead to a crystal field splitting effect between d,a
and dzyof the right sign.
Moriya (private communication) has derived a more precise version of
(6.11) which is not as useful a t this point as one might hope from its
elegant form, but which we will find important later in discussing super-
exchange numerically. Unfortunately, many of the chemical theorists,
particularly Orgel (but not Stevenss6), who have discussed this form of
ligand field theory have ignored the difference between y and y + S in
the antibonding function and in (6.11). Since the results of nuclear reso-
192 PHILIP W. ANDERSON

nance, as well as the theory, indicate that y and S are comparable, this
leads to considerable errors. Moriya has derived an elegant expression
which includes S as follows. Consider two functions 41 and 42, not neces-
sarily orthogonal, with overlap integral S and the energy matrix elements

(6.12a)

and

El2 = 1 &*H41. (6.12b)

Moriya observes that if we limit ourselves to linear combinations


* = a41 - X92, (6.13)
of these functions, the antibonding function has the maximum mean
energy of all such g’s. Applying the variational condition to

(6.14)

we obtain the expression


X2
E=E,+ ) (El - EZ), (6.15)
(a2 - X2

which in the limit of small X gives us, since a ‘v 1, an expression for E


in terms of the mean energy difference El - E2 and the percentage ad-
mixture X = y + S alone. This identity is particularly valuable for use
with nuclear resonance data, which give directly not y or S but X itself.
The “covalency” school, then, tends to ignore the direct effect of the
potential on the wave function and to concentrate on the percentage of
admixture y. This has, at first sight, a number of attractive features.
First, it is an experimental fact that some of the atomic parameters of the
ions in the complexes are reduced over the values for free atoms and ions.
Specifically, these are the spin-orbit coupling parameter146Xso and the
Slater F and G integrals which determine the multiplet splitt’ings.lssBoth
of these parameters would be decreased if the amplitude of the wave
function on the central ion were decreased. The early writers concerned
with the covalency theory noted that, neglecting S, the normalization
+
constant in (6.9) is (1 yV2)1‘2, which would tend to reduce these param-
eters. Order of magnitude estimates of ys seemed to agree reasonably well
with the effect.
A. Mukherji and T.P. Das, Phys. Rev. 111, 1479 (1958).
THEORY OF EXCHANGE IN INSULATORS 193

The F and G values can usually be obtained by a numerical fit of the


observed optical energy levels to computations from crystal field theory,
which contain the parameters Dq and the F’s and G’s. Shulman and
Sugano,ln for example, have recently made such a fit for KNiF3; five lines
gave an overdetermined fit for three parameters. The XSO values can often
be deduced from calculations of the fine structure of the ground state,
which is measured in resonance experiments. Perturbation theory gives
such splittings in terms of an expansion in XSO/A, where A is a crystal
field splitting -10 Dq (for a review see Bleaney and Bowers39).
A second attractive feature of the covalent model is the qualitative
explanation of the variation of crystal field splittings from ion to ion and
complex to complex. In particular, it is observed that crystal field splittings
tend to increase roughly by a factor of two in trivalent ions, as compared
to divalent ions, and to be even greater in quadrivalent ions. This is not
the behavior to be expected of electrostatic effects. Although the ligand
ions will be pulled closer by the greater charge in the trivalent case, this is
compensated by the smaller radius of the d function. On the other hand,
the degree of covalency will be increased considerably in the trivalent case
by the closer contact of the atoms and by a greater tendency of the ligand
electrons to transfer to a triply-charged ion. Again, Dq is observed to be
considerably smaller in the half-filled shell ions Mn+ + and Fe3+,for which
the half-closing of the d shell causes an unusually low affinity for an addi-
tional d electron. Finally, there is a tendency for the crystal field splitting
to increase with decreasing electronegativity, that is with an increasing
tendency to form covalent rather than ionic bonds; again, this is the
opposite of the prediction of the purely electrostatic theory.
While these qualitative successes seem to show that there is a con-
siderable physical basis for the idea of partial covalency, the pure covalency
theory does not stand up so well to examination quantitatively. In par-
ticular, since most of the work done with it has been of an entirely qualita-
tive and semiempirical nature, no careful analysis of the kind we have
made for the electrostatic theory can be carried out.
Obviously, the correct theory must allow both for the covalency effect
and for the electrostatic effect, as well as take into account the overlap
integral and Phillips’ theorem. The only calculation which contains all of
these physically important factors has recently been made by Shulman and
S ~ g a n ofor
l ~ the
~ substance KNiFI.
Shulman and Sugano used Watson’s182J8a Hartree-Fock functions for
Ni+ + and an analytic approximation to Froese’s1*6Hartree-Fock functions
for F-. Mean energies and off-diagonal matrix elements were calculated for
C . Froese, PTOC.
Cambridge Phil. Soc. 53, 206 (1957).
194 PHILIP W. ANDERSON

TABLEVIII. SEULMAN-SUQANO
CALCULATION

Crystal field A. x.7 A,


(10Dq) Coeff. of Coeff. of Coeff. of
Approximation (cm-1) 2s in $ z z 2p0 in +,* 2p3r in &,,

Experiment 7250 0.116 0.337 (? -0.3)


Kleiner (+ exchange) -2700 0 0 0
Point ion 1400 -
Orthogonahed 1100 S = 0.081 S = 0.111 S = 0.076
Theory 6350 0.109 0.383 0.249

the 2s) 2p*, and 2pa functions of F- and the d,z and dz,, functions on
Ni++. Only the NiFa octahedron was treated exactly, this being verified
to be a fair approximation. Three-center integrals were included; exchange
was taken into account and in many cases had an appreciable effect.
The results may be summarized in Table VIII. The experimental
values for X come from Shulman’s measurements. X is defined as the co-
eficient of the normalized molecular orbital composed of the appropriate
8 or p functions from all six ligands; X gives one a good measure of the
extent to which the appropriate function is not purely ionic. The deter-
mination of X involves measuring, either in the antiferromagnetic or para-
magnetic state, the shift in the nuclear resonance frequency of F- as we
change the mean amount or orientation of magnetization of Ni+ +. From
this we obtain both isotropic and anisotropic interactions of the F- nuclear
spin and the Ni++ electrons. These interactions are primarily the re-
+
sult of that fraction X = S y of F orbitals in the Ni+ + wave function.
From the isotropic part it is easy to abstract A,, from the anisotropic part
L,by the use of formulas for hyperfine interactions borrowed from atomic
hyperfine structure theory, and of dipolar sums. These topics are covered
in the original papers, to which the reader is referred.
The correction for 1s - 2s interference suggested by Marshall and
Stuart141 has been found not to be very large. As may be seen, the ortho-
gonalized result, that is = S, is much too small, and not far from the
naive point-ion theory, while the Kleiner theory with no admixture gives
a negative result. However, it is in fact found both experimentally and
theoretically that the values are much more than those expected from
orthogonalization alone, so that there is rather a large degree of covalency.
Somewhat more than half of the observed splitting is a consequence of the
covalency. This result provides strong support for the covalency theory,
since E‘is the most ionic of the possible ligands; one expects a considerably
greater degree of covalency in almost every other case.
THEORY OF EXCHANGE IN INSULATORS 195

