ANNALS OE PHYSICS 91, 226-274 (1975)

Soluble Renormalization Groups and Scaling Fields

for Low-Dimensional lsing Systems

DAVID R. NELSON

MICHAEL E. FISHER
Baker Laboratory and Materials Science Center,
Cornell University, Ithaca, New York 14850
Received July 18, 1974

A variety of one-dimensional Ising spin systems, including staggered and parallel

magnetic fields, alternating and second neighbor interactions, four-spin coupling, etc.,
are discussed in terms of renormalization group theory. A continuous range of distinct
renormalization groups is constructed in exact closed form, analyzed in detail, and
compared with exactly calculated thermodynamic properties. Fixed point linearization
yields relevant, irrelevant, and marginal operators. All groups yield identical “critical”
behavior (at T = 0) with 7 = 1, 6 = co, y = Y = 2 - 01, and with identical linear
scaling fields. A generalization of Wegner’s analysis to discrete groups yields explicit
power series for the nonlinear scaling fields; these are seen to depend on the particular
renormalization group and, hence, are physically nonunique. A planar, multiconnected
“truncated tetrahedron” model of effective dimensionality log, 3 is analyzed via a dede-
coration and star-triangle group revealing highly singular behavior as T + 7, = 0.

1. INTRODUCTION

The general formulation of Wilson’s renormalization group approach to the

study of critical point behavior has been presented by Wilson [l] and others [2]
and developed by Wegner [3] to discuss the corrections to asymptotic scaling [3].
The principal model so far discussed with an exactly realizable renormalization
group, in contrast to a perturbation theoretic realization by E-expansions, etc.
[410] is Baker’s model [ll] which has the same structure as Dyson’s hierarchical
model [12, 131. For this class of models, however, no exact solutions of the
renormalization group equations have been found, nor have the models themselves
been solved exactly by any other methods.
In this paper [14] we discuss a variety of models for which renormalization
groups can be constructed in exact closed form as recursion relations in finite-
dimensional parameter spaces. The simplest model is the nearest neighbor linear
226
Copyright 0 1975 by Academic Press, Inc.
All rights of reproduction in any form reserved.
 LOW DIMENSIONAL ISING SYSTEMS 221

Ising chain [15, 161; we also consider more elaborate one-dimensional Ising models.
Although the thermodynamics of these systems are readily calculable by traditional
methods (such as the matrix method), and their “critical points” all occur at zero
temperature [16] the explicit analysis of the associated renormalization groups
illuminates a number of important features of the general theory. In particular
we discuss (a) the nonuniqueness of the renormalization group; (b) its semigroup
property: (c) the behavior of the constant or “background” term in the Hamilto-
nian; (d) the existence and nonexistance of fixed points for particular types of
critical behavior; (e) the exact critical operators, including examples of relevant,
marginal, irrelevant, and pseudo-relevant variables; (f) asymptotic scaling;
(g) the “irrelevancy” and “nonlinear” corrections to asymptotic scaling; (h) the
nonlinear scaling fields, which are shown to be nonunique, and techniques for
calculating them.
In addition we discuss one model, a planar, multiconnected “truncated tetra-
hedron” Ising lattice, (which is not of simple one-dimensional form) where the
exact renormalization group equations yield precise information on the critical
(T-+ 0) behavior which seems otherwise inaccessible. Although this last model
represents the only new result concerning critical behavior as such, we believe the
study throws light on various important aspects of the renormalization group
approach and improves ones general understanding of its operation.
For the reader’s convenience we summarize at this point, the main aspects of
the paper. In Section 2 we review the renormalization group approach with
emphasis on certain general features and questions. Readers familiar with the
theory may wish to peruse this section lightly or merely refer back to it for nota-
tion, etc. as the occasion arises.
In Section 3 the simple linear Ising chain with nearest neighbor interactions is
discussed in the presence of both a uniform magnetic field, H, and a staggered
magnetic field, Ht. A renormalization group transformation is defined by using
the spin dedecoration (or “iteration”) transformation [17, 181. This yields explicit
algebraic recursion relations. Various fixed points exist, including one describing
the ferromagnetic critical point (at T = 0 with pair spin coupling J > 0) and the
antiferromagnetic critical point (at T = 0 with J < 0). The appropriate critical
exponents are correctly found by linearization of the renormalization group.
Specifically, one obtains 7 = 1, 6 = co, and y = v = 2 - d. (Compare with
Ref. [16].) Both relevant and irrelevant critical variables or operators appear. In
addition, when alternating coupling strengths J1 and J2 along the length of the
chain are introduced (Section 4), a marginal variable appears. More elaborate,
but still exactly soluble models discussed in Section 4 include the linear chain
with both first and second nearest neighbor couplings and a “ladder” of spins
including a four-spin interaction.
By considering different spatial resealing factors b, produced by dedecorating
228 NELSON AND FISHER

firstly alternate spins and secondly, a pair out of every three spins, etc., the semi-
group property of the renormalization transformation is explicitly checked. The
importance of the constant or spin-independent term in the Hamiltonian appears
in the calculations; it is seen to go to an appropriate limit at the fixed points. The
scaling predictions of the renormalization group can be checked against the exact
solutions and the appropriate scaling functions may be derived.
In Section 5 we describe two “truncated tetrahedron” Ising models. These are
respectively three- and four-coordinated planar structures containing polygons
of sizes 3, 6, 12, 24 ,... . However, despite the number of polygonal closures, these
lattices (or pseudo lattices) have a connection number [19] of only three; as a
result the critical point still occurs at T = 0. Nevertheless, the singularity there is
extremely strong; its character (in zero field) can be deduced from the exact
renormalization group which is constructed. We have not, otherwise, been able
to solve these models.
Since the exact renormalization group equations for the linear models are at
hand, Wegner’s analysis [3] of the corrections to scaling can be applied. As
demonstrated in Section 6, this enables us to generate explicitly, appropriate
nonlinear scaling fields (combinations of temperature T, magnetic field H, etc.).
However, these are seen to be nonunique: the free energy can also be extended
outside the linear critical region in scaling form using other nonlinear fields.
Wegner’s discussion [3] assumes that the renormalization group is given as a
continuous group on the spatial resealing factor. In our analysis the resealing
factors are always integral and the renormalization group is discrete. Accordingly
we present a method for calculating the scaling fields for a discrete group.
We also show, in Section 3, that it is possible to “miss” a fixed point describing
a particular critical point if care is not taken. Specifically, the antiferromagnetic
fixed point may be transformed away and an irrelevant operator is apparently
turned into a relevant one. This illustrates that a useful renormalization group
should be chosen so as to “focus” on the critical point of interest.
In a related context, we generalize in Section 7, an Ising spin renormalization
group devised by Wilson [20] thereby obtaining an infinity of renormalization
groups for the Ising chain parametrized by one continuous parameter. These
groups (which in one limit include the original dedecoration group) generate the
same low temperature fixed point with the same critical exponents, but have
different nonlinear scaling fields and different high temperature fixed points. The
distinct character of these renormalization groups is specifically related to the
presence of a variable spin-resealing factor. In general such a feature is needed
in order to realize a fixed point with a particular value of the exponent q. However,
for dimensionality d = 1 and a critical point with 71= 1, a “fixed spin” renor-
malization group, such as the simple dedecoration group, suffices as the earlier
analysis shows.
 LOW DIMENSIONAL ISING SYSTEMS 229

2. RENORMALIZATION GROUP FORMALISM

In this section we review briefly the general renormalization group formalism

[l-3]. This will serve to exhibit the features to be explored in the explicit analysis
and to introduce the notation.

2.1. Renormalization Groups

A system with given interactions at a specified temperature, T, and subject to
particular external fields H, H+,... is described by its reduced Hamiltonian

c@ = L@({s}~) = -sI”({s& , H, H+,... )/k,T, (2-l)

where ($1, = {sl , s2,..., St,..., s,,,} denotes a set of N local degrees of freedom
associated with, say, lattice sites Ri . In our discussionthese degreesof freedom,
or jield cariables, will be simple Ising spins, si = i 1. The reduced Hamiltonian,
or what is equivalent, the Boltzmann factor es, is to be a translationally invariant
function of the field variables (at least asymptotically in the thermodynamic
limit N-+ co); it will be parametrized by a set of “initial” fields or interaction
parameters {k} = {k, , k, ,...} describing the various (translationally invariant)
terms appearing in 2. It is normally convenient to take kf as the coefficient
multiplying some particular additive term in Y? but for critical points occurring
at T = 0 other choices will often be more convenient. The zeroth field, k, , may
be identified with the constant or “spin-independent” term in .?f’, say E,/k,T.
In general the set {k} must be of infinite dimensionality. The solubility of the
renormalization groups we describe residesin the fact that they may be realized
exactly in a finite field space{k, , k, ,..., k,} = K, .
A renormalization group transformation R carries a Hamiltonian LZ?(we will
omit the adjective “reduced” hereafter) into a transformed or renormalized
Hamiltonian
2’ = R[sP], (2.2)

which is again translationally invariant and may, likewise, be parametrized by

the set {k}. On a finite system the renormalization operator [w acts to reduce the
number of degreesof freedom from N to

N’ = N/bd, (2.3)

where d is the spatial dimensionality while the spatial resealingfactor b exceeds

unity. In the examples we will discuss,b will be an integer but this is not generally
necessary.
230 NELSON AND FISHER

The renormalization group must also (under all currently explored formula-
tions [Zl, 221) keep the partition function

Z,,,[i] = Tr,{e%} (2.4)

invariant so that
ZN’[TP’] = Z&P]. (2.5)

The operation Tr, denotes the trace operation, integration, summation, etc.,
appropriate to the field variables {s}~ . The free energy per degree of freedom

f[SP] = Ii% N--l In Z&F], (2.6)

thus satisfies the basic covariance relation

(2.7)

This relation ultimately leads to critical point homogeneity and scaling. The
presence of the thermodynamic limit in (2.6) and (2.7) is essential to any discussion
of critical points (which do not exist for finite N) but, in general, leads to severe
technical problems in constructing a mathematically rigorous definition of R.
In a hear renormalization group the basic correlation function

G[Ri - Rj ;rio] = (sisj)~, WO

(where (e) denotes the standard statistical average taken with 9 in the thermo-
dynamic limit) transforms according to

G[R;i@] = c2G[R/b;e@'], (2.9)

where c = c[g] is a spin resealing factor determined by R. However, nonlinear
groups have been successfully employed numerically [23], and have recently
been studied theoretically [24]. We will also exhibit an exact nonlinear renormaliza-
tion group (in Section 7).
Under iteration

FP[sq = lQ[W[S]],..., IwZ[sq = Dqrwyq], (2.10)

one obtains renormalization operators with spatial resealing factors b, = b2,..., bl =

bz which satisfy the semigroup property

Rz+z’ = pRz’
3 b z+z’ = bzbz, = bZ+Z’. (2.11)
 LOW DIMENSIONAL ISING SYSTEMS 231

2.2. Fixed Points and the Critical Spectrum

To utilize a renormalization group one looks for a fixed point S?* defined by
[wpr*] = &?* (2.12)

and presumes that one may expand as

lR[$'* + wQ] = A?* + wLQ + O(w2), (2.13)

where w is a scalar parameter and L = (SE%/&%?)* is a linear operator on

Hamiltonians. One then studies the eigenvalue problem

LQj = 4Qj, (2.14)

to obtain the spectrum of critical operators or variables Qi , and corresponding

eigenvalues dj , which in view of the semigroup property (2.11) are expected to
have the form
A,(b) = bhj, (2.15)

where the hj are independent of b.

