Z.

Physik B 36, 379-390 (1980)

Zeitschrift
f( r P h y s i k B
© by Springer-Verlag 1980

The Dynamical Renormalization Group for Master Equations:

Application to the Time Dependent Ginzburg-Landau Model
and the Two Dimensional Kinetic Ising Model
U. Deker* and F. Haake
Fachbereich Physik, Gesamthochschule Essen, Essen, West-Germany

Received August 17, 1979

We define a renormalization group transformation for the generator L of infinitesimal

time translations in near-critical systems. We argue that a convolutionless generalized
master equation must be used if formal ghost interactions of noninteracting subsystems
are to be avoided in L.
For the critical dynamics of Ginzburg-Landau models in 4 - e dimensions our transfor-
mation reproduces results previously obtained by other methods.
We apply a real-space version of our transformation to the two-dimensional kinetic Ising
model. Our findings generalize and partially contradict previous results. Due to the
nonexistence of a "canonical" form of the generator L any arbitrary value can be given to
the dynamical exponent z by an "appropriate" choice of the form of L, as long as the
real-space renormalization group transformation is implemented by finite-order cu-
mulant expansions.

1. Introduction

Renormalization group investigations of dynamical We here present a master equation technique for
critical phenomena have mostly been made by using implementing the dynamical RGT. The crucial step
Green's functions [1] or path integrals [23. Only in making the technique practical is to use a con-
rarely have master equation techniques been em- volutionless generalized master equation [11] rather
ployed 1-33. than one of the somewhat better known Zwanzig-
While for systems which have a continuous order Nakajima type [12]; it is indeed imperative to avoid
parameter the master equation method offers no par- time convolutions in the master equation since other-
ticular advantage over other methods, most other wise the independence of noninteracting subsystems
methods do not work for systems with a discrete would require a nonadditive Zwanzig kernel, i.e. for-
order parameter, like, e.g., the kinetic Ising model. mal ghost interactions in the master equation.
Usually, the microscopic discreteness of an order We shall sketch a master equation treatment of the
parameter is of no interest for discussions of critical critical dynamics of the Ginzburg-Landau model
behavior since such a discreteness is not observable with a nonconserved order parameter in a dimension
on the scales relevant to criticality. However, the slightly less than 4 1-13]. For this case the method
striking success of real-space renormalization group does not yield new results but proves convenient and,
transformations (RGT) in calculating static ex- to our taste, elegant.
ponents and even non-universal static parameters [4] Our main topic will be the discussion of the two-
has stimulated considerable interest in generali- dimensional kinetic Ising model, in which we shall
zations of the real-space R G T to the dynamical case use a real-space RGT. Here again, our master equa-
1,5-10]. tion technique yields explicit recursion formulas for
the parameters characterizing the dynamics with
* Present address: Institut ftir TheoretischePhysik der Universitgt amazingly little effort. However, for both a technical
Basel, 4056 Basel, Schweiz and a physically fundamental reason the investi-

0340-224X/80/0036/0379/$02.40
380 U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations

gation must remain inconclusive, i.e. does not pro- calculate all two-time equilibrium correlation func-
duce any reliable value for the dynamical critical tions. We have, for instance
exponent.
The unfortunate technical reason is that in explicitly (ai(t) aj(t')) =gi P(g, t - t ' [g') ~ Pst(g'). (2.1)
implementing the R G T we must rely on perturbation
Here and in the following we will use the summation
expansions in powers of the interaction strength of
convention that variables carrying a bar are to be
different cells of spins. Since the intercell bonds be-
summed or integrated over.
come the elementary bonds in the transformed sys-
We will assume the conditional probability P to obey
tem we then build up the transformed master equa-
a master equation of the form
tion by means of power series in these new elemen-
tary bonds. Consequently, we obtain fewer recursion 8
relations than there are parameters in the starting P(a, t l a')=L(a, ~) P(g, t] a') (2.2)
master equation.
The other, deeper reason for our negative result is or, in operator form,
that even for a system as simple as a kinetic Ising
g
model without conservation laws and without cou-
8t P(t) = LP(t) (2.3)
pling of the order parameter to other slow variables
there is no unique form for the generator of in-
finitesimal time translations. The general require- where the "Liouvillian" L is a linear operator.
ments available, i.e. probability conservation, detailed We may assume the set of variables al to be part of a
balance, and conservation of the independence of complete set of observables for a closed system with
noninteracting subsystems, do not suffice to fix the reversible dynamics. Then the operator L will obey
form of the generator and thus to make up for the the principle of detailed balance. This principle can
lack of information inherent in the perturbative treat- be written as the symmetry condition
ment of the RGT. We argue that the static R G T is
L(a, a') Pst(a')=L(a', a)P~t(a). (2.4)
exempt from the difficulties we encounter here since
there the requirement of the statistical independence Since the master equation (2.2) must conserve the
of noninteracting subsystems dictates the canonical, probability to find the system in some configuration,
i.e. exponential dependence of the equilibrium proba- P(6, t] a') = 1, we have the further restriction on L
bility density on the energy of spin configurations.
It follows from our unfortunately negative result that c(¢, .) = 0. (2.5)
previously reported values for the dynamical ex-
ponent z of the two dimensional kinetic Ising model It is an immediate consequence of detailed balance
do not express a physical property but rather reflect and probability conservation that the master equa-
special prejudices for the generator of infinitesimal tion (2.2) has P~t as a stationary solution, i.e.
time translations of that system.
LPst=O. (2.6)