It is interesting to combine Moriya’s identity (6.15) with Phillips’


theorem in order to obtain a rather complete, if approximate, theoretical
expression for the crystal field parameter in terms of the measured ad-
mixture coefficients and other convenient quantities such as overlaps
and energy differences. Phillips’ theorem shows that the effect of the ionic
potential and the purely overlap terms are to give the point-ion value
alone, to a good approximation. Thus we need add to the point-ion value
only the nonoverlap terms of (6.15). That is, we observe that the two anti-
bonding functions are given approximately by
$$ (antibonding) d2. - XWp, - Xls
+ Spr)pw - + Ss)s,
S dzo - (ypo (7s

A, (antibonding) * dW - Xprp, = dzy - ( y p r + S,,)pr. (6.16)


Here again S is the overlap integral, y the covalency, and X the in principle
observable admixture coefficient. By using Phillips’ theorem, according to
which the point ion value would be correct in the absence of all y’s, we
obtain
10 Dq cv (point ion value)
+ CXL - S& - (A;, - SL)I(Ed - E,)
+ (A2 - Sa2)(Ea- Es). (6.17)
We must realize, however, that Phillips’ theorem is not necessarily ac-
curate to terms having the rather small magnitude of the differences which
determine the ligand field parameter, and that (6.17) is probably over-
simplified in other ways as well.
The Shulman-Sugano calculation, which reconciles so well all the di-
vergent points of view on ligand field theory, nonetheless leaves a number
of serious questions. One centers about the reduction in term separations
and in spin-orbit coupling. The modification in the wave function near the
Ni nucleus caused by an admixture of, say, 2pa of Xu = S, + yo may be
shown to be given by the factor
(1 - 5 2 + y2)-1’2,

and thus the change caused by the overlap term is opposite in sign to the
covalency term, tending to cancel it. In the case of KNiF3, in fact, there is
no net change for practical purposes, since y and S are comparable. This
contradicts the experimental results, particularly on term separation. We
suggest that the term separation may well be sensitive to screening of the
d-d electron interaction by the F- electrons, so that this difficulty may not
be very serious.
196 PHILIP W. ANDERSON

The term separation and spin-orbit parameter difficulties may well be


* ~ the admixture y of d
solved partially by a suggestion of M u r a ~ *that
functions into the F- wave functions can be large enough to afford an ap-
preciable screening effect on the d electrons. The size of the required y is,
however, considerably greater than that observed.
Note added in proof: The final version of the Shulman-Sugano calcula-
tion gives excellent agreement on these parameters as well, further con-
firming the speculations of the covalency school.
A serious problem, and one which continues to be a difficulty with
ligand field theory, was discovered by Shulman188(see also Shulman and
KnoxlWJB1) in making comparisons of nudear resonance results on fluorides
of Mn++ and Ni++. KMnF3, KzNaCrFe,and KNiF3 give the most clear-cut
results. It had been known since the early results of T i r ~ k h a mand
~ ~of Shul-
man and Jaccarino44-46 that the hyperfine interactions in MnFz were, rather
surprisingly, predominantly isotropic. The same result is found in the much
simpler crystal structure of KMnFa, where the F is in contact with only
two Mn ions, along the same cube axis in opposite directions.
From the isotropic interaction one could derive a percentage of F- 2s
character for the Mn++ d functions, the result being about 1.4%. From
the anisotropic hyperfine interaction one could derive the difference in the
amount of 2% and 2p,, which was actually somewhat smaller, namely
-+%. On assuming, as was natural, that the 2p's were by far more heavily
occupied, this leaves 2p, less than 29, which is not only unlikely but dis-
astrous for the covalent theory of ligand fields, as well as for the super-
exchange theory.
Shulman repeated this experiment both with NiFz and with KNiF3.
In these substances, one finds that the anisotropic part is not much smaller
and that therefore the amount of 2p, is much greater than that of 2s
(the actual figures in KNiFa are given in Table VIII: 2p,,, 11%; 2s) 1.4%).
The only reasonable interpretation Shulman could offer is as follows.
In Ni++ no magnetic bonds can be formed with 2 p , electrons, because
both magnetic electrons are in d, orbitals. On the other hand, in Mn++
both p ,and p . orbitals can participate because both d, and d , orbitals are
magnetic. Thus one can suppose that the reason for isotropy in Mn+ + is
not that the amounts of 2 p , and 2p, are small but that
2p,Bh+ + * 2p,,aan++ N 11%.
That is, for some unexplained reason the d,,, p , interaction is just as great
as the d,z, p,, one. This conclusion was verified by measurements on
KPNaCrF6,in which a large p ,-type anisotropy was observed.
It is not at all surprising that the d,,, p , interaction should be com-
18' T.Murao, Progr. Themet. Phys. (Kyoto) 21, 4 (1959).
188 R. G. Shulman, Talk at Oxford Conference on Exchange, Oxford, England, 1959.
THEORY O F EXCHANGE I N INSULATORS 197

parable with the p,, interaction in order of magnitude, since the overlap
integral between dZyand p , is only 25-50% smaller than that for d,n - p,.
It is, however, extremely surprising that these numbers should cancel to
within a few per cent, as they do in the Mnff compounds. We note in
Table VIIL that the Shulman-Sugano calculation leads to a p J p , ratio of
about unity; thus it appears that their complete calculation has in some
way solved this problem. The physical basis of this solution must await
a more complete analysis of their results.
The Shulman-Sugano calculation was for the NiFe octahedral complex
alone, and gives no straightforward suggestion for explaining the trends of
the crystal field splittings with the variation of anion or cation. Since
covalency plays a major role, we expect that the explanation of the trends
may follow relatively closely that of the “covalent” school of semiempirical
ligand field theory.
In summary, then, we may say that with the work of Freeman, Watson,
Sugano, Shulman, and Moriya, the one-electron theory of the isolated
magnetic ion, that is, the basis of ligand field theory, is just beginning to
emerge from the roughest form of empiricism. As a guide for later work,
their a priori calculations show that the semiempirical covalency theory is
not off by large factors but that overlap, the Phillips-Cohen-Heine theorem
and its unknown errors, and various other even more subtle factors, play
an important role.
True many-body effects are almost entirely missing from these con-
siderations. We mentioned the possible effect that screening might have in
reducing F and G integrals. Marshall’s suggestion concerning a major
screening effect on the scale size of wave functions and on X does not
represent a many-body effect because it is contained in the Hartree-Fock
theory. It is probable that screening does not play a very great role for
calculations such as that of Shulman and Sugano because the energy
denominators in the latter are quite large. In determining such integrals
as U and internal exchange, on the other hand, it is probably of considerable
importance.