The form of dependence of Aj(b) on b also follows if R can be defined for b + 1
so that the infinitesimal generator
G = J$ (R, - U)/(b - 1) (2.16)

can be constructed. If we redefine I = In b the renormalization group Eq. (2.2)

then becomes
ds/dl = S[Y?]. (2.17)

This is the form presumed by Wegner [3] in his analysis of corrections to scaling,
etc., but we will be working with discrete groups only.
If the set of eigenoperators or critical variables Qj for the fixed point X* is
complete one may write

2 = c%f* + C h,Qj , (2.18)

thereby parametrizing SP in terms of eigenfields or (linear) critical BeIds

VG = {ho , 4 ,...I.
In the cases we will study the completeness of the scaling field expansion (2.18)
[or its equivalent] will not be in doubt; but in general one should probably expect
no more than some sort of asymptotic or weak completeness, e.g., for suitable
expectation values taken with 9 in the vicinity of &‘*. The mapping from the
232 NELSON AND FISHER

initial fields (k) to the scaling fields (h) is in general nontrivial but can often be
simplified (at least to leading order) by symmetry considerations.
One of the critical operators, say QO, may be identified as the constant or spin-
independent term with field k, or h, . Since such a term in z+? can always be
removed from under the trace in (2.4) it follows that it transforms in a trivial way
under the renormalization group so that its eigenvalue is always (1, = bd or
h, = d.
Under action of the renormalization group the expansion (2.18) yields

R[s] = 2’ = &‘* + c h,z’&Qi + 0(/Q (2.19)

which may be recast as the recursion relations

hj’ = Ll,h,[l + O(h, ) h, ,...)I. (2.20)

Iteration I times yields

h?’
3 r=z (lizhi , (2.21)

provided one stays within the linear region where the O(h2) terms in (2.19) can
be neglected. It now clearly makes sense to classify the critical variables as
(a) relevant with (lj > 1 (Xj > 0), which grow in importance, (b) irrelevant with
/1, < 1 (X, < 0) whose contribution diminishes under iteration, and (c) marginal
with (li = 1 (Ai = 0) which remain (in linear order) of constant magnitude.
Normally the most relevant operator (with largest CrJ would identify the order
parameter and its ordering field h = H/kBT. Another relevant operator would
be the “energy” and its field would be the reduced temperature t = (T - TC)/TC ,
etc. However, in the cases we will study, where T, = 0, the corresponding tempera-
ture variable must evidently be expressed in different form.
Furthermore, in general, there are a variety of distinct fixed points with distinct
scaling fields and exponents. In general different identifications will be necessary
at different fixed points. In particular we will observe infinite temperature and
infinite field fixed points where, clearly, appropriate new variables must be adopted.

2.3. Scaling and Nonlinear Scaling Fields

On using (2.21) in combination with (2.15) and the basic covariance relation (2.7)
for the free energy, one obtains, for the linear region,

f(h,, h, ,..., hj ,...>M b-zdf(bzA”h, , bzA’h, ,..., b”jhj ,... ), (2.22)

wheref[X] has been expressed as a function of the scaling fields. Since bz may be
indefinitely large one can choose 1 so that as a relevant field, say h, = t, becomes
 LOW DIMENSIONAL ISING SYSTEMS 233

small one has bz”lt = 1 whereupon this relation can be written in the standard
scaling form
f(ho ) t )..., hj )...) R3 t2-mY(ho/tbfJ)...) hj/P’ )... ), (2.23)
where
2 - a = d/X, , (2.24)

the crossover exponents are given by

(2.25)

and the scaling function is

For irrelevant scaling fields hk the crossover exponent C& is negative so that
hJt*y = hrtldjl --f 0 as t -+ 0 and one expects [21] to be able to set the corre-
sponding argument yk in (2.26) to zero and ignore the hk dependence off in the
critical region. Expansion with respect to y, should yield systematic “irrelevancy”
corrections to asymptotic scaling laws [3]. As we will show explicitly in Section 3.5,
however, “nonlinear” corrections arising from the O(h2) terms in (2.19) may well
be more important.
For a linear renormalization group, where (2.9) applies, one obtains a functional
equation for G(R) at Z-P = $‘P* with solution

G*(R) w D/R2W as R+co, (2.27)

where
c* = c[&‘*] = b-*. (2.28)

This typically critical-point decay law identifies the critical exponent 7 via

d - 2 + 71= 2w = --In c*/ln b. (2.29)

This identification shows that in using a linear renormalization group one must
usually treat c* as a sort of eigenvalue which is to be adjusted to obtain a non-
trivial fixed point describing the critical behavior of interest. In combination
with (2.19) and (2.21) one obtains the correlation scaling relation

G(R; h, , h, ,...) m R-‘d-2+“D(Rt” ,..., hj/t”J...), (2.30)

with v = l/X, , which through (2.24) implies the hyperscaling relation [22]

dv = 2 - 01. (2.31)
234 NELSON AND FISHER

Finally, the nonlinear scalingjelds

g,v4l, 4 ,...I = Ml + w, , 12,,...>I (2.32)

are introduced [3] as the exact (formal) solutions of the fuZZ, nonlinear recursion
relations (2.20) which behave as

(2.33)

where there are now no correction terms. If 2 is parametrized in terms of these

nonlinear scaling fields (assuming this to be possible) we obtain the exact non-
linear homogeneity relation

f(& 3 g1 7*..3 gj )...) = Pf(b""g, ) eg, )..., P+g, )... ), (2.34)

in contrast to the asymptotic homogeneity (2.22).

Wegner [3] has shown how to generate the nonlinear scaling fields as power
series, given differential recursion relations corresponding to (2.17). In certain
cases logarithms also appear. We will show how the same can be accomplished
for a discrete group. As mentioned, we will also show explicitly that whereas
distinct renormalization groups describe a particular critical point in terms of
the same linear scaling fields, the corresponding nonlinear fields are in general
different, and hence, nonunique.

3. DEDECORATION RENORMALIZATION GROUPS

3.1. Decoration Transformations

Decoration or iteration transformations of Ising models have been treated
generally by Fisher 1181. Using such a transformation, one can replace a central
spin (or any other physical system) coupled to two neighboring spins, s, and s2 ,
by a single bond joining the two “external” spins. The transformation is effected
by first taking a trace over the internal degrees of freedom of the physical system
coupled to the two “external” spins. One is then left with a conditional partition
function I,!I(s~, s2 ; T) depending on the two external spin variables. The assembly
is then replaced (see Fig. 1) by a single Ising bond of strength J = k,TK coupling

8L, K 8L,
0
- Ns,s2) y2* s”,
St SP

FIG. I. The generalized decoration transformation.

 LOW DIMENSIONAL ISING SYSTEMS 235

s1 and s2 , by augmented (reduced) magnetic fields 6H, = kBT 8L, and

6H, = k,T 6L, acting on s1 and sZ , and by a spin-independent function g(T)
which contributes an additive term -kT Ing to the total free energy. The trans-
formation equations, as derived by Fisher [18] are

e4K = *++#--I$+-#-+, (3.1)

e46L2 ~
e46L1= *++#+-/#--#p+, ~ *++*-+I*--*+- 3 (3.2)
and
g* = *++L#+-$-+ > (3.3)
where #,.+ = #(+ 1, + 1; T), and so on.
We apply this procedure to the nearest neighbor Ising chain of N spins in a
magnetic field [15] which is described by the familiar Hamiltonian,

2 = -J % .s+s~+~- H 5 si - NE,,(J, H). (3.4)

i=l i=l

The function E,(J, H) represents a “zero-spin” or “background” contribution

to the energy which plays a role in the description of the linear Ising chain in
terms of a renormalization group; initially, we may take E,, to be a constant. It is
most convenient to work with the reduced Hamiltonian 2 defined by

c@ = -iF]kBT = K c s~s~+~+ L c s< + NC(K, L), (3.5)

i=l i:=l
where
K = J/k,T, L = H/k,T, C = E,lk,T. (3.6)

By dedecorating every other spin along the chain [15] we now generate an
elementary example of a renormalization group transformation with b = 2. The
effect of this dedecoration transformation on the partition function Z,(C, K, L) is
given by
Z,(C, K, L) = Z,,,(C’, K’, L’), (3.7)
where K’ represents the new (effective) coupling between the remaining spins
while L’ represents the corresponding transformed ordering field. If, temporarily,
we ignore the spin-independent terms C and C’, the dedecoration group simply
maps the point (K, L) specifying the initial Hamiltonian onto a new point (K’, L’)
describing the renormalized Hamiltonian.
In the process just sketched the physical subsystem removed by decoration
was a single spin. Other renormalization groups can be constructed by applying
the dedecoration transformation to more complicated objects. In the following
236 NELSON AND FISHER

analysis we will study the effects of removing two adjacent spins out of every
three spins in a linear chain, and of removing alternate “rungs” in a “ladder”
(or double chain) of spins.

3.2. Recursion Relations for Dedecoration with b = 2

Dedecoration of every other spin in a linear Ising chain generates a renormaliza-
tion group with spatial resealing factor b = 2. The conditional partition function
for decoration is

#(.Q
, s,>= 2 cosUK(s, + ~2)+ Ll. (3.8)

It is convenient to work with the new set of variables,

w = e-4C 9 x = e-4K 3 and y = eezLm (3.9)

In terms of these variables the dedecoration relations (3.1) to (3.3) yield the basic
recursion relations

w’ = w2xy2/(1 + y)” (x + y)(l + xy), (3.10)

x’ = x(1 + Y>“/(X + YXl + XY>, (3.11)
Y’ = Yb + YMl + XV>> (3.12)

and (3.7) likewise applies in terms of w, x, and y.

Evidently the parameters x and y move in the (x, y) plane independently of w;
the spin-independent term w is thus “driven” by the interaction terms. This is a
general feature of the renormalization group. Clearly, fixed points may be obtained
by study of the last two recursion relations alone. Restricting attention to the
unit square (0 < T, H < co), for J > 0 we find a “paramagnetic” line of jixed
points at x* = 1, independent of y, corresponding to T = co (or equivalently,
J = 0). There is also an isolated “fully aligned” or “frozen,” infinite-field fixed
point at (x*, y*) = (0,O). Lastly there is “ferromagnetic” fixed point at (0, 1) as
shown in Fig. 2. Since we are interested mainly in the “critical” behavior of the
system near T = 0, H = 0, where the correlations become long-ranged, we will
study the “ferromagnetic” fixed point, (0, 1). Under iteration of the transformation
an initial Hamiltonian specified by (x0 , y,,) describes a discontinuous trajectory
in the (x, y) plane; typical trajectories are sketched (as if continuous) in Fig. 2.
On linearizing about the ferromagnetic fixed point (x* = 0, y* = 1) with
dy = y - y* = y - 1, the recursion relations become
x’ w 4x, Lly’ w 2dy. (3.13)
 LOW DIMENSIONAL ISING SYSTEMS 237

t u=n-.

0 I I I.1/ I 1-1 I t
0 0.5 1 1
x2 = e-2K
H=m
FIG. 2. Trajectories and fixed points for the dedecoration renormalization group applied
to the linear Ising chain. Isolated fixed points occur at (x*, y*) = (0, 0) and (0, l), while a line
of fixed points appears at x = 1. The dots represent sucessive applications of the b = 2 trans-
formation. Different initial conditions would give different sets of dots, all of which would fall
on the continuous curves sketched here. The trajectories are the same for the b = 3 dedecoration
group: see Section 3.5.

These relations are already diagonal so we find the eigenvalues

A, = 4, A, = 2, (3.14)
A, = 2, A,= 1, (3.15)

so that both critical fields x and dy are relevant.