We shall also assume that Pst is the only stationary

2. Definition of the Renormalization Group solution of Eq. (2.2) and is approached as t--+ c~, i.e.
We may specify the state of the system in terms of P(t) --+Ps, as t--, oo. (2.7)
the variables a~. These variables as well as the index i
may be discrete or continuous. In an equilibrium The static renormalization group transformation may
situation the statistical behavior of the system can be be defined by associating with the variables cri a
conveniently described by means of the static proba- smaller set of variables #,. The #, should have the
bility Pst(O') for finding the system in a configuration same algebraic properties as the a i and the indices v
and i should have a similar meaning. For example, v
In order to describe the dynamics of equilibrium should label wave vectors if i does; or, if i labels sites
fluctuations we may consider the two-time con- in a real-space lattice, v should do so, too, and the
ditional probability two lattices should be of the same structure. Further-
P(a, t - t'[ a') more, the length scale for the new set of variables
should be larger than that for the old variables by a
which is the probability for finding the configuration factor b > 1.
a at time t when the initial configuration at time t' To complete the definition of the static R G T we can
was a'. If we know /'st and P we can, in principle, use a matrix R(/,, a) and construct a transformed
U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations 381

static probability distribution as [4] The correlation function (2.14) then takes the form

Ps',(/~)= R(p, ~Y)Pst(a) (2.8) <~(t) #;~(o)> =,~,,(,y) P(< t l,~') ~,(~') P~t(,Y')
= <~bv(t) ~,(0)). (2.16)
or
e -w(") =R(#, g) e -~(e). (2.9) For an arbitrary transformation (2.12) the quantities
q~ and q~, will be complicated functions of the o.'s
If we require the matrix R(#, o.) to have the property without a clear physical meaning. Their correlation
R(fi, o.)= 1 (2.10) function (2.16) will, in general, not be simply related
to the basic correlation function (2.1). For the trans-
the normalization of Ps't is guaranteed and we can formation (2.12) to be useful in a description of
calculate the free energy F of the system by using critical behavior we need a proportionality of the q~
either P~t(o.) or Ps't(#) since - ~ . correlation function (2.16) to a o.i-o.; corre-
lation function with appropriate indices i and j for
e - r = e - u ' @ = e -me). (2.11)
large times and large spatial separations or wave-
In order to set up a dynamical RGT we proceed lengths characterizing the pairs of indices v, 2 and i, j.
analogously and associate with the conditional prob- The asymptotic property needed can be written
ability P(o.,tlo.') a conditional probability /5(#,tl#') schematically as
which allows the calculation of two-time correlation
<#v(t) #;,(0)> = <~v(t) ~ba(0)) ~ <o.i(t) o) (0)) (2.17)
functions of the variables #~. A convenient transfor-
mation involves the matrix R(#, o.) to relate the con-
figurations {#} and {o.} at time t similarly as in ~t --+ oO
for (wavelengths-~ oo"
Eq.(2.8) and another matrix R(o.',#') to relate the
initial configurations {o.'} and {#'}, In order to locate a fixed point for the RGT we must
rescale the time as
/5(#, tl#')= R(/~, ~Y)P(~7, t I~Y')/~(g', #'). (2.12)
t~t'=tb -z (2.18)
The matrices R and /~ cannot be completely inde-
pendent. Since both P and /5 are to be conditional and define the full dynamical RGT for P as
probabilities they must, at t=O, vanish unless the
n'(#, tl#')=/5(#, tb~l#,). (2.19)
respective primed and unprimed configurations are
equal. It follows that the matrix product R/~ must be If a physically acceptable fixed point is found for a
unity, i.e. particular value of z the #-#-correlation function
becomes identical, at the fixed point, to the o.-o.-
R(t,, e) R(~, #')= 17I 6(#~- #'~), (2.13)
v
correlation function apart from a constant factor and
the rescaling (2.18). It then follows that z is the
where a means a Kronecker symbol or a delta-
dynamical critical exponent which determines the
function according to whether the variables ~t are
time scale r for the decay of a fluctuation of wave
discrete or continuous, respectively.
vector k as ~ k -z.
For the interpretation of the renormalization group
In implementing the dynamical renormalization
transformation (2.12) we must know the relation be-
group transformation we could, by analogy to the
tween the two-time correlation function of the o.'s
usual treatment of the static case, parametrize the
defined in (2.1) and the corresponding function
conditional probabilities P and P' and extract re-
<fly(t) #~(0)> = fly/5(/7, tiff') ~T~.Ps',(fi'). (2.14) cursion relations for the parameters used from the
definition (2.12) [10]. We shall choose instead to
The latter quantity can be expressed as a two-time consider the equations of motion for P and P' and to
correlation function of certain functions of o.'s. If we construct recursion relations for the parameters which
insert Eqs.(2.8) and (2.12) in Eq.(2.14) we find the occur in these equations. The reason for this choice is
two o.-dependent quantities implicity correlated in that the equations of motion for P and P' can be
(2.14) to be formulated so as to give a description of the dy-
namics of the system which is local in time, while the
cl) (o.) = fi, R(fi, o.) (2.15) conditional probabilities P(t) and P'(t) refer to two
and times.
It is worth mentioning, at this point, that we have
4;).(o.) =/'st 1 ((9-)/~(O-, ff) 1~ Ps't(#-). quite intentionally based our considerations on the
382 U. Deker and F. Haake: Dynamical RenormalizationGroup for Master Equations

