Physics Vol. 2, No. 6, pp. 263-272, 1966. Physics Publishing Co. Printed in Great Britain.

S C A L I N G L A W S F O R I S I N G M O D E L S N E A R T e*

L E O P. KADANOFF*

Department o f P hysics, U n iversity o f I llin o is

Urbana, I llin o is
(Received 3 February 1966)

Abstract

A model for describing the behavior o f Ising models very near Tc is introduced.
The description is based upon dividing the Ising model into cells which are micro-
scopically large but much smaller than the coherence length and then using the
total magnetization within each cell as a collective variable. The resulting calcu­
lation serves as a partial justification for Widom's conjecture about the homogene­
ity of the free energy and at the same time gives his result sv' = y' + 2£.

1. Introduction

IN a recent paper [l] Widom has discussed the consequences of the assumption that the free
energy in a system near a phase transition of second order is a homogeneous function of para-
meters which describe the deviation from the critical point and has shown that this assumption
leads to consequences which roughly agree with our present numerical information [2] about the
behavior of such systems. Another paper by Widom [3] written at the same time explores the con-
sequences of an apparently quite independent idea: that the behavior of the interface separat-
ing droplets of the "wrong phase" within a system just below a phase transition should be quite
similar to the behavior of an interface separating a region of fluctuation in the order para-
meter from the surrounding medium [4], Here again information is derived which agrees with
numerical calculations and experiment [2].

Widom's ideas about interfaces are based upon physical plausibility arguments; his idea
about the homogeneity of the singular part of the free energy is not given any very strong
justification beyond the fact that it appears to work. In the present paper, the Ising model
is analyzed in a manner which is designed to throw light upon how correlations between the
order parameter in different regions of the lattice scale when the parameters describing the
deviation from the critical point - in this case T - Tc
and the applied magnetic field - are
changed. Widom's homogeneity condition upon the singular part of the free energy and some of
his results for interfaces are then derived as a consequence of these scaling arguments based
upon our model.

* This research was supported in part by the National Science Foundation under grant
NSF-GP 4937.

t A.P. Sloan Foundation Fellow.

263
264 ISING MODELS NEAR T c Vol.2, No. 6

Although the argument is carried out in Ising model language, it is clear that the arguments
could be generalized to other cases of second order phase transitions.

2. Description of Model

Consider an Ising Model in a weak magnetic field near Tc. This model can be described in
terms of two parameters

( 1)
which measures the deviation from the critical temperature, and h, a dimensionless magnetic
field defined so that flipping a single spin gives a change in magnetic energy 2h/kT. A full
solution of the Ising model would be obtained if we knew f(s, h)t the free energy per site in
the presence of the magnetic field.

To get some feeling for the behavior of the Ising model, imagine that we divided the entire
lattice into cells of L lattice sites on a side. Then each cell in this s-dimensional lattice
contains L s lattice sites. We take L to be large, but much smaller than the coherence length,
§, which describes the range of spin correlations, measured in lattice constants. Since £ goes
to infinity at e = 0 and h = 0, it is easy to find an L which satisfies this criterion.

To zeroth order, each cell can be considered to be isolated from the others and fromthe
external magnetic field. Then to zeroth order /(e, h) = fi(e) where fi(e) is a free energy per
site of a lattice of side L in no magnetic field. Since the small size of the cell tends to
eliminate the singularities from the phase transition, fi(e) should be an analytic function of
e but not of L, i.e.

( 2)
where fi°, f'i and fi2 should not all be analytic functions of L [4].

Next consider interactions of the cell a with the magnetic field. This gives a term in
of the form

(3)

The basic assumption of our analysis is that within each cell, the spins tend to line up so
that they mostly point either up or down. That is,

(4)

where Pq is either plus one or minus one. The average spin, < ct>£, defined in (4) is given by

(5)

Because of the "small" size of the cell, this depends strongly on only L, not upon e or h. Thus,
the interaction with the magnetic field takes the form of a term in exp of

( 6)
Vol.2, No. 6 ISING MODELS NEAR T c 265

Next, consider the interaction among cells. The free energy will tend to be larger if the
spins on neighboring cells are lined up. There will tend to be a smaller contribution if they
are anti-parallel. Then, in net, this tends to make a contribution to the exponential exp[-|3f]

(7)