Formally from (3.9) the ferromagnetic fixed point vaIue of w is w* = co: this
indicates that a change of variable is appropriate. If one multiplies (3.10) and (3.11)
together one obtains

(4’ = (wx)” Y”/(X + VI” (1 + xv)“, (3.16)

which near the fixed point reduces to (wx)’ = (wx)“. The appropriate fixed point
is (wx)* = 1 and the eigenvalue is A,, = 2. It is convenient to interpret this result
by imagining the temperature T is unaffected by the transformation. When one
238 NELSON AND FISHER

is near the fixed point, so that x, and dy, are small, one may take the logarithm
of (3.16) to find
(E + J)’ - 2(E + J), (3.17)

so that A,,, = 2 or X,,, = I = d; this confirms the general observation about

the eigenvalue of the constant term made in Section 2.2. Furthermore if one
chooses E, = -J, the fixed point value, one has E, = -Jl throughout the linear
region. This choice of E corresponds to setting the ground state energy of the
system equal to zero.
It should be noted that the recursion relations (3.10) and (3.11) are invariant
under the replacement y 3 y-l as expected by the symmetry of the initial Hamilto-
nian. (However, we will examine a renormalization group which does not respect
this symmetry.) If, however, one starts with x,, > 1 (corresponding to antiferro-
magnetic coupling J < 0) one finds that x0 is immediately mapped into a value
x1 -=c1. This is because the effective interaction between alternate spins in a
antiferromagnetic chain is, indeed, ferromagnetic. The effective interaction remains
ferromagnetic under further iterations. Thus no specifically antiferromagnetic
fixed point appears. Furthermore, the uniform magnetic field H is expected to
be a thermodynamically irrelevant variable near an antiferromagnetic critical
point (i.e., to have a fixed point eigenvalue (1, less than unity). After one iteration
step, however, the uniform field is transformed close to the ferromagnetic fixed
point (x* = 0, y* = I), where it is apparently thermodynamically relevant. Thus
an irrelevant variable is projected onto a relevant one. The resolution of this
paradox in treating antiferromagnetic coupling is presented in Section 3.7.

3.3. Scaling Relations for b = 2 Dedecoration Croup

The transformation of the reduced free energy

AC, K ~9 -f( W, x, ~1) = ji+i N-l In Z,(C, K, L) (3.18)

of the chain follows from (3.7) as

f(w, 4 v> = i&w’, VT’,v’)* (3.19)

If we choose E, = --J and then neglect the w- or E-dependence (as is valid in the
linear region) we obtain on iteration I times

fk 4) = 2-w% 7 MY,). (3.20)

But from the linearized Eqs. (3.13) one has

Xl = 41x and Lly, = 21 dy. (3.21)

 LOW DIMENSIONAL ISING SYSTEMS 239

We now choose I so that 221x = k < 1 for x close to the critical value x, = 0,
where k is small enough that the linearized recursion relations are valid. On
substituting into (3.20) and using the approximation dy .w 2L valid near the
critical value L, = 0, we obtain the scaling prediction

f(x, y) = f(T, H) R3 e-2KY(#wK), (3.22)

where, as before, L = H/k,T and K = Jjk,T, while the scaling function is given
by
Y(v) = k-lf(k, 2kv). (3.23)

In principle x = e-4K can take only the values k/2-2z for I integral so that k,T/J
takes only values of the form l/(c,l + c,), but since these values become closely
spaced near the critical point this is not a serious drawback. However, further
investigation is needed to check that no “ripple” is, in fact, generated in the
thermodynamic functions. The analysis of Section 6, however, shows how to
construct an analytic continuation in b.
The low temperature properties of the Ising chain follow immediately from
(3.22). Thus, the zero-field energy approaches its “critical” value UC = 0 as e-2x.
By differentiating twice with respect to L the zero-field susceptibility,
x = (aiv/aH), - a(s,yaL, is seen to diverge exponentially as e+2K as
K-l = k,T/J+ 0.
Now the exact free energy of an Ising chain with E, = -J < 0 is well known [25]
to be
f(K, L) = -K + In]@ cash L + (e2K sinh2 L + e-2”)1/2], (3.24)

which near L = 0, K-l = 0 reduces to

f (K, L) = e-2K(1 + L2e4K)1J2 + 0(L*, ec4K). (3.25)

From this we may identify the scaling function (3.23) as

Y(u) = (1 + U2)1/2. (3.26)

The spin-spin correlation function

G(R, T, H) = (s,,sR) = Z-,’ C .so,yRe*(SsI, (3.27)

{si=il)

may be analyzed similarly. (We measure R in units of a lattice spacing.) If one

takes s0 and sR as spins which are not removed by decoration, one finds the exact
recursion relation
G(R, x, dy) = G($R, x’, dy’), (3.28)
240 NELSON AND FISHER

as is intuitively obvious from the nature of the dedecoration transformation.

Evidently the dedecoration renormalization group is “linear” in the sense explained
in Section 2. Furthermore the spin resealing factor, c, is simply constant and equal
to unity. [See Eq. (2.9).] On iteration within the linear regime, where (3.13) apply,
we obtain
WC x, AY) m G(2-zR, 2%, 2z Ay), (3.29)

which leads to the scaling expression

WC x, 4) w D(Re-2K, Le2K12 (3.30)

where, with k < 1,
D(u, v) = G(u/k, k, 2kv). (3.31)
Identification of the scaling combination u = Re-2K as R/&T), shows that the
zero-field correlation length f(T), diverges as e2Kwhen T -+ 0. The correctness of
this result is easily checked from the exact expression [25] G(R, T, 0) = (tanh Q?.
The complete exact expression for the correlation function may be obtained
via the transfer matrix approach [25] which yields

G(R, K, L) = [sinh2 L + (h+/hJR]/(l + e4Ksinh2 L), (3.32)

where
h* = eK cash L -& (e2K sinh2 L + e-2K)1/2. (3.33)

The constant term in (3.32) merely represents (sJ2, the square of the magnetization.
On extracting the behavior for L, K-l + 0 the scaling form (3.30) is confirmed with

D(u, v) = (1 + zP-~{Y~ + exp[-2u(l + z?)~‘~]}. (3.34)

3.4. Critical Exponents in One Dimension

The zero-field critical exponents 01,p, y, and v are normally defined in terms of
the reduced temperature variable t = (T - TJ/Tc via

Of = f - fc N Pa, (so) - ta,

(3.35)
x - t-y, E - t-“.

In the present case, where T, = 0, these definitions clearly fail. Alternatively if

one replaces t by AT = T one would conclude that 2 - 01,y, and v are all equal
to +co. However, one can define reduced critical exponents in terms of the
correlation length by writing
Af - t--(24/v, 00) - c?lv, x - p/v. (3.36)
 LOW DIMENSIONAL ISING SYSTEMS 241

The results Of- e-2K, and x - ,$ - ezK then imply

y=v=2--0i. (3.36)

The last equality here is consistent with the hyperscaling relation [22]
dv = 2 -- 01for d = I. The first equality combined with the scaling relation
y = (2 -- q)v implies
rj = 1. (3.37)

However, this can be checked directly through the usual definition

G,(R) - 1/Rd-2fn N R1-R, (3.38)

where the last relation applies for d = 1. Since G(R) = (tanh K)R in zero field [25],
one has G,(R) z 1 which confirms (3.37). This value of 7 is also consistent with
the spin resealing factor c* = c = 1 noted above, as follows from (2.29).
The hyperscaling relation d(6 - l)/(S + 1) = 2 - 17now leads to the prediction
6 = co. This result makes good senseif one examinesthe magnetization isotherms
which follow from (3.24) or (3.22) and (3.26) namely,

M(T, H) = <so) = sinh L/(e4K + sinh2L)lj2 m L/(x + L2)lj2. (3.39)

For x or T > 0 this describes a normal, analytic paramagnetic magnetization

curve saturating at M(T, f co) = &I. For x = T = 0, however, one obtains
M(0, H) = sgn(H]: this discontinuous critical point variation is well described
by6 = co.
Finally we note that M(T, H) always vanishes as H + 0 for T # 0, even if
x < 0; thus there is no spontaneous magnetization. However, following (3.36)
one may define j3 through the scaling form [22]

M m ~-B~yW(L~Aly). (3.40)

Comparison with (3.39) yields

p=o and A == v. (3.41)

These results are consistent with (3.36) and the usual exponent relations
01+ 2p .f y = 2 and A = p + y.

3.5. Dedecoration Group with b = 3

We now discussan alternative dedecoration renormalization group with b = 3

generated simply by removing the first two spins out of every triple of spins along
the chain. Since the renormalized interactions between the remaining spins retains
242 NELSON AND FISHER

the sign of the original nearest neighbor coupling, this group provides a natural
framework for discussing antiferromagnetic as well as ferromagnetic chains.
Thus we include a staggered field

Hi = k,TG (3.41)

and take the reduced Hamiltonian to be

2 = K 2 sjsj+l + L 2 ~j + G 5 (A)’ Sj + NC(K, L, G). (3.42)

j=l j=l j=l

The object to be removed by dedecoration is now a block of two spins with

conditional partition function

#(sl , s2> = 2eK cosh[K(s, + .sJ + 2L] + 2e-K cosh[K(s, - s2) + 2G]. (3.43)

Insertion of this into (3.1) to (3.3) yields the recursion relations

w’ = w3x3Y3z3/*l*2hh, (3.44)

x’ = &dJ2/$3#4 3 Y’ = Y$4/$3 3 z’ = z*1/*2 (3.45)

where
z = e-2G = exp(-2H+/k,T) (3.46)
and

$4 = Y(1 + xz3 + 4 + Y”), $2 = Y(X + z"> + -41 + Y”>Y (3 47)

$3 = XY(l + z") + dl + XY3, #4 = xy(l + 2") + 4x + y">. .

The three recursion relations (3.45) are independent of w and determine the
fixed points. In the cube 0 < X, y, z < 1, which describes ferromagnetic inter-
actions, we find a plane of paramagnetic, infinite-temperature fixed points at
x* = 1, 0 < y*, z* < 1, as shown in Fig. 3. There is also a line x* = 0, y* = 0,
0 < z* < 1 of frozen, infinite-field fixed points and, finally, an isolated ferro-
magnetic tixed point at x* = 0, y* = z* = 1 (see Fig. 3). Linearization about
this ferromagnetic fixed point with dy = y - y* and AZ = z - z* yields

AX’ = 9Ax, Ay’ = 3Ay, AZ’ = &AZ. (3.48)

Thus the eigenvalues are

A, = 9, A, = 3, A, = 4, (3.49)
 LOW DlMENSIONAL ISING SYSTEMS 243

Z
FIG. 3. Fixed points for the b = 3 dedecoration group applied to the linear Ising chain in
both staggered and uniform magnetic fields. (The front face of the cube with Ht = 0 corresponds
to Fig. 2). Isolated fixed points occur at (0, 1, 1) and (0, 1, 0), a line of fixed points at (x = 0,
JJ = 0), and plane of fixed points appears at x = 1.

or recalling that b = 3,

h, - 2, Au = 1, A, = ---I. (3.50)

Evidently, as anticipated on physical grounds, the staggered field variable z is

irrelevant at the ferromagnetic fixed point. More strikingly the eigenvalue exponents
X, and h, are precisely the sameas in the original B = 2 dedecoration group (and
correspond to the same critical variables or eigenoperators). This confirms the
expected independence of the critical exponents on the particular renormalization
group and partially checks the semigroup property of the dedecoration class of
groups asymptoticalIy in the linear region.
We may note in passing that a dedecoration type of renormalization group for
a one-dimensional, continuous spin Gaussian model [26] yields a finite T, , and
the eigenvalues A, = 2, A, = # and A, = -4 with exponents y = 2v = 2(2 - OZ)= I
and q = 0.
It is easy to check that the recursion relations (3.44) to (3.47) are invariant
under y 3 y-1 and z 3 z-l as is to be expected. More importantly, if we let
x * x-1, corresponding to the change of ferromagnetic into antiferromagnetic
coupling, the recursion relations are reproduced except that the roles of y and z
244 NELSON AND FISHER

are interchanged. Thus the analysis of the antiferromagnetic chain exactly parallels
that for the ferromagnetic chain. In particular there is an antiferromagnetic fixed
point at T = 0, Ht = 0, at which the parallel field H (or v) is an irrelevant variable
with eigenvalue A, = - 1, while the staggered field becomes relevant with exponent
A, = 1.