conditional probability P(a, t-t'la') rather than, say, the generalized master equation must allow for a
the two-time joint probability. The convenience of solution which factorizes with respect to the then
using the conditional probability lies in the initial independent subsystems; the "Liouvillian" should be a
condition sum of terms each of which refers to one subsystem.
These reasonable requirements are not compatible
P(G 0la') = I~ b (ai - a'i) (2.20) with an equation of motion of the Zwanzig-Nakajima
i
form.
which contains, in contrast to the two-time joint In order to prove the statement just made we may
probability at t=t', no information about the static consider the dynamics of the v-th subsystem, as de-
behavior of the system. Therefore, P(a, tlcr') is com- scribed by a Zwanzig-Nakajima type equation,
pletely determined by its equation of motion.
We would also like to point out that we can always P~(0 = ~ at' K d t - t ) " ' Pv(t-'). (2.23)
0
assume the system considered to be Markovian in
nature. Consequently, the conditional probability The product
P(a,t-t'lcr) can be used for calculating even non-
stationary two-time correlation functions. In order to P(t) = [ I ~(t) (2.24)
v
generalize Eq.(2.1) to the non-equilibrium case we
just have to replace Pst(o') with the appropriate single- then obeys the equation
time distribution at time t'. However, the smaller
system described by the variables #, which the RGT
associates with the original system, necessarily has
t
non-Markovian dynamics [14]. The quantity
=2~ dt, R v ( t _ t , ) Pv(t
- ,) ~(t) -1P(t) (2.25)
P'(#, t[#') therefore is a conditional probability with v 0
respect to the equilibrium correlation function (2.14)
only; it cannot be used to calculate non-stationary which obviously cannot be rewritten in the form
correlation functions, nor, of course, to calculate (2.21) with a Liouvillian which is a sum of terms each
equilibrium correlation functions involving three or of which refers to one subsystem alone.
more times [14]. It follows that if we calculate the kernel K(t) from a
The equation of motion for the transformed con- starting Liouvillian in which all physical interactions
ditional probability P(t) which is implied by the between different subsystems are switched off, we will
starting equation of motion (2.2) and the transfor- nonetheless find ghost interactions in K(t) which just
mation (2.12) can be written down in various equiva- express the trivial fact that an equation of the form
lent forms. Especially well-known is the Zwanzig- (2.21) with an "additive" K(t) cannot have the pro-
Nakajima generalized master equation which is an duct (2.24) as solution.
integrodifferential equation of the structure [12] Of course, no such difficulty is encountered when the
convolutionless generalized master equation (2.22) is
P(t) = ~ dr'/~(t - t')/5(¢). (2.21) used since this equation does not involve any explicit
0 nonlocality in time.
Another, convolutionless generalized master equation In order to construct explicitly the generalized
reads [11] Liouvillian we could use standard projector tech-
niques [11, 12]. It is slightly more convenient, howev-
er, to just rewrite the generalized master equation
~t P(t) = L(t) P(t). (2.22) (2.22) as the operator equation

Both generalized "Liouvillians" L(t) and K(t), can be L(t) = { ~ t t5(0 } •/5(0-1 (2.26)
expressed in terms of the operators L, R, and/~.
While the generalized master equations (2.21) and
If we then use the definitions (2.12) and (2.3) we find
(2.22) are equivalent as long as the respective
the following formal expression for L(t)
"Liouvillians" are not approximated, it is easy to see
that for our discussion of the dynamical RGT only L(t) = (RE e L'/~) (R eLt/~)-1 (2.27)
the convolutionless form (2.22) is useful.
In all practical implementations of the dynamical which can be evaluated, in principle, once the oper-
RGT the system considered is taken to consist of ators R,/~, and L are specified.
appropriate subsystems the interaction of which is treat- For our discussion of critical dynamics we must be
ed perturbatively. In zero order in the interaction able to iterate the transformation (2.27), i.e. allow for
U. Deker and F. Haake: Dynamical RenormalizationGroup for Master Equations 383

a starting Liouvillian which is already time depen- is described by the Hamiltonian

dent itself. If we designate the starting Liouvillian by
L(t) we may write the generalization of Eq. (2.27) as H=½Z(r+k)
k
+lU E 3(d)(kl+k2-/-k3+k4)o'klO-kzO-k30-k4 ' (3.1)
L(O={RL(t) (exp i dt'L(t'))+-R} kl .--k4
The dynamical behavior is assumed to be governed
.{R (exp i dt'L(t'))+l~}-l; (2.28) by a Fokker-Planck equation for the conditional
probability,
here (...)+ is a time ordered product in which an
0
operator referring to any given time appears to the ~P(t)=LP(t), (3.2)
left of all operators referring to earlier times.
In the applications to be treated in the present paper with the "Liouvillian"
it will be possible to split the Liouvillian L(t)
in a
time independent, i.e. "Markovian" zero order part 0 u ~
L 0 and a time dependent interaction part L~(t)
as L=~k e- ~a_k ell. (3.3)

L(t) = L o+ L 1(t). (2.29) We have, without loss of generality, absorbed a trans-

port coefficient in the time scale. The stationary
In such cases the transformation (2.28) can be written
solution of Eq. (3.2) is just the canonical distribution
down in a form especially well suited for expansions
in powers of Ll(t),
i.e. Pst = e - ri. (3.4)

L(t)={R eL°t[Lo+ LIl(t)] (exp i dt'L~(t'))+t~ } The customary renormalization group transfor-
mation involves an elimination of all variables a k
with wavevectors in the spherical shell
• ReL°tR+Re L°t -l+expjdt'LIl(t ') , (2.30)
o A/b<lkl<A, b>l. (3.5)
where L~(t) is the interaction part of the Liouvillian The corresponding matrix R(#, 0),
in the interaction picture
e ( # , o-) = H ~(#k -- O'k/b) (3.6)
Lq (t) = e-L°t Ll(t) eL°t. (2.31) [kI > A / b

is independent of the short-wavelength variables to

Fortunately, we will only need the first few terms of the
perturbation expansion of the right hand side of be eliminated and may be looked upon as a repre-
Eq. (2.30). sentation of the partial trace operation over these
In order to complete the definition of the R G T we short-wavelength variables,
still have to rescale the time as in (2.18) and write R = tr>. (3.7)
E (t') = L(t) bz. (2.32)
We will, in the following, use the operator (3.7)
instead of the matrix (3.6) and tacitly identify the
variables #k with the long-wavelength variables
3. The Ginzburg-Landau Model Near d = 4 O'k/b •
The static R G T can then be written as
The simplest possible model for which the critical
dynamics is well understood is the Ginzburg-Landau Pst = tr> Pst
model with a nonconserved order parameter in a
or, equivalently,
spatial dimension slightly less than four [13]. We
shall carry out the dynamical R G T presented in the e-/?=tr> e -H . (3.8)
last section for this system in order to demonstrate
the practicability of our method. We here use a tilde rather than a prime to design
The variables to be considered here are the spatial the transformed distribution and Hamiltonian since
Fourier components o-k of a real scalar field a(x).
The we will, in order to obtain the complete RGT, have
wave-vectors k are restricted to a d-dimensional to rescale the new spin variables in order to find a
sphere of radius A. The static behavior of the system fixed point of the R G T [13.
384 U. Deker and F. Haake: Dynamical RenormalizationGroup for Master Equations