7. SEMIEMPIRICAL AND THEORETICAL


APPROACHES CALCULATIONS

a. Early Semiempirical Work

Since the first hypotheses of N & P concerning spin arrangements in


ferrites, and the first neutron-diffraction results, attempts have been made
to find regularities in the exchange constants deduced, or guessed, from
the observations. Often the attempts have used a theory as a guide, most
frequently the theory of Anderson’s first paper.@ I n that paper, it was
198 PHILIP W. ANDERSON

suggested that superexchange between ions having d shells which are at


least half-filled might be antiferromagnetic, whereas between shells that are
less than half-filled it might be ferromagnetic. For many years this rule a p
peared completely inapplicable. I n particular the many compounds of
Cr3+are almost uniformly antiferromagnetic.
A semiempirical conclusion suggested rather tentatively by Anderson,
and carried a great deal further by many authors, was the 180" direc-
tionality which is implied in the assumption that the major part of the
admixture involves the d,2 and p,, orbitals. This type of admixing sug-
gests that there may be a stronger interaction between two spins having
a ligand directly between them (180"position) than if the ions are at right
angles with respect to the ligand ion.
This idea was strongly supported by the consideration of the spin
arrangements that were observed in the first few years, especially if no
careful study were made of actual values of exchange integrals. The divalent
oxides-Mn0, FeO, COO, NiO, CuO-all exhibited the "second kind" of
magnetic structure1*9 in which next-nearest neighbor ions, having 180"
configurations, are aligned antiparallel. In contrast, nearest neighbors,
having 90" configurations, are indifferent. These concepts were extended
to the ferrites, in which the observed structure is almost always an "A-
site" vs "B-site" antiferromagnetic alignment, even though the B sites
are rather close to each other. The A-B connection is an approximately
125" bond through each of a pair of 0-- ions, the A-A distances being
large. It was suggested that the B-B interaction should be nearly zero
because of directionality and because this is a 90" configuration.
Again, in the difluorides in the sequence Mn++ through Ni++, the
two nearest neighbor spins along the c axis are parallel and the eight next-
nearest neighbors antiparallel. The latter have -135" bonds, the former
-90".
I t is interesting to note that this rule has been accepted so generally
that it actually proved useful in the discovery of an important tech-
nical magnetic material."
Unfortunately, this remarkably successful empirical rule has come to be
much less generally trusted, for both theoretical and experimental reasons.
In the "canonical case" of MnO, it was known from the start, in fact, that
the 90" nearest-neighbor exchange interaction was quite large. This
conclusion was drawn from the application of molecular field theory to the
ratio O/T, in this compound.s2 Molecular field theory gives an interaction
for the 90" configuration of the nearest neighbors which is twice as large

u0J. S. Smart, Revs. M&m Phys. 26, 327 (1953).


THEORY OF EXCHANGE I N INSULATORS 199

as the nnn, 180”interaction. This can be questioned on two grounds: (a)


the x vs T curve is not measured over an adequate temperature rangez9;
thus 8 is not a well-determined parameter. (b) The theory of Stern126Jn
shows that spin-wave and short-range order effects will tend to decrease
T,strongly in this region, increasing the O/T, ratio and thus the apparent
ratio of nn to nnn exchange. Recent measurements of Coles et (see
also Owenla4)on Mn ion pairs present as impurities in MgO support an
approximately one-to-one ratio, which is not in serious conflict with the
values obtained from molecular field theory for the foregoing reasons.
Thus, observed spin patterns in the MnO type of structure do not
indicate the relative magnitudes of 90” and 180’ superexchange reliably.
We should like to suggest that this is a rather general phenomenon. The
reason for the situation in MnO is that the lattice is geometrically un-
favorable for nearest neighbor ordering. Each nearest neighbor of a given
ion has four nearest neighbors which are themselves nearest neighbors of
the original ion. This makes it very hard to “fit in” an antiferromagnetic
structure. Even the best arrangement has only an 8 to 4 ratio of anti-
parallel to parallel nearest neighbors. In fact, spin-wave theory shows that
this structure is unstable if there is no next-nearest neighbor interac-
tion.126JnThus, it is natural that next-nearest neighbor ordering should be
rather more frequent.
Precisely the same kind of phenomenon occurs in the ferrites. As shown
by Anderson,lQothe B-site ordering with nearest-neighbor interactions alone
is unstable even in the Ising model. When B-B interactions are large
relative to A-B, the orderings are complex, as discussed in the chapter on
the molecular field theory. Thus the A vs. B ordering will occur in the
majority of cases, even though this does not imply that the B-B inter-
actions are particularly small.
That these difficulties occur most often in the presence of 90” bonds is
not a coincidence. In most cases the oxide lattices may be based on an
approximately close-packed lattice of oxygen ions, with metal ions in
octahedral or tetrahedral interstices. 90” bonds occur only between filled
octahedral interstices. These interstices themselves have a close-packed
structure, and therefore have the “unfavorable” nearest-neighbor arrange-
ment of the MnO lattice. Thus, any subsystem of such sites will also
tend to have bad nearest-neighbor arrangements.
It is nonetheless true that an analysis of experimental data often,
although not always, indicates a very small 90” interaction (Smart,4
Part V)-e.g., in MnF2. We can only ascribe this to compensating ferro-
and antiferromagnetic effects, as we shall discuss later.
190 P. W. Anderson, Phys. Rev. 102, 1008 (1956).
200 PHILIP W. ANDERSON

From a theoretical point of view, two factors have tended to weaken


the argument for 180" superexchange. The first is wider realization of the
fact that there is an appreciable dw - p , admixture. This was emphasized
particularly by Shulman and Knox1@JJ61in explaining the experimental
atlisotropic hyperfine coupling in fluorides. Shulman also commented on
the connection with superexchange.'@This point has also been taken up by
Marshall and and by Casselman and Keffer.19' They have made
reasonable numerical estimates of the effect and have shown that the 90"
to 180"ratio is not likely to be different from unity in Mn by an order of
magnitude. This, however, is valid only in cases in which the d,, orbitals
are occupied by spins, notably Mn+ + and Fea+.The cases Ni+ + and Cu+ +,
which involve d,2, should and do show considerably greater directionality.
The other important theoretical point is the fairly strong admixture of
s state found in the fluorides experimentally and theoretically (see Sec-
tion 6 ) . We shall see that this subtracts from the 180" exchange effect,
whereas it gives a large part of the 90" effect for d, bonds, as in Ni++.
The measurements on the fluorides suggest that the two interactions should
be comparable in them. The situation in the oxides may, however, be quite
different.