A/
Here the sum extends over nearest neighbor cells, if + / int gives the contribution to the free
energy when neighboring cells are aligned and, - K + / int gives the contribution to the free
energy when they are out of step. Sipce the direct interactions between cells which produce
/ int occur within a distance which is very short compared to the coherence length, we assume
that / int is, like fi(e), a regular function of e, but not necessarily a regular function of L.
/s/
On the other hand, K is perhaps a somewhat more subtle beast. This describes the extra free
energy that it costs to put two cells out of step. This involves, then, the rather delicate
difference between the ways cells can match up when they are in step and when they are out of
step. Nonetheless, it seems reasonable to assume that K(e, L) is also a regular function of e;
but, we assert this with somewhat less confidence than our other statements relating to this
model. In writing (7) we are asserting that the correlations among cells can be totally repre-
sented by these interactions among near neighbors and that there are no less direct interactions
that we need include in (7) as long ranged interactions. This statement, together with the
assertion that the cell can be represented by the double valued variable, pa , are the two basic
assumptions of this model.

To find the correction to the zeroth order result (1), we sum the product of (6) and (7) over
both possible orientations of each |ia . This sum is, of course, just an Ising model calculation
/V A/
with coupling constant K and effective magnetic field h. This then gives a contribution to the
total free energy

( 8)

where s measures the deviation of the new coupling constant from the critical value, and where,
from (6) the effective magnetic field in the cell problem is

(9)

Since the original Ising model problem has within it correlations over many cells, we assert
^ /V

that e and h must be sufficiently small so that the new Ising model problem retains long-ranged
wo ^

correlations in < MaWj3 > - This requires e « 1, h « 1.

Since there are L* sites per cell, equations (1) and (9) give the free energy as

(1 0 )

Equation (10) is one of the two basic relations in our analysis.-The other is obtained by
discussing the spin-spin correlation function < crrcrr ' > for the case in which the distance be-
tween spins |r - r'|, as measured in lattice constants, is much larger than L. Then each crr can
be replaced by the average of or over the cell in which it lies so that (4) may be used to re-
write the correlation function as

(ID
266 ISING MODELS NEAR Tc Vol.2, No.6

The last line of (11) follows because the u’s are described by an Ising model with effective
distance from Tc, 7,
effective magnetic field and cell h L
lattice constants long.

3. Analysis of the Model

The free energy is a singular function of 6. For example if the specific heat diverges as
e-0 [(— e)~° ] above [below] Tc, for small e and h = 0,

(12a)

In three dimensions (2) 0 < a < 1, 0 < a' < 1. Another possible behavior has the specific heat
above and/or below Tc
diverge logarithmically so that a and/or a' are zero but

(12b)

and/or

(12c)

We first attack the case in which the a ’s differ from zero and later consider the possibility
of behaviors (12b) and/or (12c).

Consider first, at h
= 0, the terms in equation (10) which are singular in e and e. If we
equate these singular terms above Tc,
we find

It follows then that

(13)

and

(14a)

(14b)

Notice that if we assert that K is a regular function of e, e, which is K - Kc, cannot have a
discontinuous first derivative with respect to e. Therefore, our assertion about the regularity
Vol.2, No.6 ISING MODELS NEAR Tc 267

of K implies
(15)

which in turn implies a = a'. It should be recognized that equation (15) is more dubious than
the other assertions of this paper. Consequently, we hold this statement aside and do not use
it in simplifying our further analysis.

More generally, when h f 0, we can equate the singular parts in e and h on both sides of
equation (10) and find

(16)

This can only be true if h is proportional to some power of L, i.e. if

(17)

and then (16) and (14) can hold true if the singular part of /(a, h) obeys

(18)

This is Widom’s homogeneity assertion [l] which we have now derived from our model.

If a = 0, and / has the form (12b), then the last term on the right-hand side of equation
(10) has a singular part

(19)

if we again employ (13) for this case. The extra term in log L J/y must be cancelled out by some
other term on the right-hand side of (10). But since this term is regular in e, the compensat-
ing term can be obtained by taking /j/2) in equation (2) to be

( 20)

or by a similar structure in sL~s/ int(e, L). When a' > 0, this term in (20) does not contribute
to the leading singularity in /(e, h)
for T < Tc
and consequently the analysis which led to
(18) can go through just as before except that a term in e2 log |e| must be added to F
and F'.