The matrix /~(#, a) which we have defined in (2.12) is, according to (3.8), the first order contribution to
can be chosen as the transformed Hamiltonian H. In order to see that
the second order contribution,
/~(#, a) = P~t(#)-i R(a, p) Pst (o-). (3.9)
IYI2 = - ½<H~ (H 1 - (H~)o))o (3.19)
The corresponding operator is just
enters similarly in L we rewrite those terms in (3.15)
= e-n/e -fI (3.10) which stem from the expansion of/~ as
and obviously obeys the identity R/~ = 1. - ( L i eL°t(Ha - - (Hi)o) e-L°~)o
In order to carry out the dynamical R G T explicitly t
we treat the quartic part of the Hamiltonian as a
perturbation. The corresponding zero order term in
the Liouvillian is • ( L i eL°(t-e)(Hi _ (H1)o) e-L°(t-C))O. (3.20)

8 e_Ho 8 e+Ho The first term on the right hand side in the above
identity just is

= k~ {~k (r -~~2) 0"k-}-80.k~-0.. k}" (3.11) -- (L~(H 1- (H~) o)) o --- < 8 8/~r2 (3.21)
T 8~k 8%

It describes a Gaussian Markov process without in- so that the transformation (3.15) takes the form
teractions between variables with different wavevec-
tors. Consequently, we have the operator identity 8 e~ 8 e/i

tr> C Lot = e L°t tr> (3.12)

+ i dt'(Lx {eL°(t-t')(L1 -- (L1)0) e-L°(t-t')
0
where L 0 is identical in appearance with L 0 expect
that it contains only long-wavelength variables, + ~ eL°(t-t')(H1 _ ( H i ) o ) e - L°(~-c)}) o + O (U3). (3.22)
8 - 8 -
Lo=~--e -no e u°, (3.13) It is interesting to note that the first term in (3.22)
k 80"k 80"-k gives L(t) at the initial time, t = 0, exactly. This state-
< ment follows from P ( t = 0 ) = l and the definition
/~o =½ ~ (r +ka) auo- k. (3.14) (2.26),
k
By using the identity (3.12) we may write the transfor-
L(t = 0) = tr> L(t = O)
mation (2.30), up to corrections of order u a, < 8 e_n(~en)e_n+ a
=~tr> \Sa_ k
L(t)=Lo + (L1)o + i dt'(Li eL°(t-C)(L1-(L1)o)
0 < 8 e /? 8 d?" (3.23)
.e-Lo(t-r))o--(LieLot(Hi--(Hi)o)e-L°t)o , (3.15)
In order to further evaluate the remaining terms in
where the angular brackets denote the average over (3.22) we use the identities
short-wavelength variables with the zero-order ap- a{(z)-- e-Lo~ ake+Lo~
proximation of/~ as a weight, 8
= e-(~+ k2" ak +(r + k2) - a {e-(k +k2)~_ e+(r+k2,}
( - . . ) o = tr> (...)/~o = tr> (...) e -H°+R°. (3.16) 8a_k
and
The first two terms in Eq. (3.15) can be written in the 8 8
form e -L°t e+L°~=e+("+k2)~ (3.24)
8a k 8a k"
8 e_(Ho+Ht ) 8 e~°+& (3.17) The averages in (3.22) are thus reduced to moments
of the Gaussian ensemble /~0- They give rise to the
where three additive contributions to the transformed
Liouvillian which we denote by LA(t), L~(t), and
/tl = (Hi) o (3.18) Lc(O,
U. Deker and F. Haake: Dynamical RenormalizationGroup for Master Equations 385

L A(t) = - 6 u 2 i dr ~< A(k, ~) e(' + k2)~

0 k

8 e- rio 8 e~ ° (3.25)
8ak 8a_ k

LB(t) = -- 18 u ~ dr ~ B(K, ~c)e°'+k~)~

0 {k}

8 8 e~,O,
• 80._k 10"k20"/3(--'~)e-~/o 80.k 4
G
and Fig. 1. The diagrams arising in O(e2) of the dynamical RGT of the
Ginzburg-Landau model
L c ( t ) = - 9 u 2 dr ~ C(ki, z) e(~+k~)~
o (k}

. ~O'--kl O'k20"k3 0"/4(__,'C) 0"/5(__27) g-/go _ _ e Ho. variables a as

(~o_ k 6
t ' = t b -z, (7~k=ffk/bb--½(d+2-~1), (3.29)
In these formulas we have used the abbreviation a[,(t)
explained in (3.24). Moreover, the multiple wavevec-
E(t', {e'})=bzL(t, {e}),
tor sums are restricted by the conservation law ~ k i
i
iterate the transformation, and determine the expo-
=0. The coefficients A, B, and C occurring in (3.25) nents q and z so that a physically acceptable fixed
are wellknown from previous discussions of the criti- point is reached. We need not present the discussion,
cal dynamics of the Ginzburg-Landau model [-133. which, from this point on, completely parallels exis-
They can be expressed by means of the elementary ting ones [-1, 13]. The results obtained are wellknown.
correlation function The transformation yields a fixed point, in a dimen-
sion
G0(k, t)= <ak(t ) (7 k(0)> 0 = (r'-~ k2) -1 e -(r+k2)t (3.26)
d=4-e, s<O, (3.30)
as
> with
A(k, t ) = ~ Go(q1 , t) Go(q2 , t) G o ( k - q 1 - q 2 , t),
(q) r* = O (g), u* ~- 7[2 8/18, r/= 82/54

