VOLUME 69, NUMBER 11 PH YSICAL REVI E% LETTERS 14 SEPTEIUIBER 1992

Self-Organized Critical Forest-Fire Model

B. Drossel and F. Schwabl
Physik-Department der Technischen Universitat Miinchen, D-8046 Garching, Germany
(Received 30 June 1992)
A forest-fire model is introduced which contains a lightning probability f.This leads to a self-
organized critical state in the limitf 0 provided that the time scales of tree growth and burning down
of forest clusters are separated. %e derive scaling laws and calculate all critical exponents. The values
of the critical exponents are confirmed by computer simulations. For a two-dimensional system, we show
that the forest density in the critical state assumes its minimum possible value, i.e. , that energy dissipa-
tion is maximum.

PACS numbers: 05.40.+j, 05.45.+b, 05.70.Jk

Recently, Bak, Chen, and Tang [1] introduced a
forest-fire model which they assumed to show self-
organized critical behavior. Their forest-fire model is a
probabilistic cellular automaton defined on a d-dimen-
sional hypercubic lattice with L sites. In the beginning
each site is occupied by either a tree or a burning tree or
it is empty. The state of the system is parallely updated
by the following rules: (i) A burning tree becomes an
empty site. (ii) A green tree becomes a burning tree if at
steady state, a burning forest cluster must be so large
that trees grow at one end while the fire burns the other
end; i.e., the diameter of a burning forest cluster is pro-
portional to 1/p, and consequently the size of a fire front
is also proportional to 1/p.
The model becomes critical when a mechanism is in-
cluded that allows for small forest clusters to burn also.
We therefore introduce a "lightning parameter" f
fourth rule: (iv) A tree without a burning nearest neigh-
and a

least one of its nearest neighbors is burning. (iii) At an
empty site a tree grows with probability p.
If the system size is larger than the correlation length
of the fire, the model assumes a steady state with finite
fire density. Bak, Chen, and Tang found that the fire-fire
correlation function obeys a power law in two and three
dimensions. They concluded that the system is critical in
the limit p 0 where the fire correlation length diverges.
Bak coined the expression "self-organized criticality" to
describe the behavior of extended dissipative systems
which assume a critical steady state independent of the
initial state and without tuning of parameters to a special
value. The most prominent example of a self-organized
critical system is the so-called sandpile model [2].
Grassberger and Kantz [3], as well as Mossner, Dros-
sel, and Schwabl [41, performed computer simulations of
the forest-fire model with values of p smaller than those
used by Bak, Chen, and Tang. These simulations show
that the forest-fire model is not critical but instead it be-
comes more and more deterministic with decreasing p
and develops regular, spiral-shaped fire fronts. The size
of these spirals, as well as the distance between them, is
of the order 1/p. The temporal fire-fire correlation func-
tion oscillates regularly with a period proportional to 1/p.
By contrast, a critical system should contain fire fronts of
all sizes up to the correlation length, and their temporal
correlation function should show a power-law spectrum of
frequencies.
The reason that there are no fire fronts smaller than
1/p is the following: Trees that are next neighbors belong
to the same forest cluster. A tree only catches fire when
one of its neighbors burns. So a small forest cluster can-
not be ignited; therefore it grows until it becomes part of
a burning cluster. Since the fire burns constantly in the
bor becomes a burning tree during one time step with
probability f.In order to understand how criticality
arises in this extended forest-fire model, let us first simpli-
fy dynamics and assume that a whole forest cluster is
burned down instantaneously, i.e., during one time step
when one of its trees is struck by lightning. In this case
dynamics are invariant (except for a change of the time
scale) when fand p are multiplied by the same factor.
Then there is only one relevant parameter f/p in the sys-
tem. Let p be the mean overall forest density in the sys-
tem in the steady state. The average number of lightning
strokes in the system during t time steps is tfpL . The
average number of trees growing in the system during t
time steps is tp(1 p)L . Conseque— ntly, the average
number of trees destroyed by a lightning stroke is
s = (f/p) '(1 —p)/p.
In order to avoid finite-size eff'ects, the number of sites L
must be chosen much larger than the largest forest clus-
ter. For any finite value of f/p, the value of s is then in-
dependent of L. Equation (1) represents a power law
six: (f/p) for small values of f/p if limftp Qp( l.
This is the case for d ~
2 as we conclude from the follow-
1 for small values of f/p in d )
ing consideration: If the mean forest density p were near
2 dimensions, the largest
forest cluster would contain a finite percentage of all trees
in the system, and the average number of trees burned by
a lightening stroke would diverge in the limit L ~in
contradiction to (1). We therefore expect a critical point
in the limit f/p 0. The special case d =1 will be treat-
ed later.
Let us now return to the real forest-fire model where a
forest cluster is not burned down instantaneously but dor-