3.6. Scaling with an Irrelevant Variable

The spin-independent term w can be handled as before. If the initial zero-spin
energy is taken as E,, = -J then (E + J)z = 3z(E + J),, remains zero throughout
the linear region. If attention is concentrated on the equation of state M(T, H, H+)
rather than the free energy one can, in fact, avoid all consideration of the spin-
independent term. This is because the recursion relation for the magnetization is
simply
M(x, Lly, AZ) = M(x’, dy’ AZ’) (3.51)

independent of w. Of course this is just another reflection of the simple linearity

of the dedecoration groups with c = 1 already noted in Section 3.3 in connection
with the correlation functions; it shows that the renormalization group trajectories
form curves of constant magnetization. On using (3.48) for the linear region this
yields the scaling expression

M(x, dy, AZ) m B(Le2K, Ge-zK). (3.52)

This result may be compared with the exact result

A4 = e2K sinh 2L(2 + 2 cash 2L cash 2G + egKsinh2 2L + e--4Ksinh2 2G)-lJ2.

(3.53)

Comparison with (3.52), near the fixed point x = 0, L = G = 0, shows that the
scaling function is
B(v, v’) = v/(1 + v2 + v+~)~/~. (3.54)

Since G is an irrelevant variable v+ = e-2KG --f 0 as T+ 0 and so asymptotic

scaling is obtained by replacing B(v, v+) by B(v, 0). By expanding B in powers of
v+ one is tempted to conclude [following the argument sketched after Eq. (2.26)]
that the leading corrections to the asymptotic scaling form due to the irrelevant
variable G should be represented by

+ O(G4x2)/. (3.55)
 LOW DIMENSIONAL ISING SYSTEMS 245

However, expansion of the exact result (3.53) neglecting corrections of order L2

(arising from the sinh 2L and cash 2L terms), yields a true correction factor

j1 _ G2(l + x)
(3.56)
I 2{1 + (LZ/x)] + G(G4$

which in order G2 is larger by a factor of about x-l = e4K, which actually

diverges as T--f 0. At first sight this severe discrepancy is quite puzzling. However,
in deriving the scaling relation (3.52) from the exact recursion relation (3.51) we
used only the linearized recursion relations (3.48). We conclude that the most
significant corrections are “nonlinear” corrections rather then “irrelevancy”
corrections. We will discuss such nonlinear corrections in Section 6. For the
moment, we merely observe that replacement of the scaling field x by

I = x cosh2 G ,Q x(1 + G2) (3.57)

enables one to use the reduced or asymptotic scaling form

M(T, H, H’) SWB(LW, 0) = [l + (P/L2)]-112, (3.58)

and reproduce the leading part of the correction (3.56) correctly.

It is clear that for an antiferromagnet we merely have to interchange Hand H+,
or L and G, in the foregoing analysis.

3.7. Pseudorelevant Variables: Antiferromagnetism with the b = 2 Dedecoration

Group
The renormalization group treatment of the antiferromagnetic chain led to the
difficulties noted in Section 3.2. Specifically if one starts with

-1 = X0 = ep41K' < 1
x0 (3.59)

as appropriate to antiferromagnetism, then one iteration of the recursion relation

(3.11) yields

x1 = (l + X0(1 + Yo12
xoyo)(xo + yo) = 4Xo[l - 2x0 + ..*I, (3.60)

which represents aferromagnetic coupling which remains so after further iteration.

On the other hand the initial parallel field variable y. = 1 + dy, , which should
represent an irrelevant variable at the antiferromagnetic critical point, X0 = 0,
is by the recursion relation (3.12), transformed again into a parallel field. However,
a parallel magnetic field is a relevant variable at the ferromagnetic fixed point
246 NELSON AND FISHER

and we accordingly expect the scaling relation (3.22) to hold with K merely
replaced by 1K I. Two derivatives with respect to L would then indicate a sus-
ceptibility diverging strongly at T--f 0.
This conclusion must be false! To find the gap in the reasoning let us rewrite
the recursion relation for y in terms of the critical point deviation dy = y - 1;
this yields
Al,’ = 2441 + BAY) = 2dy[l - x + @ly + . ..I. (3.61a)
1+x+x4

so that in the ferromagnetic case dy, = 2~ly, near the critical point, as expected.
However, using (3.59) for antiferromagnetic coupling yields

dy, = 2% ~Ydl + UYO) 2X, dy,[l - X, - @ly, + ..*I. (3.61b)

1 + x, + dy,_ =
It is now clear what happens: after one iteration the field L, w &ly,, is replaced
by L, m &dy, m X0 L3y, w 2e-41KIL, . We may then use the scaling relation (3.22)
with K replaced by 1K 1 and L replaced by $L, which yields

f(K, L) M e-21RI Y(L,e 21~1-41~1) = e--2lK/ y(&-2lKl). (3.62)

As T + 0 the argument v = L,e-21KI vanishes for any fixed L, . Thus the variable
L, is only a pseudorelevant variable. One step of the iteration has multiplied it by
the factor X0 = e-41Kl which, “by accident,” vanishes strongly at the critical point
and so converts L, from a relevant into an effectively irrelevant operator.
This analysis indicates the potential importance of the first few iteration steps
in obscuring the true nature of the critical behavior. It illustrates that for a straight-
forward interpretation one must always attempt to chose a renormalization group
which focuses on the critical behavior of interest and represents it by a fixed point
of corresponding character.

3.8. Attracting Fixed Points and the Free Energy

If all its critical variables are irrelevant a fixed point may be termed an attracting
fixed point since all trajectories flow into it while none leave it. It is evident from
Figs. 2 and 3 that the paramagnetic high-temperature fixed points form such an
attracting set. Since it has no relevant variables, an attracting fixed point cannot
describe genuine critical behavior. However, since all, or a large class, of trajectories
flow into an attracting lixed point one can use the basic recursion relation [(2.7) or
(3.19), etc.] to find an expression for the total free energy.
We will illustrate the procedure by analyzing the simplest, zero-field situation
 LOW DIMENSIONAL ISING SYSTEMS 247

with H == H+ = 0 or y = z = 1. The attracting fixed point is then x* = 1 or

K = 0. After iteration the basic relation reads

f(% 2x0) = 2-zf(wl 34. (3.63)

As I--f co the attracting fixed point is approached. The partition function near
the fixed point is easily found to be

ZN(G 7 Kz) - 2” exp[NC + $NK,2 + O(NK,4)], (3.64)

so that, as I+ co,

f(wz , Xl> = In 2 + C1 + $K,2 + ...,

M - ln(&wt’4) + O(Llx’), (3.65)

where dx = x - 1 NN -4K. By combining with (3.63) we conclude that

(3.66)

This demonstrates explicitly what is, in fact, a general result [I], namely, that the
free energy can be expressed wholly in terms of the development of the spin-
independent term w1 (or C,).
To evaluate (3.66) we introduce the variables
1 - xl12
21= &(wx)l/4 and v = tanh K =
1 + xl/2 ’

in place of w and x. The recursion relations (3.10) and (3.11) [with y = I] then
become
zl’ = U2(1 + q/(1 + ?Y), (3.68)
v’ = v2. (3.69)

The simplicity of this second recursion relation provides, of course, the justification
for this choice of variables. Since x2 + 1 the result (3.66) becomes

f&l 3wO) = - I~+JI2F In uI . (3.70)

Iteration of (3.68) and (3.69) reveals that

In u2 = 2’ In u0 + 2” ln(1 + 23 - ln(1 + L$“). (3.71)

 248 NELSON AND FISHER

For o0 < 1 (i.e., T # 0) substitution in (3.70) yields the final solution

f(% >X0) = - In[u,(l + vO)] = - * In trO + ln(xk’4 + ~;l’~)

= Co + ln(2 cash K), (3.72)

which is the known exact answer [25].

It is clear that the same procedure will work more generally although the
difficulty of solving the recursion relations analytically may be insurmountable [27].
In principle, however, a numerical solution could still be used to explore the critical
region along these lines.

4. FURTHER LINEAR MODELS

4.1. Alternating Coupling Strengths

By enlarging the space of Hamiltonians to allow for nearest neighbor coupling
constants which alternate along the linear Ising chain one discovers examples of
a fixed point with a marginal operator. (Examples of relevant and irrelevant opera-
tors were, of course, discussed in Section 3.) As explained in Section 2, a marginal
operator is characterized by a renormalization group eigenvalue (1 = 1 or h = 0.
We will utilize the b = 3 dedecoration group in which two out or every three
spins are removed. The recursion relations may be derived quite straightforwardly
as explained in Section 3.1. If K, and Kb are the alternating coupling constants,
the recursion relations in zero field are most simply expressed in terms of the
variables
v, = tanh K, and vb = tanh Kb (4.1)

as introduced in (3.67). The recursion relations are then

0, ’ = va2Q)) VO’ = Vb2V,. (4.2)

Linearization of the equations about the ferromagnetic fixed point v,* = Q,* = 1
yields
Au,’ = 2dv, + Au, + O(Av2),
(4.3)
dub’ = AZ,, + 2Avb + O(Av2).

By diagonalizing the linear part of these equations the critical fields are seen to be

h, = Au, + Au, and h, = Au, - Au, (4.4)

 LOW DIMENSIONAL ISING SYSTEMS 249

with eigenvalues A, = 3 and (1, = 1 or

A, = 1 and A, = 0. (4.5)
Thus h, ,V K, - Kb is a marginal field; the corresponding eigenoperator is
essentially
P2 = c d%i+1 - &i-l>- (4.6)

(The value A, = 1 is consistent with the earlier result A, = 2 if we note that

Au c? --2x1/2.)
The marginality of h, can be understood by noting that there is actually a line
(or curve) of critical fixed points given by u,* = l/z+,*, which just touches the
physical region 1 v, 1, I vl, j d 1 at v,* = &, * = 1. This is illustrated in Fig. 4

FIG. 4. Trajectories and fixed points for a linear Ising chain with alternating coupling strengths
J1 and Jz . A line of fixed points is determined by v1v2 = 1 [with ZQ = tanh(J&J)J, and an
isolated attracting fixed point occurs at (0,O).

which also shows the trajectories. The fact that the other fixed points lie outside
the physical region, means, correctly, that one cannot achieve criticality and long
range correlations unless dv, and dvb --f 0, i.e., unless both K, and Kb -+ co; but
one can still have J, # Jb as k,T + 0. Nevertheless the eigenvalues at the other
fixed points are the same as about the physical point. This can be seen most easily
by defining new variables CLL= v,/v, * for p = a or b. The recursion relations (4.2)
are unchanged in terms of the fiU so that the subsequent analysis, including (4.5),
remains valid with U, replaced by V, .
250 NELSON AND FISHER

The invariance of the eigenvalues and, hence, of the critical exponents along
the line of fixed points is in contradistinction to what is presumed to happen in
the two-dimensional eight-vertex or Baxter model [28, 291 where the eigenvalues
vary along the line of fixed points so that the critical exponents vary continuously
with the marginal field l-281.