b. The Goodenough-Kanumom' Rules

A considerably more satisfactory system of semi-empirical rules was


developed over a period of years by Goodeno~gh~~~J93 and Kanamori. lg4
They were stated most clearly by Kanamori. Wollan et uZ.lg5 have made
important applications.
These rules have the important feature that they take into account the
occupation of the various d levels as determined by ligand field theory.
In addition,
The importance of ligand fields was first suggested by P01der.l~~
they are directly related to the prescriptions about the sign of super-
exchange given in Anderson'sBBpaper. The angular relationship with the
intermediate ligand ion does not play as important a role.
First, one observes which of the d orbitals contains the magnetic elec-
trons according to ligand field theory. This is simple in the octahedral
ions that are isoelectronic with Cra+,Mn+ +, and Ni+ +. The first has mag-
Ig1 T. N. Camelman and F. Keffer, Phys. Rev. Letters 4, 116 (1960).
191 J. B. Goodenough, Phys. Rev. 100, 564 (1955).
loa J. B. Goodenough, Phys. and Chem. Solids 6, 287 (1958).
lg4J. Kanamori, Phys. and Chem. Solids 10, 87 (1959).
1%E.0.Wollan, H. R. Child, W. C. Koehler, and M. K. Willrimon, Phys. Rev. 112,
1132 (1958).
198 I>. Polder, J . phys. radium 12, 274 (1951).
THEORY OF EXCHANGE I N INSULATORS 20 1

netic electrons in d,, orbitals alone, the second has them in all, and the
third has them only in d.2. For other ions, ligand field theory must be
studied in relation to any distortion from cubic symmetry in the sur-
roundings.
I n a form slightly modified from the original Goodenough-Kanamori
formulation, the rules are the following.
1. When the two ions have lobes of magnetic orbitals pointing towards
each other in such a way that the orbitals would have a reasonably large
overlap integral, the exchange is antiferromagnetic. There are several
subcases:
a. When the lobes are d,, type orbitals in the octahedral case, par-
ticularly in the "180" position" in which these lobes point directly towards
a ligand and each other, one obtains particularly large superexchange.
b. When d,, orbitals are in the 180" position relative to each other, so
that they can interact via p , orbitals on the ligand, one again obtains
antiferromagnetism.
c. In a situation involving a 90" ligand in which one ion has a d 2
orbital occupied and the other a d,, occupied, the p , for one is the p, for
the other and one expects strong overlap and thus antiferromagnetic ex-
change. (Actually this is not a case discussed in these terms by any of the
authors mentioned, but we feel follows logically from the same principles.)
2. When the orbit,alsare arranged in such a way that they are expected
to be in contact but have no overlap integral, most notably, a d,* and a
d,, state in the 180" position for which the overlap is zero by symmetry,
the rule gives ferromagnetic interaction. This interaction, however, is not
usually as strong as the antiferromagnetic one. Some of the authors men-
tioned put this condition in an equivalent form by saying that the effect
of an empty d orbital in "contact" with a magnetically filled one is ferro-
magnetic.
These two rules seem to explain almost all the spin-pattern data for a
wide variety of substances. The reader is referred to the original papers and
to Nagamiya' for specific applications.
It is already clear from the foregoing that the explanation of the
regularities is quite straightforward when regarded from the point of
view of the theory of Sections 4 and 5. The first point we notice is that,
according to the theory, the antiferromagnetic kinetic exchange effect can
only arise between two orbitals which have a finite transfer integral b
connecting them. Moreover the exchange integral is proportional to b2.
This means that the two orbitals must have the same symmetry in the
region of overlap, which is the same as the requirement of Rule 1. If the
orbitals do not have the same symmetry, so that there is no transfer inte-
202 PHILIP W. ANDERSON

gral b, that is if they are orthogonal by symmetry, it is still entirely possible


for them to have a potential exchange integral J . Therefore we expect only
ferromagnetism, albeit a somewhat weaker effect.

c. Further Regularities Suggested by Theory

It may well be worthwhile summarizing a few of the qualitative con-


clusions to which the theories of Sections 4 and 5 would lead and which
are, on the whole, borne out by experiment.
First, let us consider a straightforward case of 180' exchange through a
ligand such as F- or 0--. For simplicity b e consider the cases of d3, d6,
and d8 ions only. The others will have intermediate values, which to a great
extent are controlled by the precise symmetry of the lattice.
(1) da. (Cr3+). Here dW, d,,, and d,, are occupied magnetically, d.2 is
entirely unoccupied. If z is the ligand axis, d,, and d,. will mix with p ,
and p , r orbitals, which will lead to two couplings which might be written
schematically as 2[2(b,) 2 / U ] .At the same time there is only ferromagnetic
potential exchange between xz on one and yz on the other and vice versa.

{[TI
We may call this J,. The resulting interaction energy is
E(da) a 2 2(b,I2 - 2 J , } .

Of course, the fact that d,, on atom 1 contains some small admixture
of d,. on atom 2 suggests that there should be a d,, - dzy exchange effect.
We expect this to be small; probably most of the exchange effect comes
from the wave function on or near the ligand, which for dzyvanishes almost
exactly.
2. ds. (Mn+ +). Here there is considerably greater complication. The
d, orbitals contribute the whole of (7.1) to begin with. There is also a
d,2 - d,2 contribution which, although only a single term, probably is
dominant, because of the large overlap of z2 - pu and 9 - s. This term is
(2bn2/U).On the other hand, there is an even greater multiplicity of po-
tential exchange terms in this case, since d,* has an exchange integral
J+w with both d,. and d , and vice versa. This makes four interactions in
all; thus the net result for dSis

Moriya (private communication) first called attention to the fact that


there are an unusually large number of ferromagnetic components in (7.2)
when compared to either (7.1) or the case associated with d8.
THEORY OF EXCHANGE IN INSULATORS 203