1992 The American Physical Society 1629

YOt UME 69, NUMBER 11 PH YSICAL REVIEW LETTERS 14 SEPTEM BER l 992

ing a finite time interval T(s) which depends on the num- with s, „ix (f/p) . Here, k and r are new critical ex-
ber of trees s in the forest cluster and will be determined ponents, and the cutofT functions C (x) and C(x) de-
below. Given a value of f/p, the forest dynamics of the crease monotonically from the value 1 for x(&1 to 0 for
system are essentially the same as in the case of instan- x»1. The logarithmic correction to N(s) for r =2
taneous burning down provided that p is so small that a guarantees a finite mean forest density in the limit f/p
forest cluster is burned down before trees grow at its 0. The exponent p is the fractal dimension of a forest
edge, i.e. , if I/p» T(s) Eq.uation (I) remains valid, and cluster. Equation (2) and the definition of the correlation
we expect critical behavior in the limit f/p 0 with length lead to the following scaling relations:
p((T '(s). Figure I shows the system in the steady X=vp, d=p(r —I). (3)
state with f/p =0. 1.
In order to show that the system really is critical in this Since p & d, the exponent r must obey i ~ 2, as already
limit, we calculate the scaling relations and the critical stated above. The mean number of trees destroyed by a
exponents which follow from the assumption of scale in- lightning stroke is
variance. We then confirm the values of the critical ex-
ponents by computer simulations. The radius R(s) of a sm,. „', 3&r &2f
forest cluster is the square root of the mean quadratic dis-
tance of the cluster members from their center of mass.
s, „/ln(s, „), r =2. (4)

We define the correlation length g by (=R(s), and the We already derived [Eq. (I)] s eep/f. Since
there is only
exponents v and z by gee (f/p) " and T(s) eel . The one diverging length scale in a critical system, one has
condition that forest clusters burn down rapidly then
sixs, „. Together with Eqs. (3) and (4), this leads to
the following values of the critical exponents [5]:
reads p(&(f/p)". Let N(s)ds denote the number of
clusters consisting of s trees in a system of a given size. r =2, p =d, v= I/d.
For brevity we introduce the moments m„= 1" s"N(s)ds f Since r =2, the power law for s,
„noted after Eq. (2)
of this non-normalized distribution function. Equation
(I) implies that N(s) does not decay faster than a power acquires a logarithmic correction factor and now reads
law in the limit f/p 0. Since the density of trees in the
system is finite, N(s) decays at least as fast as s . The
s,,„ee (f/p) '
ln(s, „)ee (p/f) In(p/f) .

cluster distribution N(s) therefore obeys a power law in Even for d = I, the scaling laws and critical exponents
the limit f/p 0. Under the scaling transformation derived are valid since the factor p/(I — p) entering Eq.
x x/b, f/p b' 'f/p, the normalized cluster distribu- (I) diverges only logarithmically, as we conclude from
tion N(s)ds/mo and the function R(s) must be invariant. the following considerations: The number of forest clus-
This implies the following scaling laws for N(s) and ters in the system is constant in the steady state. It in-
R (s): creases by 1 when an isolated tree grows, and it decreases
by I when lightning strikes a tree. Consequently Lp(1
C'(s/s, ,„), r & 2, —p) —2pmo=fLp. In addition, we have the following
'X'
cx
C'(s/s, „)In '(s, „), r =2, two relations that are valid in any dimension: L p=m]
and (p/f)L (I —p) =mz. From these equations we fi-
R (s ) ee s '~" C'(s/s, „), nally obtain p/(I —p) a:ln(sm, „). The scaling law for
N(s) again contains a logarithmic correction factor
aaa~amr
O
Dn
0
0 0
O
9 0 400
OOOOOD
—aaaqm-qra-~
0 0 000 0
D04000 oae
o oooo ooa o
0 OO
000
0
9 OO 9
—

o
— a
o oo
9
0
mr "Jr" --'
o 9
9 oo oo 0
000 0
0
90
0
9 009D 0 0 0 0 9000
n
0
00 ~
0
— 'a a '-aaa
0
9
0 00
00
ooooo
000 0
——
9
—-aaa
OOOO
o oo oooo
OO DOO O 0 0 00 DOO
0
00000
000 00
0 OOOO
ooooao
0 On
Do

9 o 000
OD
DII
ln '(sm, „); the power law for s,. „, however, is exact in
one dimension.
ooo 0 00
44 0
noooo o ooooo
0 0
oo o o oo
0
oo
9 o 0 ( 00
o
00
00
00
o 0
0 OO
0
oo
OOD
oo o o o
0 0 OOOO oooo
0 00 0 0000 0 0
o
0 OD 00 000 0 000 0 000 000 0 0 0 0 0 0 00 0
OO
0 OoOO 4 00 9009O On 0 O 9 0
0 0 0 0 OOO 0 DOOOOO O 0
On a 0 O OOOO O OO
JOO 00 9 on 0 0 0 0 0 0 0 0 0 0000000 00 400 0 0 00
DODO 9 0 09 O OO 00 O 000 0 0000440 000 0 00 00 00 000 0