4.2. Second Neighbor Interactions

The Hamiltonian of the linear chain with nearest neighbor interactions Jr and
second neighbor interactions JZ in zero field is

(4.7)

If we rewrite SiLTi+, as (s~s~+~)(s~+~s~+~)

and introduce new spin variables

Ui = SiSi+l 3 (4.8)

the Hamiltonian assumes the form

Af = -J1 i cri - J, 5 uiui+l . (4.9)

i=l i=l

Furthermore for an open chain the ui take the values &l quite independently.
Thus the problem of second neighbor interactions in zero field has been reduced
to that of first neighbor interactions of strength J = J, in a field H = J1 .
We can thus take over the analysis of Section 3.2 for the b = 2 dedecoration
group removing alternate 0 spins. In particular the recursion relations (3.10) to
(3.12) apply but with

x = e-4K2 =
exp(--4J2/kG9 and y = pK1 = exp(--2J,/k,T). (4.10)

The original ferromagnetic fixed point x* = 0, y* = 1, now corresponds to zero

first-neighbor coupling (so that the chain decomposes into two disconnected
second-neighbor chains with Jz > 0). But this fixed point is unstable to perturba-
tions with dy = y - y* # 0 (or J1 # 0) and so the system crosses over to the
fixed point at x* = 0, y* = 0 (originally termed the “frozen” fixed point). However
one cannot directly linearize in terms of x and y about this fixed point owing to
the factor x/(x + y) in (3.11). This difficulty can be circumvented by eliminating x
in favor of
z = xy = e--4Kz--2K~, (4.11)
 LOW DIMENSIONAL ISING SYSTEMS 251

which leads to the new recursion relations

Y’ = (Y” + Nl + 4, (4.12)
z’ = Z(1 + y)2/(1 + z)“. (4.13)

(Note that we use z here with a different meaning than in Section 3.5). These
equations are easily linearized to yield

y = z + w2, z2), (4.14)

z’ = z + O(y2, z2), (4.15)

from which critical fields are

h, = z and /j 2 = y - z = e-‘K’(l - e-4K2>

, (4.16)
with eigenvalues (1, = 1 (h, = 0) and (1, = 0 (X2 = -co). Thus the first variable
is marginal while the second might be termed totally irrelevant! Evidently h,
vanishes faster than any exponential, which suggests that K2 converges extremely
rapidly to zero. To check this we form the recursion relation for the variable

u = h,/z = (y/z) - 1 = edK2- 1 (4.17)

and find
2.4’= z?z/(l + z + UZ)2, (4.18)

from which one can show that U’ is always less than U. Furthermore once u2z < 1
or K, < &, the value of u goes very rapidly to zero, which implies that K,
also vanishes rapidly (even though ICI may be large). The fixed point is thus
described by only nearest neighbor interactions.
Since the remaining variable h, = z is only marginal rather than relevant, as
might have been expected, we learn nothing immediately from the general linear
analysis of the renormalization group given in Section 2. As a matter of fact z is
weakly relevant (for J, > 0) since its recursion relation can be written

z’iz = (1 + -j+Jz = (1 + -+)2 > 1, (4.19)

for z f 0; however, this variation does not contribute significantly since h2’ m h22
so that hhE)goes rapidly to zero.
Nonetheless we can obtain information if we recall the recursion relation (3.10)
for the constant term w = e-4c. This becomes

w’ = (1 + v)2 (;27y2)(1 + z) a (lWTz)4 ’

252 NELSON AND FISHER

where the second part follows since h, = y - z is so strongly irrelevant. Iteration,

using the fact that z is marginal, so that ztz) m z, then yields

,$.(Z) m [wz2/(1 + z)“]“” z-2(1 + z>*, (4.21)

while the free energy renormalization relation gives

f(w, y, z) m 2-Ef(w’l’, z, 2). (4.22)

Now we choose I so that wtz) = k and so obtain

f(w, y, z) M [ln wz2 - 4 ln(1 + z)]f(k, z, z)/ln[kz2/(1 + z)“]

5% -4[C + 2K, + Kl + z + U(z2)1 Q(z), (4.23)

where we have used (3.9) for w, and (4.11) for z, while Q(z) stands for the residual
function of z in the first line. However, we know on general grounds that the
constant term C can enter intofonly linearly and with coefficient unity! It follows
that -4@(z) = 1 (at least for z < 1) and so we finally conclude

f(w, Y, z) m C + 2K, + Kl + z + U(yz, z2), (4.24)

where the correction of order yz is anticipated directly from (4.19). Recalling the
identifications K = K, and L = Kl , obtained by going to the 0 variables, this
expression agrees precisely with that following from the exact result (3.24) in
the limit T + 0 with J = JZ and H = J1 fixed. We conclude that even though
the standard linearized renormalization group analysis of Section 2 fails, the
general formalism can still be used to derive the asymptotic free energy from
behavior near an appropriate fixed point.

4.3. Braced Ladder

To illustrate a more complex but still exactly realizable renormalization group
involving many-spin interactions we consider the “braced ladder” of spins
illustrated in Fig. 5. This consists of two parallel one-dimensional chains of spins
with interactions along the nearest neighbor bonds (sides and rungs) and second
nearest neighbor bonds (braces) plus a four-spin interaction of strength J4 = k,TK,
around each braced square (as indicated in the figure by the dotted loops). With
the notation of the figure, and supposing all bonds to be ferromagnetic (Ji > 0),
we may define
xi = e-4K: = exp(-4Ji/k,T). (4.25)

If we now remove alternate pairs of spins (which form a ring) as indicated in

the figure, we generate a b = 2 dedecoration group as before. Even if J4 was
 LOW DIMENSIONAL ISING SYSTEMS 253

FIG. 5. Dedecoration transformation for a braced ladder of spins; J4 denotes a four-spin

interaction.

initially zero we would be forced to introduce a four-spin coupling at this stage.

This effect of “induced interactions” is, of course, a very general feature of
renormalization groups which, indeed, normally prevents their construction in
explicit closed form. If we define

$1 = $(l + Xl”Xf) + X1X&Xq , (4.26)

#2 = &(x1” + x32) xz3 + x,x,2x,x, (4.27)

$3 = ~1x*2~3(x2 + x,), (4.28)

$4 = x,x3(1 + x,xI>, (4.29)

$5 = &x,x,3x,x,[l + XIX2 + XIX3 x2x312, (4.30)

the required recursion relations become

x;” x
lfi3~4h4 9 XL2 = ~21fi3/&%~4 > (4.31)

x;2 = b%4/~1~3 9 xi” = ~52/~1+2~3~4~ (4.32)

These formulae define trajectories for the Hamiltonian in the four-dimensional

space (x.~, x2 , x3 , x,), which eventually terminate in a line of attracting para-
magnetic fixed points given by x1 = x3 = xq = 1, 0 < x2 < 1.
For special parameter values, specifically x2 = x3 = x4 = 1 and x1 = x2 =
xq = 1, the equations degenerate into versions of the simple linear chain dedecora-
tion group discussedin Section 3. However, at the interesting ferromagnetic fixed
point x1 = x2 = x3 = xq = 0 the equations are nonanalytic as in the linear second
neighbor case discussedabove.
Investigation indicates that after a few iterations x, becomesclosely equal to x, .
In fact the condition x1 = x3 or Jr = J3 is preserved by the recursion relations,
as must be so on the grounds of symmetry. We conclude that AJr3 = Jr - J3 is
an irrelevant variable and confine ourselves to the case x1 = x, . However, the
recursion relations are still nonanalytic about the ferromagnetic fixed point.
254 NELSON AND FISHER

Accordingly we work with x1 and the variables

3’ zzz x2/.x1 = ,-2(K~-Kl), (4.33)
z = x2x4 = e--?(Kz+KJ
> (4.34)

and obtain the new recursion relations

Xl’ = 2/z x,(1 + z)‘l”/(l + 2x,% + x,4)1/2, (4.35)

Y’ = (WJ + Ml + 4, (4.36)
z’ = &z(l + x12 + 2x,9)/(1 + z)(l + 2X,22 + x14). (4.37)

From these it is evident that x1 is a relevant variable with X, = 4, while w is an

irrelevant variable with h, = -1. Linearization of (4.36) with (4.37) shows that
w = y - 22 is a totally irrelevant variable with X, = - co. What happens therefore
is that the difference between y and 22 goes rapidly to zero so that J4 is forced
to equal $(ln 2) kBT - Jz which, for small T involves a change of sign of the
four-spin interaction.
The significance of the eigenvalue X, = & can be understood in terms of the
linear Ising chain if we study the correlation length,

.$ - t-u, (4.38)

where Y = I/h, and t represents the temperature-like variable e-4K1. For the
simple chain, h, = 2 and we predict f - t-!’ m e2K1. The recursion relations
(4.35)-(4.37) indicate that K, quickly diverges whereupon the sides of the ladder
become “locked” together. The four-spin term K4 now only contributes a constant
to the free energy, and the ladder reduces to an Ising chain with an effective nearest
neighbor coupling 4J, (assuming J1 = J3). This is precisely the information
conveyed if we apply (4.38) with h, = 4:
,$ N t-” m e+8K1. (4.39)

As in Section 4.2 we could go on to analyze the free energy but this does not
seem worthwhile.

5. TRUNCATED TETRAHEDRON MODEL

The truncated tetrahedon model is a planar Ising model of coordination number

three defined through the decoration and star-triangle transformations [17, 181.
It is simply related to another planar model that is four-coordinated. To describe
the truncated tetrahedron “lattice” we start in zero order with a tetrahedron of
 LOW DIMENSIONAL ISING SYSTEMS 255

FIG. 6. Conversion of a tetrahedron of Ising spins into a truncated tetrahedron lattice of

order n = 1 by successive application of decoration and star-triangle transformations.

four Ising spins. This system is then decorated by putting two spins on every bond
as indicated in Fig. 6. To complete the transition to the first order lattice, a star-
triangle transformation is made in order to remove each of the original four vertex
spins. We are finally left with the first order “truncated tetrahedron” shown in
the last part of Fig. 6.
The next order lattice in the hierarchy is obtained in a precisely similar fashion.
To generate a lattice of order n + 1, we decorate every bond in a lattice of order IZ
with two spins and then make a star-triangle transformation to remove all spins
at three-coordinated vertices. If we define the order of the initial tetrahedron of
spins to be n = 0, then a lattice of order n is three-coordinated, and contains

FIG. 7. Section of an infinite truncated tetrahedron lattice.

256 NELSON AND FISHER

N = 4 x 3” spins and 6 x 3” bonds. We pass to the thermodynamic limit by

letting n --f co. A section of an infinite truncated tetrahedron lattice is shown in
Fig. 7. A lattice of order n will contain polygons of sizes 3, 6, 12,..., 3 x 2”, but
will always have a connection number [19] of only 3. Thus the circumference [19]
of a lattice of order n + 1 will be twice that of a lattice of order n; but a lattice
of any order can still be cut into two pieces of arbitrary size by cutting three bonds.
For this reason we may guess that the transition temperature will still be at T, = 0.
However, the critical behavior should be quite distinct from the linear chains.
The lattice described above is related to a similar lattice of coordination
number 4. If we take a truncated tetrahedron lattice of order it and decorate every
bond with a single spin, and then make a star-triangle transformation at every
three-coordinated vertex, we obtain a four-coordinated lattice. The sequence of
transformations used to arrive at this four coordinated “fully” truncated lattice
from the truncated tetrahedron lattice is illustrated for a lattice of order 1 in Fig. 8.

FIG. 8. Conversion of a truncated tetrahedron lattice (of order n = 1) into a “fully truncated”
lattice of coordination number four by applying a decoration and star-triangle transformation.
Note that only a single spin is inserted by the decoration while in Fig. 6 two spins are inserted.