(3). d8. (Ni+ +). Here nothing but d., is occupied and we obtain only
E(d8) = 2bU2/U. (7.3)
In comparing (7.3)) (7.2)) and (7.1) we see that (7.2) has only two
rather small extra antiferromagnetic terms, but it does have six ferro-
magnetic integrals more than (7.3). As Moriya has suggested, it is very
likely, as a result, that the d6 exchange integral is even somewhat less than
the da integral, which would make the relatively low value of the exchange
in MnFz and MnO relative to NiF2 and NiO easier to understand (see the
discussion of this problem in Anderson'o). On the other hand, Eq. (7.2)
probably is somewhat larger than (7.1)) because the extra term bU2/U
probably outweighs the additional J,is. Thus we come to the fist conclu-
sion, that
Jds > Jd5 > Jd* (180") (7.4)
for 180"bonds.
At 90") on the other hand, we reach quite a different conclusion. In
this case the d8d8 integral has only the s-admixture antiferromagnetic
component which seems both empirically and theoretically to be rather
small (see Section 6). I n this case the p, function for one cation is the p ,
for another and contributes only a ferromagnetic effect. The case of d6, on
the other hand, leads to a relatively large antiferromagnetic result from the
interaction of d,,, on one anion, with d,a as well as d,a,l on the other.
Finally, in the case of 90" alignment for the da octahedral system some of
the interaction probably comes from direct overlap of the dzuorbitals with
each other. This is a variable effect and is difficult to estimate. The p ,
function on the ligand for one cation will be a p, function for the other,
so that the effect obtained through the ligands is ferromagnetic. Indeed
ferromagnetism is often observed.lg4We have
Jd5 << Jda (90") ; (7.5)
J ~may
s change sign.
As we have mentioned, in the best-known case of the oxides the rela-
tion (7.5) is well borne out by the relative magnitudes of the interaction
in NiO and MnO as far as d8 and ds are concerned. On the other hand,
JW is surprisingly small in MnF2.
Thus far this discussion has assumed that transfer integrals and other
atomic quantities vary only with orbital symmetry. There are, of course,
major chemical differences to be considered.
The most important and best-documented difference is that between
divalent and trivalent cations. The fact that the crystal field parameter
usually is about twice as large for the trivalent ion as for the corresponding
divalent ion indicates that the trivalent ion is considerably more covalent
204 PHILIP W. ANDERSON

than the divalent one, especially in the covalent form of ligand field theory
(see Section 6). This implies that b will be substantially larger for the tri-
valent than for the divalent case, so that
Jtrivalent > Jdivslent, (7.6)
other things being equal. This rule is borne out in a comparison between
Mn+ + and Fe3+oxides. The various ferrites, garnets, and hematite show
that the exchange integrals for Fe3+are several times as large as those for
Mn+ +. (For evidence, see Anderson.70)
Rule (7.6) has an interesting effect in countering the influence of rule
(7.4). It happens that most substances in which ?r interactions are im-
portant-notably CrS+ compounds-are triply ionized, whereas the d,,
interactions usually involve divalent ions, except in the case of Fea+.
Thus, for a long time, there appeared to be no ?r - u difference, and no
qualitative difference between the two halves of the d shell. We would now
interpret this as a case of cancellation between rules (7.4) and (7.6).
Rules (7.4) and (7.6) leave the Fea+ ion in a unique position as
the only common trivalent ion with u interactions. Thus it is no coincidence
that all the technically important insulating ferrimagnetic matrials with
high Curie points and relatively stable properties contain Fe+ + + ions as
an important component.
The other important chemical ingredient is of course the ligand. The
role of the ligands is considerably more obscure than that of the cations,
as we saw in our discussion of ligand field theory. One of the reasons for
this is the rather surprising fact that the crystal field parameter Dq does
not vary rapidly with the expected degree of covalency of the cation-
ligand bond. For instance, Dq for F- is only a few per cent smaller than
that for 0--, which is essentially identical with that for the oxygen end of
the water molecule. Nonetheless, a priori we should expect a rather rapid
increase in covalency along the series F--0- --8- --Se- -, and there-
fore a corresponding increase in b and in the antiferromagnetic exchange
integral, as is indeed observed. This must be considered to be a rather ad
hoc assumption, however, in the absence of nuclear resonance data for
any of the more covalent series members and in the absence of any evi-
dence arising from Dq. Nonetheless, we may consider this a rule from a
semiempirical point of view, placing it in the form
JF- < Jo-- < Js-- < J s e - - . (7.7)
It should also be noted that theoretically one should guess that both s
and ?r bonding effects should be reduced relative to the simple u effect in
the higher members, as is observed experimentally in the series MnO-
MnS-MnSe.
THEORY OF EXCHANGE IN INSULATORS 205

d . Quantitative Calculations of Superexchange

This section can be extremely brief, for very little truly quantitative
work has been carried out in superexchange theory.
We can dismiss, with only a few words, almost all of the quantitative
work undertaken before the basic principles were made clear by Kondo,1*8
L o ~ d i n , and
l ~ ~ Anderson.70 Aside from rough estimates in Kramers112and
Anderson’s papers,6athe first quantitative calculations are to be found in
Yamashita’s p a p e r ~ . ~These
~ ~ J ~follow
* the ideas of Andersone6very closely.
As we explained earlier, Anderson’s mechanism corresponded to only one
of many terms in the higher-order perturbation series method of analyzing
the problem. I n particular, the Mn-0 transfer integral is correctly used
in Yamashita’s calculation to get the degree of admixture of the p , and d,
function, that is the quantity we called “A” in Section 6. However, the
0-Mn exchange is then calculated by a procedure which includes only
direct overlap. Thus while the answer turns out to be correct in order of
magnitude, there is no reason to expect even a reasonable degree of ac-
curacy.
The quantitative work of Pratts7suffers from a basic difficulty in under-
standing the role of orthogonality-particularly, when the assumption is
made that one can find an orthogonal set of one-electron functions which
diagonalize the one-electron Hamiltonian and still are localized. Within
this framework, Pratt, in contrast to Nesbet,’” has carried through a
correct calculation. Thus we should expect his results to be rather small
and not of the right qualitative character. We have already discussed
Nesbet’s work within the same framework.
Koide et ~ 1 . appear
~ 9 ~ to have calculated one of the terms of what is
called the indirect exchange effect in the present paper. Their estimate of
this term is too large and probably of the wrong sign, the main difficulty
being the neglect of the McWeeny orthogonality effect in the evaluation
of certain integrals. The formalism is rather complicated and tends to
restrict too rigidly the manifold of excited states which are taken into
account.
Much more complete attempts were made by Yamashita and K ~ n d o ’ ~ ~
and Keffer and Oguchi140within the old framework. Yamashita and Kondo’s
calculation was complete in principle, but it was nearly impossible for them
to come to any quantitative conclusions except as to order of magnitude.
because of the large numbers of parameters which enter into the ab initio
calculation from atomic wave functions. It also happens that they over-
lg7J. Yamashita, Progr. Theoret. Phys. (Kyoto) 12, 808 (1954).
198 J. Yamashita, J . Phys. Soc. Japan 9, 339 (1954).
leg S. Koide, K. P. Siha, and Y. Tanabe, Progr. Thewet. Phys. (Kyoto) 22, 647 (1959).
206 PHILIP W. ANDERSON