But, when a' is also zero, the same term, (20), must cancel out log L
terms both above and
below Tc.
The coefficient in front of this term cannot change at Tc
since fi
is, by hypothesis,
regular in e. Then when a = a' =0, and hence v = v' by (14), we conclude with Widom that

(21 )

It is relevant to notice that this equality holds in the Onsager solution of the two dimensional
268 ISING MODELS NEAR Tc Vol.2, No.6

Ising model [4]. It is interesting, but perhaps quite besides the point to notice that it is
also true for the A-transition [5] of He 4 and in at least one anti-ferromagnetic transition [6],

The arguments of Widom are designed to use the homogeneity of the singular part of /(e, h),
as exhibited in equation (18), to derive relations between the average magnetization,

( 22 )

the spin susceptibility

(23)

and the parameters, a, a', v and v' which we have already defined. If we define (3, y, y' and 5
in the conventional manner [l, 2]

(24)

it follows from (18) that

(25a)

(25b)

(25c)

Since the parameters in (25) are all experimental quantities, these relations could be checked
if we could find an Ising model in nature.

In the course of this analysis, we also find that vp can be written in terms of experimental
quantities as

(26)

This result is useful in the analysis in the spin-spin correlation function of equation (11),
which equation can now be written as

(27)

for T > Tc. The right-hand side of (27) will only be independent of L if g is a homogeneous
Vol.2, No. 6 ISING MODELS NEAR Tc 269

function of its arguments of the form

(28)

Consequently, the coherence length that we discussed at the beginning of this paper must be at
h = 0

(29)

and at e = 0

(30)

The result (29) makes contact with the notation of Widom [3] and Fisher [2] and permits us to
identify v and v' as experimental quantities. Therefore, equations (14a) and (14b) may be
viewed as experimental relations.

There is one more experimental relation to be obtained. From (28) as e and h go to zero, g
is proportional to |r - r'|"2H\ This relation is conventionally [2] written as

(31)

so that, from (26), we can write

(32)

which is our final experimental relation.

4. Discussion and Comparison with Other Theoretical Results

One can view the preceding work as developing relations among the nine "experimental" quanti-
ties a, a', |3, y, y', 6, v, v' andr\.Equations (14a), (14b), (25a), (25b), (25c) and (31) are
the six such relations that we consider to be direct consequences of our model. The relation
(15), v = v', which implies a = a', y - y' is a somewhat more tenuous conclusion. With (15) in-
cluded, we have seven relations among our nine parameters; if (15) is rejected there are only
six such relations. In the first case, then, there are three free parameters; in the second two.
270 ISING MODELS NEAR Tc Vol.2, No.6 |

s
None of these relations are new. Essam and Fisher [7] argued for (25b) on the basis of I
homogeneity considerations not totally different from those of Widom Cl]
who made full use of ]
the homogeneity conjecture to get (25a), (25b), (25c) and (31) as well as (15). Besides, (31)
can be viewed as a tautology: a definition of the coherence length. In this sense, the relation

which follows from (31), (25c), and (14) can also be viewed as trivial.

There exists an alternative derivation of equation (14a) due to Pippard [8] and Ginsberg [9],
Imagine that we were in a situation of zero magnetic field in which the temperature lay just
above Tc, i.e. e > 0. Then imagine that a temperature fluctuation occurred which took a region
of side X of the material into the ordered state. This would cost a free energy of the order
of the volume of the region times the difference in free energy density between that at T
and
that at Tc, i.e.

(33)

But, the probability that such a fluctuation will occur is proportional to the exponential of
the cost in free energy, divided by kTc.Therefore, the maximum possible value of X, which is
the coherence length, is given by setting (33) to be of order kTcor

(34)

Therefore, we find at once v = (2 - oc)/s.

Despite this simple derivation equations (14) are not unassailable. One could argue that,
as s - ®, we know that the molecular field approximation is valid. This gives v = 1/2, a = 0,
and (14) fails. But, it may well be that as s - ®, the molecular field approximation becomes
valid closer and closer to Tc\ and for all finite s there remains a region within a very small
neighborhood of Tc for which the molecular field approximation fails. Then (14) can still be
true for all s, in the strict limit as g - 0.