B({k}, t ) = ~ Go(q, t) Go(k 1 + k 2 - q , t), (3.27) Z = 2 +(6 ln~-- 1) t] + O(/33). (3.31)

q

C({k}, t)= O(Ik 1 + k 2 + k 3 l - A / b ) Go(k 1 + k 2 +k3, t),

4. The Kinetic Ising Model in Two Dimensions
where O is the unit step function. The usual diagram-
matic representations of the coefficients A, B, and C The two dimensional kinetic Ising model, in contrast
are depicted in Fig. 1. to the one dimensional one [-3, 6] does not allow for a
We should point out that, up to corrections of order rigorous treatment of its critical dynamics [,,7-10]. To
u 3 for L(t), the Hamiltonian /~o appearing in our present knowledge there is not even a systematic
Eqs. (3.25) may be replaced with the full transformed approximation scheme available for calculating the
Hamiltonian /t. It then becomes obvious that our fixed-point Liouvillian and the exponent z.
final result for the transformed Liouvillian, We shall illustrate the above statement for the tri-
angular lattice. We specify the R G T in analogy with
8 ~ 8 - the procedure used by Niemeijer and van Leeuwen
L ( t ) : F, 7 7 e - e/g+LA(t)+L~(t)+Lc(t), (3.28)
k UO k 8a_k [4-] to treat the static critical behavior. The lattice of
sites is regarded as a lattice of cells such that each
implies, as it must, the canonical distribution cell comprises three sites which are nearest neigh-
e x p ( - / ~ ) to be a stationary solution of the general- bours to each other (see Fig. 2). A new Ising variable
ized master equation (2.22). /~ is associated with each cell. In calculating the
The remaining steps of the renormalization group iterated Liouvillian E from the starting one, L, we
analysis require that we rescale the time t and the account for all intracell interactions exactly while we
386 U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations

The spin flip rates F~ and F~j are then bound to be

functions of the flipping spins and the local fields at
the respective sites which are defined as

B i = - o-i1(1 -Fi) H. (4.7)

We may choose

F i = a e -aiBi ¢p(Bi) (4.8)

and
Fij=(fl aiaj + Y) exp(-ai Bi-ajBj + K alaj)
• ½(1 + F~ Fj) z(B~)x(B;) (4.9)
with functions ¢p(Bi) and )~(Bi) which are even,
Fig. 2. Sites and cells in the triangular lattice qo(n~)= ¢p(-Ui) , z(B~)=z(-Bi) (4.10)

and take the value one for vanishing argument,

treat the intercell interactions perturbatively. It will
suffice to consider the corresponding perturbation cp(0) =Z(0)= 1, (4.11)
expansions up the first order terms.
The Liouvillian to be investigated by the RGT can but are arbitrary otherwise.
conveniently be written with the help of spin flip A simple special case of (4.8) is Glauber's choice [15]
operators F~, which are defined so as to reverse the F~= c~(1- tanh oh B~). Other authors have worked with
i-th spin when acting on an arbitrary function f({a}), the exponential F/= c¢exp ( - a i Bi) which corresponds
to ~0(Bi)= 1 [7, 9]. A certain formal analogy with the
F i f({a}) = f({a})]o,~ _~,. (4.1) Ginzburg-Landau Liouvillian (3.3) used in Sect. 3 for
continuous spins suggests the form
In order to describe multiple spin flips we will need
products of such operators referring to sets a of lal AiFi=~xAie-U Aie H with Ai=½(1-Fi);
sites (or cells), (la[ =number of sites contained in the
the latter form corresponds to ~o(B~)= cosh Bi.
set a),
No special choice for the functions (P(Bi) and z(B~),
Fa = H Fi" (4.2)
however, is physically distinguished from any other
iEa one. In principle, the renormalization group transfor-
mation should yield a unique fixed point from L* of
The most general Liouvillian has the form the Liouvillian. Unfortunately, however, when the
L = - Z ½(1 --Va) fla' (4.3) RGT is implemented by means of perturbation series
a in terms of the intercell interactions, the arbitrariness
of the starting Liouvillian L implies a certain arbi-
The transition rate F~ for the ]al-fold spin flip must trariness of E and thus of L* as long as the series are
obey detailed balance, i.e. the requirement truncated in finite orders. This can be seen as fol-
lows.
(1 -Fa) FaP~t= 0 (4.4)
In first order in the intercell interactions the static
for all sets a but is unrestricted otherwise. and dynamic RGT yields K'-expansions of the form
In a first order calculation, since correlations between
n.n.b. )
pairs of neighbouring sites will generate correlations
P~',(/~)=const 1 - K ' ~ ~/~z (4.12)
of pairs of neighbouring cells only, the Hamiltonian v,A
and the Liouvillian we must study will involve nea-
n.n.b.
rest neighbour bonds according to
n,n.b,
H = - inns, = - K ~, O"i O'j (4.5) F~'~(p) = fl'/G #~ + 7'. (4.14)
i,j
and Since the bond strength K in the original problem
n.rl.b. has been defined with respect to the exponent in the
L=-~½(1-F~)F~- ~ ½(1-F~Fj)Fw (4.6) canonical distribution P~t(a) we must read the first of
i i,j
U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations 387