estimated the polarization and Nesbet-Goodenough terms. They did not


make use of the McWeeny relationship.16 The Keffer-Oguchi paper also
attains no really quantitative conclusion because of uncertainties in wave
functions.
K o n d ~ was
’ ~ ~the first to present a numerical estimate of the exchange
integral using ideas related to those of Section 4. His calculation is limited
to the ordered antiferromagnetic and ordered ferromagnetic states, in
which it is of course valid to proceed, as he did, with a pure band-theoretic
calculation. In the antiferromagnetic state Kondo allowed an admixture
coefficient C k between the wave functions on atoms with upward and
downward spin in the superstructure. He gives an essentially similar
expression to ours for this coefficient, which we have called b / U . His
final energy difference between the purely ferromagnetic and purely anti-
ferromagnetic states contains a quantity of the form - b2/U for the largest
term. He estimated that this has a value of 0.0069 ev for NiO, without
claiming great numerical accuracy. His estimate of b originates in perturba-
tion calculations of the second-order transfer integral Ni-0-Ni. He
used a reasonabIy good energy denominator but a rather approximate
nodeless Slater function for the d orbital of Ni. He obtained -0.045 for
the admixture parameter A, which is certainly in a reasonable range.
Attempts to calculate superexchange quantitatively through the new
formalism naturally have centered on the calculation of the effective
transfer integral b. One may estimate the quantity U in various ways but
the estimates do not vary by more than a factor 2. K o n d ~for , ~instance,
~~
takes the atomic value reduced by e2/Rij, the latter representing the elec-
trostatic correction for the fact that in superexchange the electron transfer
takes place not to infinity but to the next atom. Kondo’s result for Ni++
is 12.55 ev. Anderson reduces this number further by making an electronic
polarization correction and a correction for reduced covalency, obtaining
about 6.3 ev. This probably is an excessively low value. Some of the elec-
tronic polarization may not relax at such a high rate, and the covalency
correction is probably not justifiable. In any case U enters only to the first
power, and in most cases can be estimated to be -9 f 2 ev.
In principle the calculation of b simply involves a one-electron problem
of ligand field theory. In practice, however, it is very difficult. Thus far two
attempts have been made to solve this problem. The first, by Anderson,’O
represents an uncritical application of the “covalent” variety of ligand
field theory. More recently Moriya, Shulman, and Sugano have carried out
a calculation based on the Shulman-Sugano ligand field theory of KNiFa.ln
Both calculations make use of the fact that, even though the ligand
field parameter 10 Dq may not be identical to b, the parameters are closely
THEORY OF EXCHANGE I N INSULATORS 207

related. To see this relation it is perhaps simplest to present the analogue


for "b" to Moriya's identity of Section 6 for 10 Dq (Eq. (6.15)). First we
observe that b can be calculated either as the off-diagonal matrix element
between two neighboring ligand field theory wave functions $(r - Ri)
and #(r - Ri),or as half the diagonal difference in energy between the
symmetric and antisymmetric linear combinations, namely,
b = [+(r - Ri) I X I J.(r - R i l l
= t{CJ.(r- RJ + rL.(r - Rj) I X I $(r - Ri) + #(r - Rj)]
- [#(r - Ri) - $(r - RJ I X I rL(r - Ri) - $(r - RJ]}
= +(E+- E-). (7.8)
We know by symmetry, however, that the two combinations are of the
form
J.i +
$i = .+/a(4,(1) 4d(2)) - 4%4a +
(7.9)
*. - *.3 = .-/-\/Z(4d(l> - 4dd(2)) -

The factors of .\/z are necessary to make A, and A 8 approximately equal to


their values in the localized functions J.i. We also know again that these
are the antibonding functions, and thus we can apply the condition that
the energy be a maximum with respect to variation of A, and obtain, by
analogy with Moriya's' previous identity,
E- S E[+d(l) - 4d(2)] + A,'(& - E,)
E+ 'v E [ & ( l ) &(2)] A,"(Ed - E8).
A rough approximation is obtained by assuming that, in accordance with
Phillips' theorem, the energy difference of the two d combinations cancels
the straight orthogonality terms. Thus
b = (Ap', - Sp 'U) (Ed - E p ) - (A,' - 88') (Ed - En). (7.10)
Here the coefficient A is the sum of the overlap and covalency coefficients,
A,, = s
pa + Ylw
(7.11)
A, = s, + Ye,
and is the coefficient measured in the nuclear resonance experiments.
These A's are those in Table VIII, multiplied by a factor l/G. Those in
Table VIII correspond to molecular rather than to atomic pu and s or-
bitals.
208 PHILIP W. ANDERSON

As we saw in Section 6, the crystal field parameter 10 Dp is given by


10 Dq = (point ion value)
+ [(rw) (&u + Aw) - Y P * ( ~ P+* Apr>I(Ed - EP)
+ (YJ (8,+ A,) (Be - E p ) . (6.17)
The covalent term is the sum of contributions similar to those in (7.10).
In assuming the validity of the pure covalent theory Anderson effec-
tively neglected (a) the point ion term in 10 Dq; (b) the s and ?r terms both
in (7.10) and (6.17). b and 10 Dp then become identical.
These assumptions are not too seriously in error in the oxides, if not in
the fluorides, because the different terms neglected have various signs.
The point-ion term and, particularly, the s-admixture, tend to increase 10
Dq relative to b, whereas the ?r bonds tend to hiwe just the opposite effect.
In actual fact, the computed TN obtained by Anderson in NiO was
920") compared to the observed value of 520". If one used a somewhat
more realistic value of 7-8 ev for U rather than 6.3 as he did, the agreement
would be even closer, to within 50%.
Anderson also worked out a comparison of the exchange integrals in a
number of oxides and fluorides of various transition metal ions. He assumed
the simplest form of covalent ligand field theory, and neglected any direct
exchange effects. The rough comparison proved to be reasonably satis-
factory because the relative values of b can be expected to follow those of
10 Dq fairly closely, but it was not quantitatively exact. Little that is not
contained in the series of rules we have listed in Section 7b can be learned
from this comparison. We refer the reader to the original paper.
A much more accurate comparison with theory was made by Moriya,
Shulman, Sugano, and Anderson in the KNiF3system (private communica-
tion). In this case the A's in (7.10) are essentially quantities measured by
nuclear resonance and thus the only theoretical values necessary are the
overlap terms and the numbers Ed - E, and Ed - EP.Nevertheless, since
the purely theoretical predictions for the A's were in essential agreement
with experiment, those values (last line, Table VIII) were used, in order
to make the calculation completely nonempirical. Shulman and Sugano
calculated the energy differences and found them to be 7.6 ev ( p ) and
30.5 ev (s). The predicted J is 76" as compared with Smart's experimental
J of 45°K.The difference could easily be accounted for by the subtraction
of a ferromagnetic potential term of smaller magnitude, which tends in
the correct direction and may be of the right magnitude. Thus the agree-
ment is adequate, since the theory may contain as much as -30% errors
in the b values and in U.The most likely sources of error are the p - d
THEORY OF EXCHANGE IN INSULATORS 209