Precisely this effect is expected to occur in the superconductor [8-10], Here, the molecular
field approximation (Landau-Ginzberg theory) is known to be valid until you get very close to
T c. This theory introduces a coherence length

(35)

where § 0 is a very large number ~ 104 lattice constants for a pure superconductor. Over most
of the temperature range §l .g . is larger than the coherence length defined by (34). Then the
molecular field theory remains valid because fluctuations in the order parameter are not im-
portant. But very close to Tc, the coherence length of (34) will surpass that of (35), and then
the molecular field theory will fail. In this tiny temperature region about Tc, equations (14)
are expected to be true if the analysis of reference [lO] is correct.

To check the seven relations among physical parameters we compare with known results for
two and three dimensional Ising models. A variety of numerical results have been obtained in
both two and three dimensions through evaluations of the singularities of power series. These
are reviewed in references [2,11-13]. In the two dimensional case there are also a variety of
analytical results available [14-16] all based upon Onsager's solution [4]. For the two dimen-
sional case, equations (28) would imply that the spin-spin correlation function is of the form
Vol.2, No. 6 ISING MODELS NEAR Tc 271

when the magnetic field vanishes and |r - r'| » 1. This result agrees with the conclusions of
reference [16].

The second column of Table 1 lists the known values of the nine experimental parameters for
the two dimensional Ising model. These all check exactly with Widom’s relations.

TABLE 1

Two dimensional case Three dimensional case

Parameter Previously calculated Reference Fit to data

Value Reference value

a 0 4 0.0 < a < 0.2 2 0.085

a' 0 4 0.0 < a' < 0.06 2 0.085

P 1/8 13 0.303 < p < 0.318 2 0.332

y 7/4 2, 11, 16 1.250 ± 0.001 2 1.250

9
Y 7/4 2, 16 1.23 < y' < 1.32 2 1.250

6 15 13 5.2 ± 0.15 13 4.78

V 1 2 0.644 ± 0.003 2 0.638

v' 1 16 2 0.638

*1 1/4 14 0.060 ± 0.007 2 0.045

The first column listed under the three dimensional case gives information which has been
obtained via Pade and other numerical methods. The last column, labeled ’Pit to data*, gives a
set of values for these parameters which agrees with all seven of our relations including the
more doubtful statement v = v \ This fit agrees with all the known numerical results to within
about two and one-half standard deviations. These results indicate at least that the conclu-
sions drawn from our model are not grossly inaccurate.

Acknowledgment

I would like to thank Dr. P.C. Martin and Dr. Brian Josephson for very helpful comments.
272 ISING MODELS NEAR Tc Vol.2, No.6

References

1. B. WIDOM, J. Chen. Phys. 43. 3898 (1965).

2. M. FISHER, J. Math. Phys. 5, 944 (1964).

3. B. WIDOM, J. Chets. Phys. 43, 3892 (1965).

4. For the two dimensional Ising model, the Onsager solution (L. ONSAGER, Phys. Rev. 65, 117,
1944) bears out equation (2).

5. M.J. BUCKINGHAM and W.M. FAIRBANK, in Progress in Lots Temperature Physics III, (ed. C.J.
Gorter), Ch. 3. North Holland, Amsterdam (1961).

6. J. SKALGO, Jr. and S.A. FRIEDBERG, Phys. Rev. Letters 13, 133 (1964).

7. J.W. ESSAM and M.E. FISHER, J. Chen. Phys. 39, 842 (1963).

8. A.B. PIPPARD, Proc. Roy. Soc. Lond. A216, 547 (1953).

9. D.M. GINSBERG and J.S. SHIER, to be published.

10. E.G. BATYEV, A.Z. PATASHINSKII and V.L. POKROVSKII, Zh. Eksp. Teor. Fix. 46. 2093 (1964).
English translation in Sov. Phys.-JETP 19, 1412 (1964).

11. C. DOMB, Advanc. Phys. 9, 34. 35 (1960).

12. M.E. FISHER, J. Math. Phys. 4, 278 (1963).

13. M.E. FISHER, in Proceedings of the International Conference on Phenomena Near Phase
Transitions, to be published.

14. B. KAUFMAN, Phys. Rev. 76, 1232 (1949); B. KAUFMAN and L. ONSAGER, Phys. Rev. 76, 1244
(1949).

15. G.F. NEWELL and E.W. MONTROLL, Rev. Mod. Phys. 25, 353 (1953).

16. L. KADANOFF, to be published in Nuovo Cimento.