these RGT equations, (4.12), as an expansion of a than the above argument suggests. Such hopes of
transformed canonical distribution P~',(#) ours have been frustrated by the following calcu-
= const e-w. lation which shows that any arbitrary value can be
No other form of the static distribution than the obtained for z by choosing an "appropriate" starting
canonical, i.e. exponential one is admissible for a Liouvillian.
renormalization group discussion of criticality for the In order to explicitly define the RGT we choose the
following wellknown reason. The renormalization operator R as a product over all cells [4]
group deals, when the transformation is iterated re-
peatedly, with smaller and smaller parts of the physi- R = [ I ½(1 + #v q~ (a)) (4.15)
v
cal system under investigation. The additivity of the
energy and the multiplicativity of P~t for noninteract- with
ing parts of the system then dictate the canonical
~v(O-) = N ( a v l + a v 2 + ava - fa~l a~2 or,3). (4.16)
form of Pst-
We may similarly argue that the Liouvillian L(t) Here the avi are the three original spins in the v-th
must be additive for noninteracting subsystems, as we cell and N and f are two parameters to be de-
have indeed done in Sect. 2. We cannot, however, termined presently. Note that ~ is the quantity
infer from such general considerations any further defined in (2.15). Similarly, we choose the operator/~
restrictions which would uniquely determine a as
"canonical" form for the transition rates Fa, i.e., es-
pecially, for the above functions q)(Bi) and 7~(B~). /~ = es,(a) es't (#)- * IF[ ½( 1 + #~ q3 (a)). (4.17)
v
Unfortunately, nothing can be inferred about these
functions from the recursion relations (4.13) and The identity R/~ = 1 implies certain restrictions on the
(4.14) either. The latter give, on their right hand sides, functions 43(a) the simplest and most important one
K'-expansions of the transition rates for arbitrary of which is
choices of (p(Bi) and )~(Bi).
The transformed parameters c(, fl', and 7' are func- 1 = ~(c7) 4,(g) P~t(g). (4.18)
tionals of the undetermined functions cp and 7~. It
A more complete discussion of the implications of
follows that the dynamical critical exponent z, which
the identity R/~= 1 will be represented in the Appen-
can be calculated from the recursion relations for the
dix.
parameters c~, fi, and 7, will depend on the choice we
Since we shall carry out the RGT up to first order in
adopt for qo and )~.
the intercell bonds we can satisfy the restriction (4.18)
The difficulty just outlined persists in higher orders of
by the choice
the perturbation series, because the number of pa-
rameters which we must admit but cannot determine 4~(a) = ~(a), (4.19)
increases as the order of the truncation of the per-
turbation series grows. The difficulty can be avoided if we choose the normalization parameter N accord-
only if the starting Liouvillian and Hamiltonian itself ing to
are also represented by a truncated bare K-expan- 1 = ~b,(~7)2 Ps(t°)(~). (4.20)
sion. In such a realization of the RGT all bonds,
whether intercellular or intracellular, are treated This equation reads explicitly
equally. The transformation must then yield an un-
ambiguous fixed point Liouvillian L* since terms not N = [ e 4 K ( 3 - - f ) z + 3(l + f ) z ] - } [ e 4 K + 3] ~. (4.21)
generated in E can be dropped in L. However, this
We still have to fix the parameter f The socalled
complete K-expansion is, although systematic, not
majority rule, f = 1, which has proven appropriate for
useful either since it leads to unacceptably bad results
the static RGT [4] is not the best possible choice for
for all critical exponents.
our purpose [9] since it would lead to a non Mar-
It is worth mentioning that other known approxima-
kovian transformed Liouvillian already in zero order
tion schemes for evaluating the RGT, like the cluster
in the intercell coupling.
expansion [4] and the technique of bond moving
In order to make the zero order part of the transfor-
[16] meet with similar failures since they, too, yield
med Liouvillian,
less structure in E than is, in general, present in L.
One might, of course, hope that the dynamical ex- L o (t) = (RL o e L°t No) (R e L°t/~o)- 1, (4.22)
ponent z depends only weakly on parameters con-
tained in ~0(B~) and )~(B~) and that the recursion time independent, i.e. Markovian, we proceed just as
relations (4.13, 4.14) are thus more useful in practice we did in dealing with the Ginzburg-Landau model
388 U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations

and impose on the operator R the analogue of the that our first-order R G T allows for a Markovian
identity (3.12), i.e. fixed point.
The construction of the recursion relations for the
R e L°t = eL°tR. (4.23) four parameters K, ~, fi, and 7 from the R G T (4.27) is
greatly facilitated if we use Eqs. (4.13, 4.14) to derive
By using the independence of different cells in zero
the identities
order we can infer from the identity (4.23) that the
quantities ~(o-) must be left-hand eigenfunctions of c~' = 5 (#v E #,)av- 3 (#v #4 E/2, #;),,
Lo,
c~t K t = 5(#v
1 t
#4 E),v (4.31)
• ~(8) Lo(~, a) = - 2 ~ ( a ) . (4.24)
t 1 t t
fi = 2 ( # v #-;~/~)av - - ( # v / ~ #X)av
This eigenvalue equation determines the parameter f 7t 1 / t
Among the two eigenfunctions that solve Eq. (4.24)
and have the form (4.16) the one with the smaller
in which the pairs of cells occuring are pairs of
eigenvalue 2 is the relevant one for our purpose since
nearest neighbours and the averaging prescription is
we should associate the slowest eigenmode ~ of the
y1 ~ for all cells. In evaluating the averages we use
v-th cell with the cell spin #v. /t=±l
In exploiting the condition (4.24) we observe that the Eq. (4.30) and consider the double-spin flip rates F~s
intracell contributions to the local field at the i-th site which involve spins in different cells as of first order
can take on the values 0, _ 2 K only. Consequently, in the intercell interaction. We then find the recursion
the two functions (p(B~) and x(B~) are, in zeroth order, relations [171
fully specified by Eq. (4.11) and the two parameters
K'=2KN2[e4rC+3]-z[(3-f)e4IC+l+fl z (4.32)
(p(2 K) = (p, x(ZK) = )6 (4.25)
b-z c( = 2 (4.33)
We then easily derive from (4.24) the following two
equations for f and the eigenvalue 2 b-z fl'= 8 fl N 2 [e 4K + 3] - 2 [(1 - f ) Z e2 t¢+ 1 + f l 2 (4.34)