energy difference (the s term is small in this formulation) and failure of


Phillips’ theorem.
The final conclusion, then, is that the calculation seem to come close
to giving a quantitative verification of the theory. However, it seems wise
not to claim better than about 100% accuracy for the theory in view of the
uncertainties.

VII. Double Exchange

The entire literature of the theory of double exchange, which is con-


cerned with the exchange process involving free or quasi-free d-band car-
riers in mixed-valency oxides postulated by Zener in 1951, consists of
three papers: those of ZenerlBB Anderson and Hasegawa12*and deGennes.201
The last paper contains a remarkably complete and clear summary of the
previous work; thus we can be quite brief here and refer the reader to that
paper for details.
Zener’s theory focussed attention on a series of measurements by
Jonker and Van Santenm2-m on the series (Lal-zCa,) (Mn:LMn:+) 0 s .
The oxides at the extreme ends of this series (x = 0 and 1) are quite

mediate compositions, in this case between x -


good insulators and are purely antiferromagnetic. For a range of inter-
0.2 and 0.4, the con-
ductivity is, however, several orders of magnitude higher, and the materials
axe ferromagnetic. There appeared, in fact, to be a rough correlation
between the conductivity and the ferromagnetic Curie point.
Zener’s explanation of this was brief but, apparently, correct. As we
have seen, each ion will remain precisely Mna+ at x = 0, real hopping
motions of the electrons being prohibited by the Mott correlation effect.
At finite values of x, however, the Mn4+ions, which represent holes in the
Mna+shell, are much more free to move, and give a large conductivity. If,
however, we assume that the “hopping integrals” b which allow the holes
to move are rather small compared to the intra-atomic exchange integrals
which hold the spins of all d electrons on a given ion parallel, the hopping
can only take place between pairs of ions on which the Mn4+ion core spins
are parallel. Only then will the energies on the two sites match. This
provides a type of exchange mechanism which holds the two spin cores
parallel, for the free holes can of course gain kinetic energy in an amount
e b by hopping. We estimate a magnitude --b per free carrier.
*O0P. W. Anderson and H. Hasegawa, Phys. Rev. 100, 675 (1955).
zolP.-G. de Gennes, Phys. Rev. 118, 141 (1960).
zm G. H. Jonker and J. H. Van Santen, Physica. 16, 337 (1950).
G. H. Jonker and J. H. Van Santen, Physica 16, 599 (1950).
*M G. H. Jonker and J. H. Van Santen, Physieo 19, 120 (1953).
210 PHILIP W. ANDERBON

Anderson and Hasegawa examined much the same mechanism in a


more complete formal theory of the two-ion situation. Their use of Serber’s
method to discuss the transfer process was unnecessary in view of the work
of Section 4 concerning the transfer integral b for superexchange. One
need only treat the problem of two magnetic ions at R1 and R2 in which d
orbitals &(r - R1) and &(r - Rz) are connected by a one-electron matrix
element b12, which on modern estimates might have a value -0.5-1 ev
rather than -0.1 ev as they assumed.
It is probably valid to assume as the investigators did that the re-
mainder of the electrons, the (dz,)3Mn4+core in the particular case under
study, are strongly coupled to give a net Bpin vector S (2 in this case) and
that only the extra dz: electron is mobile. With this assumption the two-ion
problem may be solved exactly. It is instructive to discuss the principle of
that solution.
Under these circumstances only four states are connected by matrix
elements. These are the states in which the surplus electron is on ion 1 or 2
and has spin parallel or antiparallel to the ion core spin S1 or S2of the ion
on which it finds itself. It is then shown that the Hamiltonian matrix is

Here J I is the intra-atomic exchange integral which couples the spin of


the extra electron to the ion core spin S;cos 0/2 is effectively the cosine
of half the angle between the spins Sl and Sz. The latter may be given the
precise quantum-mechanical definition

where s is the spin of the extra electron. This reduces to the correct value
when we treat the spins S1 and SZclassically, as de Gennes did in the con-
sideration of a macroscopic spin system.
The meaning of (VII.l) is rather obvious. It says that whenever the
two core spins are not parallel, the transfer integral, which carries the
electron from one site to the other without reversing its spin, connects the
state parallel to S1on ion 1 with both the states on ion 2, whether parallel
or antiparallel to S2.
THEORY OF EXCHANGE I N INSULATORS 211

It is possible to solve for the eigenvalues of (VII.l) exactly, but the


result is not of much interest. There are two limiting cases:
1. J I S << b. In this case the exchange effect represents only a weak
perturbation on the motion of the electron. The extra electron is analogous
to a free electron in a metal, and the two-ion model is not very edifying.
The problem should be treated in terms of polarization of the free electron
gas.
2. b << J I S , or at least b sin (O/2) << J I S . This is the true “double
exchange” situation in which the unique features of the mechanism stand
out: In this case it clearly is correct to treat only the lower eigenstates of
the intra-atomic exchange interaction, and to neglect the mixing with the
+
J I ( S 1) term; the result is
E = - J ~ S b cos (e/2). (VII.3)
This is the energy associated with modification of the transfer integral b
to take account of the fact that the two states between which hopping
takes place are not parallel. The physical process is that the “ d band”,
which is just the symmetric and antisymmetric combination of the states
on the two atoms in our simplified case, is wide when the spins are parallel,
and narrow when they are antiparallel. In this limiting case the gen-
eralization to a larger system is very simple. We modify the transfer
integral for the carriers by the factor cos O/2, O/2 being the angle between
neighboring spins.
In the discussion of the actual mixed crystal system Anderson and
Hasegawa were probably in error in assigning an unrealistically small
value to b. As we have seen from superexchange, this must be -1 ev,
not -0.1 ev; thus the observed transition temperature k T , is not of order b,
as one might naively predict. The small value of b leads, for example, to
the incorrect prediction that the high-temperature susceptibility does
not obey a Curie-Weiss law C / ( T - O ) , but instead follows the ordinary
Curie law C/T.That is because for a given value of cos 0/2 (VII.3)
predicts that the energy of as many states will be raised as lowered.
De Gennes has used both the same mechanism and formula (V11.3),
but has given a much more complete and correct discussion of the physical
situation. For one thing, he points out that b >> kT,.Therefore, only the
states of lower energy with the negative sign in (VII.3) are appreciably
occupied.
De Gennes’ discussion is divided into three parts. The first is concerned
with the ground state, taking into account the fact that double exchange
will always compete with antiferromagne tic superexchange interactions
212 PHILIP W. ANDERSON