Now the decoration and star-triangle transformations yield an explicit relation

between a lattice of order n and a lattice of order it + 1. However, because of the
properties of the star-triangle transformation, this relation can be obtained only
for zero magnetic field. Let 2, be the partition function of a truncated tetrahedron
lattice of order n with zero constant term E,, (there being N = 4 x 3n spins in this
 LOW DIMENSIONAL ISING SYSTEMS 257

lattice). We assume there is a uniform spin-spin coupling of strength J, and

introduce the variable,
v, = tanh(J,/k,T), (5.1)

as before. Then, we find

ZdhJ = k(bJ13n-1Z-1(%-1) (5.2)

with
g(v) = 256(1 - u + 2v2)/(1 + 0)” (1 - u,)~(1 v3y (5.3)

and, with v replacing v, and v’ replacing u,-~ ,

V’ = zP/(l - v + 0”). (5.4)

If we take the logarithm of (5.2), divide by N = 4 x 3”, and take the thermo-
dynamic limit y1+ co, we obtain for the free energy the relation

f(v) = 39(v’) + & In g(v). (5.5)

Evidently we have constructed a renormalization group. The second term,

depending on g(v), arises because we defined 2, , and hence f, with zero constant
term, E,, . If we regard the dimensionality of the truncated tetrahedron as d = 1,
on the grounds that it can be cut into indefinitely large pieces by only three cuts,
then we should take b = 3. On the other hand, if we measure distances along the
bonds, as is quite natural, we find that each step of the iteration corresponds to
a length resealing of only b = 2. This can be seen from Fig. 7 where the number
of points along the “side” of a basic triangular figure (on “tetrahedral face”)
goes up as 2” while the number of points increases as 3”. i3y this argument the
dimensionality of the lattice is d = log, 3 N 1.5850; however, although the
dimensionality is, by this measure, larger than unity the critical point will still
be at T, = 0. To see this, we examine the recursion relation (5.4). This relation is
compared in Fig. 9 with the analogous relations for the linear Ising chain dedecora-
tion groups b = 2 and b = 3 which are u’ = v2 [by (3.69)J and 8’ = v3 [say, from
(3.45) with y = z = 11. All formulae have a infinite temperature attracting fixed
point v := 0, and an unstable low temperature, ferromagnetic fixed point u = 1
corresponding to T, = 0. However, since the slope of the graph for the truncated
tetrahedron approaches unity as v --+ 1 we are dealing, as in Section 4.1, with a
marginal operator, rather than a relevant one.
If we set
v=l-6, fi m ze-2” - 2&K + . . . 7 (5.6)
258 NELSON AND FISHER

I- I I I ,

V’

0.5-

FIG. 9. Recursion relations for two different models expressed in terms of o = tanh(J/ksT).
Curves (a) and (b) represent the b = 3 and b = 2 dedecoration groups for the Ising chain, while
(c) is the corresponding curve for the truncated tetrahedron model. Both models have fixed
points at v = 0 and v = 1.

we obtain the recursion relation

jj’ - fi zzz fi2[1 - $(I - fi + q-q. (5.7)

When V is small, i.e., near its critical value, we may approximate this relation by
the differential equation
dC/dl w ii2, (5.8)

(see the Appendix for a justification of this step), where G(I) is the result of 1 itera-
tions. This equation has the solution

V(l) = i&/(1 - r&J. (5.9)

Now, since b = 2, if we measure distances along the bonds, the correlation

length transforms as
[[V(Z)] = 2-C@,). (5.10)

On writing V = 15(Z)and eliminating I with (5.9) one finds

(5.11)
 LOW DIMENSIONAL ISING SYSTEMS 259

so that the correlation length diverges as the exponential of an exponential when

T-t Tc, namely as

t(T) - exp[&(ln 2) exp(2J/k,T)]. (5.12)

This indicates the strong degree of cooperativity in the model.

By casting the recursion relation (5.4) into an integral equation, we can formally
extract the I dependence. Repeated iteration of (5.4) gives

l-l
Vl = u. + c <I?;, - VT,)/(l - jj,, + fi;,>. (5.13)
I'=0

Solving this discrete analogue of an integral equation for i$ iteratively, we obtain

1’1= co + ho2 + Z(Z- 1) V”3 + O(V,4). (5.14)

The thermodynamic behavior near T, = 0 can be found directly from the

basic relation (5.5). Expanding In g(V), we obtain

In g(V) = 4 In 2 - 6 In E;+ 3ti + F2 + O(I?~), (5.15)

which leads us to form the expansion

f(C) = A In C + B + CO + Dr! + 0(03), (5.16)

Substituting this expression into (5.5), we obtain

f(E) = -2 In V + & In 2 + $$ C2 + 0(03), (5.17)

from which we conclude that the energy varies as

U(T) w -$J + $.Je-2K + 3Je-4K (5.18)

as T-0.
Although this behavior does not look particularly anomalous one should
recall that the lowest order temperature dependence of a two or three-dimensional
lattice of coordination number q would be z = exp(--2qJ/k,T), as follows easily
by overturning spins from the fully ordered state [25]. This would yield a leading
variation of z = e-‘jK, whereas (5.18) yields d U - zI/~; this makes the anomalous
behavior evident (a similar analysis of the linear chain yields d U - z1i2, again
indicating “critical” behavior).
260 NELSON AND FISHER

6. CORRECTIONS TO SCALING AND NONLINEAR FIELDS

6.1. Introduction
The problem of corrections to asymptotic scaling has been treated by Wegner [3]
in terms of nonlinear scaling fields. We will apply his ideas to some of our simple,
exactly soluble models. It will be convenient to work with the variable x = e-4K
as before and the complementary variables

j7 = 1 - y = 1 - e-2L, ,?j = 1 - z = 1 - e-20.

(6.1)

[See (3.9) and (3.45).]

Consider first the case of a linear Ising chain in a uniform magnetic field (Z = 0).
In terms of the variables x and 7, the recursion relations (3.11) and (3.12) for the
b = 2 dedecoration group have the Taylor series expansions:

x’ = 4x - 8x2 + 12x3 + xj” +

. . .
3

(6.2)
7 = 2y - y2 - 2yx + 3xy2 t 2yx2 + *-- .

For the case b = 3 [see (3.45)], the expansions are

x’ = 9x - 48x2 + 208x3 + 6y2x + -.. ,

(6.3)
7’ = 3j - 3y2 - 8jx + 20xJ2 + 24x2j + y3 + .a. .

The coefficients of these Taylor series are obviously b-dependent, and the
leading coefficients have the form bAc,bAu,with X, = 2, and X, = 1 (as before).
This form is expected from the semigroup property of the renormalization group.
We will present a procedure for determining the b-dependence of the higher
order terms in the recursion relations, given the recursion relations generally for
a particular b.
The scaling prediction of the renormalization group for the one-dimensional
Ising model is that, in the linearized region about the fixed point, the magnetization
is a function only of the ratio y”/x. (We consider the magnetization rather than
the free energy to avoid worrying about the constant term for the moment.)
Following Wegner, we try to find nonlinear scaling fields gZ(x, J) and g,(x, 3)
which correspond to “exact eigenfunctions” of the renormalization group operator
Iwb. These nonlinear fields are (i) to reduce to x and 7 near the fixed point and
(ii) to transform simply under the renormalization group according to

(6.4)
 LOW DIMENSIONAL ISING SYSTEMS 261

The result of iterating these recursion relations is obvious. The scaling prediction
is then that the magnetization is a function only of the ratio gU2/g, . This prediction
should hold exactly for all values of g, and g, , not just in the linearized region
about the fixed point (g,*, g,*) = (0,O).
Wegner [3,30] has presented a procedure for finding the expansions of g, and g,
in powers of x and 9 given the recursion relations in d@rential form. We review
this briefly. Suppose we know the derivatives with respect to b evaluated at b = 1,
of the recursion relations for a set of fields (hi}, namely,

The a,,,(b) are b-dependent coefficients appearing in the original recursion relations
for b > 1, and the hi are the eigenvalues of the renormalization group. We assume
we have chosen the hi so that the recursion relations are diagonal to first order,
as indicated in (6.9, i.e., the hi are linear scaling fields as in Section 2. The problem
is to find nonlinear scaling fields gi which behave in the simple manner indicated
in (6.4) under the action of the renormalization group. Wegner [3, 301 assumes
that the hi can be expanded in a power series in the gi , as

(6.6)
ik

and then finds that the desired coefficients bijk are given by the set of equations

(hj + h, - A,) bijk = (%j,/W,,, . (6.7)

More complicated expressions are found [3] for the higher order coefficients in (6.6).
For the special case where hj + hk - A, = 0, Wegner finds that logarithmic
corrections must be introduced [3, 301.
All the renormalization groups we have discussed involve only discrete values
of b. Thus, it is impossible to determine required partial derivatives like
(aa,,,/ab),_, . Accordingly we will develop a method for determining the nonlinear
scaling fields gi when only a discrete renormalization group is given. The method
follows Wegner’s general approach.

6.2. Nonlinear Scaling Fields for Discrete Renormalization Groups

We assume, as in (6.9, that the recursion relations for a set of fields have
already been diagonalized to first order, so that we may take
m 1
hj’ = b”‘hj + 1 7 C ajr&7,(,) . (6.8)
n=2 n. l(n)
262 NELSON AND FISHER

Here, Z(n) is a multiindex which represents iliz **. i, , and

lZ&) = h,,h,z *-. bin . (6.9)

The fixed point is given by hi = 0 (all i). The coefficients Q,(,) are in general
functions of b, but we suppose that they are given for a particular b > 1. The
nonlinear scaling fields gj are, as before, to satisfy

gj’ zz bAjgj . (6.10)

Since the gi should reduce to the hi in first order, we try to express them as

(6.11)

The cjlcn) are to be determined, and are expected to be independent of the scale
factor b.
We will calculate the gj by obtaining a sequence gj2)(hi), gj3)(h,),... such that
gj”)(h,) satisfies (6.10) to nth order. The function gj”(hi) is an nth order polynomial
which agrees with gjvpl(hi) to U(h~~-‘). To calculate g(2)(hi) we form the product

h’I(Z) = bA’@‘h(2) + WQ,,), (6.12)

where we use the notation

hI(n) = hiI + ... + hin * (6.13)

Forming the linear combination

(6.14)

and adding this to (6.8) we obtain

zzz b”’ [hj + 3 C (bpA’ajr(z) + bAr’2’-A’cjr(z)) ha,,

I(2) 1+ n=3
f ; 2
. I(2)
&,h,) .
(6.15)
The quantity
(6.16)
 LOW DIMENSIONAL KING SYSTEMS 263

will satisfy (6.10) to second order provided

CjI(2) = clj,(2)/(bAl - b”““‘). (6.17)

The a$,& result from adding the higher order terms in (6.14) to the higher order
terms in (6.8).
Suppose in general that we have recursion relations of the form

gp’)(hJ = b”j,gj”“‘(hi)+ (6.18)

Repeating the procedure indicated above we can derive the fields

gjm+l)(hJ = gfyh,) + @ ; 1),. ,& Wm+1hm+1) 3 (6.19)

which are correct to order m + 1 provided

Cj,(m+l) = LZjTLtl,)/(bA’ - bA’im+lJ). (6.20)

Thus, we have obtained the desired sequence of functions gjm+l)(hi) which

should converge to the scaling fields g, . The method breaks down if, at some
stage, XI(3 = hj, which is precisely the case where Wegner [3, 301 rinds the need
for logarithmic corrections.
We note finally that in order for the ciro) to be independent of b as required,
the relation (6.17) dictates the b-dependence of the recursion relation coefficient
ail&b). Once we have evaluated the number cjrcz) for, say, b = 2, we simply
obtain q,(,)(b) = cjr&b”j - bAl(Z)). The appropriate b-dependence of higher order
coefficients can be found in a similar fashion. We will in fact check explicitly for
the dedecoration groups that the same results for the gj follows for both b = 2
and b = 3.