2 =½(1 + f ) c~-¼(1- f ) ( e 2/¢ - 3 e -2K) c~~0 b-z 7' = 8 7 N2 [ e4/¢ + 3]- 2 [(1 _ / ) 2 Z eeK + (1 +f)21
+ ¼(eK+ 3 e - 3K) (X2 + 1) (fl + 7) (4.26) • [(1 - 4 N 2) Z eZK + 11. (4.35)

2 f = ~(1 + f ) a - ¼ ( 1 - f ) ( e 2 K + e-2K) a q~ These recursion relations, supplemented by the con-

+ ¼(eK- e - 3K) (g2 + 1) (fl + 7). (4.27) ditions (4.21, 4.26, 4.27) define the R G T in the param-
eter space we have chosen for the Liouvillian L.
We may deduce from the identity (4.23) and the Obviously and as already stated above, these equa-
principle of detailed balance (2.4) - which we assume tions leave two parameters undetermined. We take
to hold in zero order in the intercell interaction - the the open parameters to be q~ and X-
further useful identity In solving the above R G T equations we must de-
termine the exponent z such that a nontrivial physi-
L0/~o =-fro Lo- (4.28) cally acceptable fixed point c~*, fl*, 7*, K* exists. We
The analogue of (4.28) for the Ginzburg-Landau find two types of solutions:
model has been involved implicitly in the last section, (i) The first is a pure single-spin flip fixed point,
also• We here need it, together with (4.23), in order to
write the perturbation expansion of the transformed 3*=7*=0, ~* 4=0,
Liouvillian (2.28) or (2.30) in the form, up to the first K* = K*((p), z = z(q)). (4.36)
order term
As is shown in Fig. 3, the dynamical exponent z takes
L(t)= RLR. (4.29)
on arbitrary values when we allow the parameter (p
Up to the first order in the intercell coupling the full to vary freely.
R G T (2.24) now reads (ii) The second fixed point found contains double-
spin-flip processes,
b - ~E ( t b -~) = RL(t) 1~. (4.30)
7*=0, fi*+O, ~*+0,
Therefore, the transformed Liouvillian will be time
independent if the starting one is. We may conclude K* = K*(q), Z), z = z(q~, Z). (4.37)
U. Deker and F. Haake: Dynamical Renormalization Group for Master Equations 389

z In order to further discuss the stability problem we

have plotted the recursion relation of the ratio fi/~ for
3
the case ~ 0 = z = l [17] in Fig. 5. The intersections
with the straight line (fl/e)'=(fl/c 0 at (fl/a)*=0 and
2 (fi/e)* ~ - 0 . 1 9 correspond to the single-spin-flip and
1 the double-spin-flip fixed points, respectively. Ob-
viously, the single-spin-flip fixed point is the stable
l I -
I 2 one. This result contradicts [9].
The important albeit negative result of the present
Fig. 3. The dynamical exponent z for the single-spin-flip fixed point
investigation is the large indeterminacy of the ex-
ponent z. We should add that the uncertainty of z is
even larger than Figs. 3, 4 might suggest since our flip
rates (4.8, 4.9) are by no means the most general ones
possible. Even within the plausible assumption that F~
o and Fu should depend on the flipping spins and the
~0.6 respective local fields other choices than ours are
/ I possible and no less acceptable.
", -. ~ 1.5 It seems to us that a more successful realization of
the dynamical R G T for the two dimensional kinetic
Ising model would require a radically different ap-
proximation scheme for evaluating the transfor-
, 7--,:-
; mation (2.28) for L(O.
We would like to acknowledge helpful discussions with David
Fig. 4. The dynamical exponent z for the double-spin-flip fixed Nelson, Dieter Forster, and H. Thomas. We gratefully ac-
point. The labels at the full curves give the values of Z. Only below knowledge financial support by the Deutsche Forschungs-
the dashed curve the double-spin-flip fixed point is stable gemeinschaft.

R; Appendix
We here discuss a few formal aspects of the dynami-
cal RGT. Although we phrase the statements below
for the case of Ising variables, adand #~)=_+1,
attached to a lattice of sites i (and v), they hold equally
.I R~ or with slight modifications for other cases as well.
The general R G T can be characterized, instead of by
the matrices R and/~, by their moments

• ~(a) = fia R(fi, a)

(~a(O') = Pst 1 (o-)/~(o, fi)/~a Pst (fi), (A. 1)
Fig. 5. The recurslon relation for R# = fi/~
where a is an arbitrary set of sites v and

We have plotted the dependence of the exponent z on #a = H #v' (A.2)

vffa
the parameters (p and Z in Fig. 4.
We should point out that the undetermined param- The moments ~a(a) and ~a(a) have correlation func-
eters q0 and /~ not only determine the value of the tions which can be calculated as averages over the o-'s
exponent z but also govern the stability of the two or over the #%
fixed points found. In Fig. 4 the domain below the
<#a(O # b ( t ' ) 5 = < ~ a ( t ) ~ b ( t ' ) 5 , (A.3)
dashed curve corresponds to the range of stability of
the double-spin-flip fixed point (ii), while above that as we have shown in Sect. 2.
curve the single-spin-flip fixed point (i) is stable. It is In order to find the restrictions o n ~b a and q3, which
interesting to see that perturbations of 7 never lead to follow from R/~ = 1 we denote by
instability so that the parameter 7 could have been
set equal to zero form the outset. aob=awb\ac~b (A.4)
390 U. Deker and F. Haake: Dynamical Reuormalization Group for Master Equations