of a similar order of magnitude. He shows that the most usual state will
have a canted antiferromagnetic arrangement with a resultant ferromag-
netic moment. The second part develops a rather ingenious version of
molecular field theory which is appropriate for the behavior at higher
temperatures, and shows that a canted-antiferromagnetic or canted-
ferromagnetic transition point TI usually occurs below the true N&l
or Curie point, TN or T,,of the transition from the paramagnetic to the
ordered state. The third part seeks to correct the oversimplification which
occurs in all the previous work of assuming the carriers are free, and shows
that neither impurity-trapping nor self-trapping of carriers will affect the
results qualitatively.
The argument of the f i s t part is very simple and worth summarizing.
The author considers a layer structure, in which each layer has a fixed
angular deviation 8 from the previous layer and the succeeding layer, to be
adequately general. Then in the case that b >> kT the double exchange
energy is
Ed = -NxZb cos ( 8 / 2 ) . (VII.4)
Here x is the carrier concentration and 2 the number of bonds per atom
be tween successive layers. There is also, naturally, a superexchange
energy between layers given by
E, = S ~e
N ~ J cos
(VII.5)
= ~ COS* (e/2) - 1).
N Z J S (2
We can see immediately that cos 8/2 = 0 (8 = 180") never gives a
minimum of Ed + Em, so that the pure antiparallel behavior does not
occur if b is finite. The actual process of minimization gives

(VII.6)

if E 2 1. If E is greater than 1, B = 0, which corresponds to a ferromagnetic


state.
De Gennes points out that a fixed angle 80 between layers does not
determine the state. Equation VII.6 can be satisfied with either a helical
arrangement, a two-sublattice canted arrangement, or even a purely
random collection of vectors, each having an angle 80 with the previous one.
In general, however, we can expect anisotropy forces to stabilize the two-
sublattice canted arrangement.
It is important to note that this case is quite the opposite from that of
Yoshi1nori-Kaplan8~+4which leads to a helical arrangement coming from
next-nearest neighbor forces. In this case the canting effect is the result of
THEORY OF EXCHANQE IN INSULATORS 213

the unusual dependence of (VII.4) on the angle, and not of multineighbor


forces. There is no reason to expect helical arrangements to be more
stable than canted ones in the present case.
As de Gennes remarks, a number of experimental data are well ex-
plained by two-sublattice canting :
1. Lack of saturation in high fields. Clearly the canting can be increased
indefinitely by high fields. (Experiments by Wollan and K ~ e h l e r . ~ ~ ~ )
2. Neutron intensities: both ferromagnetic and antiferromagnetic
neutron lines are predicted and 0bserved.m The dependence of the ratio of
intensities on x, the concentration, predicted through (V11.6) is observed
experimentally. This shows that the foregoing theory is valid in the range
0 < x < 0.25.
The second part of de Gennes’ theory is too complex to duplicate here.
The essential result is that as kT increases and as the direction of the
individual spins begins to fluctuate, the mean interaction energies E d a
(cos 0/2) and E, a (cos 0 ) decrease at different rates. Hence the two
components of the effective local field change differently with temperature.
The result always turns out to be such as to decrease the angle of canting
and to lead eventually to ferro- or antiferromagnetic ordering at higher
temperatures. Figure 12 (from de Gennes) shows a typical phase diagram
giving the transition temperatures and ordering arrangements vs. the
concentration x, as predicted by de Gennes’ theory. The transition points
T I, T,,and TN are complicated functions of the various parameters, but
it is interesting to note that the paramagnetic Curie point in the Curie-
Weiss law
x = C/(T - TP) (VII.7)
obeys the standard relationship for molecular field theory, namely
T p = -TN or +T,, (VII.8)
so long as b >> k T , which is satisfied in all cases of interest.
The essential justification given in the third part of deGennes’ paper
for treating bound or self-trapped carriers as if they were free is as follows.
The effect of a bound carrier is to produce the canting and other influences
we have been describing in its immediate neighborhood. Since the anti-
ferromagnetic superexchange is present throughout the lattice, however,
the distortion or rotation of the magnetization is propagated in a very
long-range manner, in fact as 1/R2, as a result of the process discussed
by Suh1206 in an analysis of the long-range interactions of nuclear spins
206 E.0.Wollan and W. C. Koehler, Phys. Rev. 100, 545 (1955).
ao) H. Suhl, Phy.9. Rev. 109, 606 (1958).
214 PHILIP W. ANDERSON

b b
X-

FIG.12. Phase diagram for a double-exchange magnetic material, showing stability


regions of canted, ferromagnetic, and antiferromagnetic states (after de GennesZo1).

caused by polarization in ferro- and antiferromagnets. Such long-range


interactions tend to couple the various carriers together and produce much
the same effect as a uniform distribution of free carriers.
In summary, de Gennes’ remarks, although applied by him only to the
( La-Ca) MnOa system and, in passing, to (Cr-Mn) Sb, probably apply
to a wide range of systems having quasi-free carriers with bandwidths
that are not large compared to intra-atomic exchange integrals. Numerical
calculation of b and the superexchange integrals would not differ in any
important respect from those given in Part VI. Only the use made of the
parameter b is different. The basic point which is emphasized and justified
is that the dependence of the interaction on cos e/2 rather than cos 0 repre-
sents a real departure from the Heisenberg Hamiltonian which has quite
observable effects.

ACKNOWLEDGMENT0

In addition to the very generous help which J. S. Smart gave me with Part V, I
should like to acknowledge a large number of useful suggestions and conversations with
R. G. Shulman, S. Sugano, and T. Moriya, m well m permission to use their unpublished
work. C. Herring has looked over many parta of the manuscript and made useful sugges-
tions. Conversatiom with V. Jaccarino and F. Keffer have also been helpful.

You might also like