6.3. Application to the Linear Ising Chain

Applying the method outlined in the previous subsection to the recursion
relations (6.3) and (6.4) we can deduce the nonlinear scaling fields for the linear
Ising chain. We find

(6.21)
 264 NELSON AND FISHER

The same results are found starting with either the recursion relations for b = 2
or those for b = 3. Thus g, and g, are indeed independent of b, as desired.
The expression for the magnetization obtained from the exact solution is, when
expressed in terms of x and 7,

M = [l + 4(x/77(1 - y)]-‘I”. (6.22)

If we calculate gz/gV2 from (6.21) we obtain

&/&/2 = (x/Y2)[l - 3 + w3, X2Y,XY2,r311* (6.23)

Thus, terms proportional to X, ~7, x2, and y2 cancel exactly, and the scaling
property of the nonlinear fields is verified explicitly to this order. Expansion of
the nonlinear fields after substitution in the scaling function (3.54) (with U+ = 0)
thus reproduces all the corrections to asymptotic scaling.
We can also treat the case when both uniform and staggered magnetic fields
act. The Taylor series expansions of the recursion formulae (3.45) are

x’ = 9x - 48x2 + 208x3 + 2xZ2 + 6xjj2 + --a , (6.24)

j’ = 3y - 3J2 - 8xj + 20xy2 + 24x27 + ji3 + *.a, (6.25)

z’~fz+~~2+~xz+~z3-~~2r+~X2~+~xZ2+~~2z+ . ..*

(6.26)
Using the methods developed, one finds the nonlinear scaling fields are

gz(x, 7, Z) = x + ; x2 + + x.2 -kx,:2+$3+ . ..)

- -
g,(n,~,z)=y+~Y2+~x~+~xy2+~X2~+~~3+..., (6.27)

- -
g3(X,y,z)=~+~22--x~--x~2--x2~+~~3--il12,:+ .. . .

Again, one can take the appropriate ratios of these scaling fields and verify that
they agree with the exact expression (3.53) found for the magnetization, to the
order to which we have calculated.
Note that, although these nonlinear scaling fields are unique with respect to
the particular renormalization group we are using, they are certainly not unique
as regards their ability to represent the exact-solution in a complete scaling form.
For example, let /3(x, j) be an arbitrary analytic function subject only to /3(0,0) = 1.
Then taking the case of the linear Ising chain in a uniform field, we could replace
 LOW DIMENSIONAL ISING SYSTEMS 265

g, by & I= /P(x, 3) g, , and gv by 6, = 6(x, p) g, , and the ratio &/i,” would have
the same value as gm/gy2 and so reproduce the correct magnetization. These new
“nonlinear fields” would equally diagonalize the linearized recursion relations
(although they need not come from the solution of a full set of recursion relations).
We will actually show in Section 7 that a continuous range of nonlinear scaling
fields exists for a range of distinct renormalization groups.
A natural question to ask is whether one can identify in closed form the func-
tions expanded in (6.27). This we have not been able to do in general, but the
following special cases can be found

(6.28)
g9(0, j, 0) = -ln(l - 7).

More generally, however, the convergence of these series (in some appropriate
sense) is an open question.

6.4. Corrections to Scaling and the Spin Independent Term

So far we have discussed corrections to asymptotic scaling only in terms of the
magnetization. This was done for reasons of convenience; consideration of the
free energy is complicated by the effects of the constant or spin-independent term.
For completeness, we will show generally how the effects of the spin-independent
term together with those of irrelevant variables can be taken into account; detailed
calculations for the one-dimensional Ising model will not be carried out.
Denote the spin-independent part of the Hamiltonian by h, . If g, , g, , g, are
the nonlinear scaling fields discussed previously, then we have
fks 3g, , gz , 4,) = b-‘f@“go , bg, , b-h, 0) + h,(b), (6.29)
and choosing b = b* = ljgtJ2 we find

.f( gx 7 g, 9 gz 3 ho) = 81,” Y( g,/d’*, gz d”“> + Mb *I. (6.30)

To find the b-dependence of h(b), we follow the procedure developed in 6.2,

applying it to the linear Ising chain with z = 1.
It was convenient to express the recursion relation for the spin independent
term [see (3.16)] as
(wx)’ = (WXY y”/(x + VI2 (1 + yx)“, (6.31)
where h, = -4 In w. The fixed point is given by (xw)* = 1. On defining
u = 1 - xw, this may be written as
u’ = 2u + 4x + O(u2, x2, XJ) (b = 2). (6.32)
266 NELSON AND FISHER

To determine h&b*), it sufficesto find the b-dependenceof this recursion relation.

The corresponding scaling field g, is found to be

g, = u - 2x i- O(u2, x2, x7). (6.33)

Application of the method explained after Eq. (6.21) yields

u’(b) = bu + 2(b’ - b)x + O(u”, x2, xy). (6.34)

Expressing x and 3 in terms of g, and g, , we can find h,(b*) in terms of the scaling
fields, and determine the additional contribution to the free energy on the right-
hand side of (6.30).
The division of the free energy given by (6.30) into a piece which scales and
an extra term, is a general feature of the renormalization group [3, 321. One does
not expect such a decomposition to be unique; the nonlinear scaling fields will,
in fact, be seento be nonunique in Section 7. (There is a trivial nonuniqueness of
scaling fields, due to the possibility of constructing them around the different
tied points of a renormalization group problem, but we will refer in Section 7
to a nonuniquenessassociatedwith different Hamiltonianflows.)

7. SPIN RESCALING RENORMALIZATION GROUPS

In the dedecoration renormalization groups discussedso far, the spin resealing

factor c[%] (see Section 2) has been fixed at unity, consistent with the fact that
7 = 1 in one dimension. There is, however, a different approach due to Wilson [20],
which involves a variable spin resealingfactor. Of course, to find a low temperature
ferromagnetic fixed point for our one-dimensional models the spin resealing factor
must approach unity as T + 0 in zero field. Using Wilson’s approach, we will
generate a range of distinct, new renormalization groups for the Ising chain.

7.1. Generalized Renormalization Group Transformation

The idea of a renormalization group transformation is to eliminate a specified
fraction of the degreesof freedom of a system by performing some sort of partial
trace. The degrees of freedom eliminated could be alternate spins in an Ising
chain, or the high momentum modes of a Brillouin zone [l]. If the transformation
is chosen sensibly, fixed points describing critical behavior and the various critical
exponents can be found.
One way of writing a general renormalization group transformation is to define
a new reduced Hamiltonian *‘[s’] by

exp(s’[s’]) = Tr{B[s’, s] exp(s[s])j, (7.1)

 LOW DIMENSIONAL ISING SYSTEMS 267

where
Tr{ } = C ... (7.2)
Sl=fl s1\1Z*,{ >

for discrete Ising spins. The spins s’ replace the old spins s, and are a “thinned
out” set of degrees of freedom which describe the renormalized Hamiltonian 2’.
A minimal requirement that (7.1) describe a renormalization group transformation
is that the transformation preserve the partition function,

Tr’{exp(*‘[s’])} = Tr{exp(X[s])}, (7.3)

This leads to a condition on P[s’, s], namely,

Tr’(.!Y[s’, s]} = 1. (7.4)

The Wilson approach embraces the Kadanoff idea [25] of block spin variables
directly. With a block of two adjacent spins sBL and sZlc+l in the linear chain, we

S2(k-1) .S?!.
....?;ktl..........:
.. . ... . . . . . ..
-. :,.
'2(k+2)
. .. . ..

i..?( . . . . f..' i. .!( . . . .,! ,:. :\.'.( . . . . ,!.i : \

"'i"
\ I \ \
\/ '1 I/ \ ,I \
gj e @

"IL Sk slitl

FIG. 10. A generalized renormalization group transformation.

associate a new spin variable sit’ (see Fig. 10) which can assume the usual values
Sk’= i 1. We write the function P[s’, s] in the pair factored form
N/2
Pb’, $1= JJ m;; S2k 9 SPkll) (7.5)
k=l

with the condition

p(1 ; SZk , sZk+l> + p(- 1 ; s2k , S2kfl) = 1 (7.6)

to insure that (7.4) holds. Since P(s’; s1 , s2) is defined only on a space of eight
states, we may write it generally as a product of exponentials containing eight
independent constants [ 181,

W; s1 , x2) x exp(ps’ + w’s1 + q2s’s2 + IS’+S~ + p + &s, + q2s2 + EJ~s~). (7.7)

268 NELSON AND FISHER

The normalization condition eliminates four of these constants, and forces P

to take the form

P(s’; sl , s2) = exp(ps’ + qls’s, + q&s2 + rs’s,s,)/2 cosh(p + qls, + qts2 i- rs,s,).
(7.8)
The term exp(rs’.s,s,) in the numerator leads to a new Hamiltonian not in the
space of those with nearest neighbor interactions and magnetic fields, so we set
Y = 0 to obtain finally

PCs’; xl , s2> = exp(ps’ + w’s, + q2s’s2)/2 cosh(p + qlsl + q2s2). (7.9)

7.2. Realization of Transformations in Closed Form

In order to do the sum (7.1) we write the denominator of P(s’; s1 , s2) as

cosh(p + qlsl + q2s2) = g exp[Kw, + &s, + SL,s,l, (7.10)

where g, K, 6L,, and 6L, are given by the formula (3.3) with #(.sr , s2) =
cosh(p + q1 , s, + q2sZ). We can now express (7.1) as
ND
exp(s’[s’]) = Tr n g-l exp[(K - K) S2kS2k+l + CL - %) s2k + (L -6L2) SZk+l
1k=l
+ Ks2ks2k-1 + psk’ + qlsk’s2k + qas,‘s,,+,l (7.11)
1
For most values of p, q1 , and q2 , this transformation removes the Hamiltonian
from the original parameter space of nearest neighbor couplings and uniform
fields. If, however, we require

e4K = [cash 2p + cash 2(q, + q,)]/[cosh 2p + cash 2(q, - q2)], (7.12)

then R equals K, the summations can be performed analytically, and we stay

within the original parameter space. The recursion relations for K and L are then

e2L' = e2P*l/$2 , (7.13)

where

#I = eK cosh(ql + qn - 6L, - 6L2 + 2L) + e-K cosh(q, - q2 + h, - A,),

$2 = eK COW1 + q2 f SL, + 6L, - 2L) + e-K cosh(q, - q2 - SL, + a~,),
h = eK cosh(ql - q2 - 6-h - SL, + 2L) + e-K cosh(q, + q2 + 6~, - a~,),
$4 = eK COW, - q2 f SL, + SL, - 2L) + e-K cosh(q, + q2 - 6~, + 6~~).
(7.14)
 LOW DIMENSIONAL ISING SYSTEMS 269

The constraint (7.12) determines a two-dimensional surface of allowable

renormalization groups in the space indexed by p, q1 , and q2 .
When the parameter p is nonzero, we obtain an example of a nonlinear
renormalization group [24] (see Section 2). This is seen by examining the effect
of the transformation on the magnetization

Ok’)’ = &nh(p + wzk + q2szs+dj (7.15)

and the spin-spin correction function

(Sk’SO’) = <taMp + wzk + q2SPkfl) tanh(p + es0 + q2s1D. (7.16)

Here (*>’ denotes an expectation taken with respect to the primed spins. It is
easy to show that (7.15) and (7.16) do not have the simple scaling properties
associated with linear renormalization groups and discussed in Section 2. In fact,
a nonzero value of p breaks the symmetry of a zero field nearest neighbor Hamilto-
nian after one renormalization group iteration, and artificially shifts the magnetiza-
tion as indicated by (7.15). Because of these features, the parameter p appears to
be the Ising spin analogue of the spin “shift” used in calculations with continuous
spins below T, [8]. A linear Ising chain is always above T, = 0, so it is doubtful if
useful renormalization groups can be obtained from (7.13) with p # 0.