the set of sites which lie in a or b but not in both a 3. Kuramoto, Y.: Prog. Theor. Phys. 51, 1712 (1974); and the
and b. The algebraic properties of the Ising variables works on real-space techniques to be mentioned below
Deker, U., Haake, F.: Z. Physik B 35, 281 (1979)
may then be represented by 4. Niemeijer, T., van Leeuwen, J.M.J.: Physica 71 A, 17 (1974)
Kadanoff, L.P., Houghton, A.: Phys. Rev. B 11, 377 (1975)
/2 a • ]2b = # a o b . (A.5) 5. Kinzel, W.: Z. Physik B 29, 361 (1978)
Achiam, Y.: J. Physik A 11, 975 (1978)
The condition R/~=I now requires that the quan- 6. Deker, U.: Z. Physik B 32, 283 (1978)
tities #a and #a obey the algebraic identities (A.5) in 7. Achiam, Y., Kosterlitz, J.M.: Phys. Rev. Lett. 41, 128 (1978)
the mean, Achiam, Y.: J. Physik A 11, L 129 (1978)
Achiam, Y.: Phys. Rev. B 19, 376 (1979)
8. Suzuki, M., Sogo, K., Matsuba, I., Ikeda, H., Chikama, T.,
((J~a 4 b ) = (]~a(O) ~b(0)) = (#aob)
Takano, H.: Prog. Theor. Phys. 61, 864 (1979)
=(¢aob) 9. Mazenko, G.F., Nolan, MJ., Valls, O.T.: Phys. Rev. Lett. 41,
500 (1978) and to be published
= (~bnob). (A.6) 10. Ma, S.-K.: Phys. Rev. B 19, 4824 (1979)
11. Fulifiski, A., Kramarczyk, W.J.: Physica 39, 575 (1968)
If, especially, the matrix R(#, a) represents a coarse- Bates, Jr., C.W.: Phys. Rev. 188, 529 (1969)
graining transformation of the type Tokugama, M., Mori, H.: Prog. Theor. Phys. 55, 411 (1976)
Hashitsume, N., Shibata, F., Shingu, M.: J. Stat. Phys. 17, 155
R(#, a)= I~ 6(#~- ~b~(a)), (A.7) (1977)
Shibata, F., Takahashi, Y., Hashitsume, N.: J. Stat. Phys. 17,
the identity R/~ = 1 can be satisfied by the choice ~b~ 171 (1977)
H~inggi, P., Thomas, H., Grabert, H., Talkner, P.: J. Star. Phys.
= ~ since the 4~ are Ising variables themselves, in 18, 155 (1978)
that case. Examples of transformations of the kind 12. Zwanzig, R.: J. Chem. Phys. 33, 1338 (1960)
(A.7) are simple decimations and the transformation Nakajima, S.: Prog. Theor. Phys. 20, 948 (1958)
of Niemeijer and van Leeuwen [4]. Haake, F.: Springer Tracts Mod. Phys. 66, Berlin, Heidelberg,
New York: Springer 1973
The RGT we have treated in Sect. 4 also allows, 13. Halperin, B.I., Hohenberg, P.C., Ma, S.-K.: Phys. Rev. Lett. 29,
although it is not of the kind (A.7), for the choice 1548 (1972) and Phys. Rev. B 10, 139 (1974)
De Dominicis, C., Br~zin, E.: Phys. Rev. B 12, 4945 (1975), see
(/}a(O') = l ~ ~ v ( O ' ) = ~[~a(0") (A.S) also [2]
v 14. Hiinggi, P., Thomas, H.: Z. Physik B 26, 85 (1977)
since we carried the perturbation expansion up to Grabert, H., Talkner, P., H~inggi, P., Thomas, H.: Z. Physik B
29, 273 (1978); B 26, 389 (1977)
first order only. We could therefore satisfy the con- 15. Glauber, R.J.: J. Math. Phys. 4, 294 (1963)
ditions (A.6) cell by cell by requiring 16. Migdal, A.A.: Zh. Eksp. Teor. Fiz. 69, 1457 (1975)
Kadanoff, L.P.: Ann. Phys. (N.Y.) 100, 359 (1976)
((b~) = 1. (A.9) 17. The values q~=X= 1 correspond to the choices taken by Maz-
enko et al. in [9]. It should be noted that the meaning of our
A strong motivation for the choice (A.8) is, beyond parameters c~and fl is identical to that of the parameters e and
simplicity, that it preserves detailed balance. If de- /3 in [9] whereas the respective 7's differ slightly; the parameter
tailed balance holds for the original system, i.e. if 7 in [9] measures the strength of a term in L reading, in our
notation, Ai Aj exp ( - a i B i - a j B~ + K ai a j), where A i = ½(1 - F~);
because of the identity A I A j = ¼ { ( 1 - F ~ ) + ( 1 - F ~ ) - ( 1 - F ~ F ~ ) }
(~ba(t) ~ b ( t ' ) ) = (~bb(t) ~ a ( t ' ) ) , (A.10) such a term accounts for, besides double-spin-flip-processes,
somewhat peculiar single-spin-flips which we do not include in
we immediately conclude from (A.3) that property for our Liouvillian. The slight difference in ? disappears at the
the new system, fixed points where ?* =0 in all cases

(~ta(t) #b(t')) = (#b(t) #a(t')). (A.11)

Uli Deker
Fritz Haake
Fachbereich Physik
References Universit~it Essen
1. Hohenberg, P.C., Halperin, B.I.: Rev. Mod. Phys. 49, 435 Gesamthochschule
(1977) and references therein Postfach 6843
2. Bausch, R., Janssen, H.K., Wagner, H.: Z. Physik B24, 113 D-4300 Essen 1
(1976) Federal Republic of Germany

Responsible for advertisements: L. Siegel, Kur ffirstendamm 237, D- 1000 Berlin 15, Telephone: (030)8821031, Telex: 01-85411 - Springer-Verlag Berlin. Heidelberg. New York
Printers: Universitiitsdruckerei H. Stiirtz AG, Wtirzburg - Printed in Germany -- © by Springer-Verlag Berlin Heidelberg 1980