7.3. A Continuum of Renormalization Groups

With p = 0 we can readily analyze a continuum of renormalization groups
dependent on a single parameter. The condition (7.12) now reduces to

ezK = cosh(ql + q2)lcoWql - q2), (7.15)

and we imagine cosh(q, + q2) is chosen to fulfill this requirement with

cosh(q, .- q2) left as a free parameter. Using the usual variables x = e-4K, y = e--2L,
we obtain the e-dependent recursion relations

x82(1 + y)4 + (1 - P) 4v( 1 + >I)2

x’ = (1 + y2)2 + 4x( y3 + y) + 4X‘$2 - (1 - B”x)(l - y2)2 ’ (7*16)
1 + y2 + 2J?X - (1 - &x)1/2 (1 - JJ”)
(7.17)
y’ = 1 + y2 + 2yx + (1 - e‘%)l/2 (1 - v”) ’

where
6’ = llcosh(q, - 4, o<eGl. (7.18)

For 0 = 0, we recover the original dedecoration recursion relations (3.11) and

(3.12). The flows and fixed points for 0 = 1 (ql = q2) are shown in Fig. 11. There
270 NELSON AND FISHER

.
0 0.5 1
& = g2K
FIG. 11. Trajectories and fixed points describing the linear Ising chain under the action of a
spin resealing renormalization group (with 0 = 1). There are fixed points at (0, l), (1, l), and (0,O).

are fixed points located at (x*, JJ*) = (0, l), (1, 1) and (0, 0). The trajectories
differ markedly from those for the dedecoration group (see Fig. 2) at large values
of X. In particular, the “paramagnetic line” of tied points at x = 1 has vanished.
All groups with 0 -=c0 < 1 have the same general structure as depicted in Fig. 11.
Linearization of (7.16) and (7.17) about the ferromagnetic fixed point at (0, 1)
yields
A, = 4, A, = 2, (7.19)

A, = 2, Ar = 1, (7.20)

independently of 0. Thus the eigenvalues and ferromagnetic fixed point are inde-
pendent of the particular group chosen to describe the physics. This invariance
of physically significant eigenvalues is, of course expected on general grounds [33].
Although both the sic’ and the s, attain only the values fl, by examining the
way the magnetization transforms (an equivalent result derives from treating
the spin-spin correlation function) we can see that there is an effective spin resealing.
Analyzing (7.15) with p = 0 we find

(Sk’) = (1 - xP)l/Z (S‘&. (7.21)

 LOW DIMENSIONAL ISING SYSTEMS 271

The spin resealing factor (discussed in Section 2) is thus

c[&] = (1 - ?+)l/2, (7.22)
and it changes with each iteration step. However, for any value of 8, c approaches
unity as the ferromagnetic fixed point is approached, which is consistent with
rj = 1.
Just as in Section 6 one can now calculate e-dependent nonlinear scaling fields.
The results are

(7.21)

g,(e; 4 3 = Y + i 72 + (f + i e2).q + (a + & e2)xj2

+ (f+~e2+&8+2~+$3+ .. . . (7.22)

These expressions demonstrate explicitly that the nonlinear scaling fields are
nonunique; they are e-independent only to order x and 7 so that the linear scaling
fields are preserved. [For 0 = 0 they reduce to the previous results (6.21) as they
should.] Thus although the linear scaling fields have a definite physical significance,
the nonlinear fields cannot have a general significance. Indeed, the existence of
distinct renormalization groups with differing global Hamiltonian flows necessarily
implies distinct nonlinear fields.

APPENDIX: DIFFERENTIAL EQUATIONS FROM DISCRETE RECURSION RELATIONS:

TRUNCATED TETRAHEDRON MODEL

It is interesting to determine to what extent the differential approximation (5.8)

to the discrete recursion relation (5.7) for the truncated tetrahedron model is
valid. Consequently, we present here a systematic procedure for calculating
corrections to (5.9) and calculate the first correction term.
The recursion relation (5.7) may be written as
u = v + 3 - fi4 + O(65). (Al)
We will treat only the truncated recursion relation
iqI + I) = E(l) + 52(l). W)
Making the change of variable
w = v,/c, fio = E(O), (A3)
272 NELSON AND FISHER

we obtain
w(Z + 1) = w(Z) - 60 + tiJ[w(Z) + Q)]. (A4)
Approximating the function w(Z) by the solution of &v/d1 = --i$ as was done in
Section 5 gives w(Z) M I - i&Z. We obtain a correction to this result by substituting
the expression
w(E) = 1 - u,z + g(Z) (A5)
into (A4). The resulting recursion relation for g(Z) is

go + 1) = g(l) + &vl - v,z+ fro+ g(Ol. w

We now approximate g(Z) by the solution of dg/dl = Go2/(1 - V,Z+ a,) and obtain

w(Z) = 1 - V,Z- 6, ln(1 - i$) + O[G02ln(1 - U,Z)]. (A71

Clearly, one can continue this process of successive approximations indefinitely.
However, the next term produced by iteration of the procedure is of the same
order as the error introduced by truncating the original recursion relation. Thus,
higher order terms in the expansion (Al) have to be taken into account.
The extra term in (A7) shows the correlation length of the truncated tetrahedron
model varies as

5 --xpMln 2) expW/bJ) + (In WVb7T, c48)

which should be compared with (5.12).

ACKNOWLEDGMENTS

Conversations with Dr. T. L. Bell are greatly appreciated. We are also grateful for support
by the National Science Foundation, in part through the Materials Science Center at Cornell
University, and by the Ford Foundation through the Six Year Ph.D. Program.

REFERENCES

1. K. G. WILSON, Phys. Rev. B4 (1971), 3174-3183, 3184-3205, see also K. G. WILNN AND
J. KOGUT, Phys. Rept. 12C (I974), 77.
2. Some recent reviews are S.-K. MA, Rev, Mod. Phys. 45 (1973), 589-614; M. E. FISHER,
Proceedings of the International Conference on Magnetism, Moscow August 22-28, 1973;
Rev. Mod. Phys. 46 (1974), 597-616; F. J. WEGNER, Proceedings of the 8th Finish Summer
School in Theoretical Solid State Physics, August l-10, 1973.
3. F. J. WEGNER, Phys. Rev. B 5 (1972), 4529-4536.
 LOW DIMENSIONAL ISING SYSTEMS 273

4. K. G. WILSON AND M. E. FISHER, Phys. Rev. Letters 28 (1972), 240-243.

5. M. E. FISHER AND P. PFEUTY, Phys. Rev. B 6 (1972), 1889-1891.
6. F. J. WEGNER, Phys. Rev. B 6 (1972), 1891-1893.
7. K. G. WILSON, Phys. Rev. Letters 28 (1972), 548-551.
8. E. BRBZIN, D. J. WALLACE, AND K. G. WILSON, Phys. Rev. Letters 29 (1972), 591-594.
9. M. E. FISHER, S.-K. MA, AND B. G. NICKEL, Phys. Rev. Letters 29 (1972), 917-920.
10. E. BRI?ZIN AND D. J. WALLACF, Phys. Rev. B 7 (1973), 1967-1974.
11. G. A. BAKER, JR., Phys. Rev. B 5 (1972), 2622-2633; JR., AND G. R. GOLNER,
G. A. BAKER,
Phys. Rev. 31 (1973), 22-25.
Letters
12. F. J. DYSON, Commun. Math. Phys. 12 (1969), 91-107.
13. The hierarchial models have also been discussed from a renormalization group viewpoint by
Ya.G. Sinai and P.M. Bleher. See P.M. BLEHER AND YA.G. SINAI, Commun. Math. Phys. 33
(1973), 2342.
14. Brief accounts of aspects of this work have been presented at the Conference on the Renormal-
ization Group, Temple University, Philadelphia, May 29-31, 1973 (see the Proceedings
pp. 22-23) and at the 19th Ann. Conf. on Magnetism, D. R. Nelson and M. E. Fisher, AIP
Conf. Proc. No. 18, Magnetism and Magnetic Materials 1973 (AIP, 1974), pp. 888-890.
15. (a) In independent work announced at the Conference cn the Renormalization Group,
Temple University, May 29-31, 1973, (Proceedings p. 20), L. P. Kadanoff has studied the
linear Ising chain with nearest neighbor coupling by a simple renormalization group technique.
However, his aims and emphasis have been substantially different from ours. (b) In subsequent
work, R. E. PRANGE, Phys. Rev. A 9 (1974), 1711-1715, has similarly discussed the one-
dimensional Landau-Ginsburg or continuous spin-field model. (c)In work based on Ref. 1141,
S. KRINSKY AND D. FURMAN, Phys. Rev. Letters 32 (1974), 731-734, have used a dedecoration
group to discuss spin one Ising chains. See also Phys. Rev. B 11 (1975), 2602.
16. The one-dimensional Landau-Ginsburg models have been discussed analytically by R. Balian
and G. Toulouse [announced at the Nobel Symposium No. 24, Lerum, Sweden, June 12-16,
1973; see “Collective Properties of Physical Systems” (B. Lundqvist and S. Lundqvist, Eds.),
Academic Press, New York 1973, p. 451 for continuous values of n, the number of isotropic
spin components (n = 1 corresponds to Ising-like systems). For n < 1 they find a nonzero
transition temperature. See also R. BALIAN AND G. TOULOUSE, Amz. Phys. (N. Y.) 83 (1974),
28-63.
17. S. NAYA, Progr. Theoret. Phys. (Japan) 11 (1954), 53-62; I. SYOZI AND H. NAKANO, Prog.
Theoret. Phvs. (Japan) 13 (1955), 69-78.
18. M. E. FISHER, Phys. Rev. 113 (1959), 969-981.
19. J. W. ESSAM AND M. E. FISHER, 42 (1970), 272-288.
Rev. Mod. Phys.
20. K. G. WILSON, private communication.
21. Problems relating to the failure of hyperscaling relations, such as dv = 2 - oi, (see Ref. [22])
may lead one to consider more general formalisms. See M. E. Fisher in Proc. Conf. on the
Renormalization Group, Temple University, May 29-3 1, 1973.
22. For a general discussion of hyperscaling, etc., see M. E. FCSHER,Collective properties of
physical systems, in “Proceedings of the Nobel Symposium,” No. 24, (B. Lundqvist and
S. Lundqvist, Eds.), Academic Press, New York 1973.
23. (a) TH. NIEMEIJER AND J. M. J. VAN LEEUWEN, Phys. Reu. Letters 31 (1973), 1411-1414:
(b) PhJUiCQ 71 (1974), 17-40.
24. T. L. BELL AND K. G. WILSON, Phys. Reo. B 10 (1974), 3935.
25. C. DOMB, Advun. Phys. 9 (1960), Nos. 34, 35.
26. The Gaussian model was introduced by T. H. BERLIN AND M. KAC, Phys. Rev. 86 (1952),
821-835.
274 NELSON AND FISHER

27. For an interesting approach to extracting the Ising chain free energy from the dedecoration
renormalization group, see R. Priest, University of Maryland Preprint, 1974.
28. R. J. BAXTER, Phys. Rev. Letters26 (1971), 832-833, and Ann. Phys. (N.Y.) 70 (1972), 193-228.
29. L. P. KADANOFF AND F. J. WEGNER, Phys. Rev. B 4 (1971), 3989-3993.
30. F. J. WEGNER AND E. K. RTEDEL, Phys. Rev. B 7 (1973), 248-256.
31. L. P. KADANOFF, Physks 2 (1966), 263-272.
32. D. R. NELSON, Phys. Rev. B [in press].
33. F. J. WFGNER, J. Phys. C7 (1974), 2098.