The distribution n(T) of the lifetime of a fire also

oa 90 0 0 '&009 On 000 09 OOODDOOO 000 Do 0 9
&9 9 0 0 90
a ooonoo 0
o Doo o J o 9
' D
0
0 OOO ~ 0 000 0 94 0 00
OOOO 0
000 00OOOOOOO 0 0 00 00
4 0 0 00 0 oo
o
0 0 9 0
OOO 0
0 OOO 00009
0 0
0 000 0
Qn
J On 4 OOOO
09 99 00
000
0 9 9 oooa 0 J9 On no 00 Q O 0
9 0 0
0 00
000 00 4 0 0
OOP
00 0 OOOO 09
00 9 0 ~
0 oooo
OO O O 0 O0 Oo0 0 00 000OOO 0 00
0
00 0 0 900 0Do
D Q OD
000 0000000000
000 4 9 9 0 0 0000 0
9
JQQ
9 90900 ooo
00 0 DOOODO
900 0& n no ' 0'&
0090 Ono 0 0
9
9 D 0 000 090
QDJ DJ

obeys a po~er law: The probability that lightning strikes

&
0 0000 0 0 0 00 0 0 OO0 O 90 Ooo O OODOO
040 0 Q0
Do
QOOOOOOODO
00 0 9 00
QO I 9 0 9 0 000009 00 0
"I 0
oo
9 0
0 ooaooa o o oo o oooooo o oao
0 090000 9 o oo o oo a 4
0 0 00 00 0 0 0 0 0 0 04
0
0 o
0
o ooo o ooooooo
0000 0D'& 0 0 00 0 D
D o o oo o o
o
. 0 09 9 00 DODO 0 00
0
J 0 OD0 0 0 00 Oon Da ao 0 0 0 0009 90 00

a cluster of size s is proportional to sN(s) ~s '. The

O 9
090 0 00900 00 0 0099 00 D 9 D D DO 0 9 D 04
7 JO 0 000 000 0 & 0 09 DO 09 ~ 0 0 9 oo 4
'&0 0 F 90 0 0 OO 0 0 0 0 9
0 0 0 0 0000 OO4
OO Oo O O DD 0
090 909 99 0& 9 9 9 ~ 0 0009'& 09 000 0000DDD 9
0 On 0 9 &000 0
oo
9 9
00 9 000000000 0
0
~ 0 0
0 00 00 0
00 000 0 09
0 0
I& 9 0 9 90 0 00 0 0 0 00 00000 0 00 0 000 O 0 9000
0 QJQ 09
9 00 0 9 00 00
0 9 000000 0
OOOO
00 00000 00 0 0 9 0 00 0 0
0 00 09 00OOOO 000 00Q 00 009

cluster burns down in T(s) ee R (s) ~ s '

Dn In 9 0& OOOO 0 0 0 OOOO 0 Oo 0 000
00 0 Jn 0 0 000 0 00 9 000 0
0 OO
OOOO 000
0 DODO 00 D
0 00 0 OOOOOOOO 0 0 9 0
009 OOOO 0 0
90 999 000

time steps
0 OOD
Jo'& D& 000 ooooooaoo oo o o On D 0 0 D

'
0909
000 0 00909 9 OODOODDOOOO 0 9 oon o 0 0 0
0 00000 0 0 9 00 OO
00 OOOO oo0 0
OO DD 0 OOODOOOO oan O 9
D 0 0 00 OOOO 00 OO

with a= l. The mean lifetime

a 0 0 9
oo o 0
O 000 00 OOO
Jnn 00 0 0
oooo0 one noo 000 ooo0 0 0 ao o 09 00
moo &0 o 9
'&o
Oo»&oo oo o
000
ooo0000000 o o o D 00
9 0 ~

which implies n(T) ee T

90 90 0 0 00
Q
09 90 ODOD
DDDDD
090 00 0009 OOOODoo DD
ao o
o00 909 0
*09o QD ' 0 o 0 990
90 000 0
o o Oon 00 9 90
9 0
D 9 0 9 DD 000
0 9 0
09
0 0 0 90
0
09 0 90 OOOI»O 00 o o 0
D
00 9 noo 0 Do
OO
00 QDD 0
Do
00 0 00
'I 0
DD 9 0
anon o ~ ooo oon ooo oo o o o oooao oo 9 0 9 o Dnno oeo 0 o o 9 ~ o o*
~ 9 J 9 ooo 0 099 0 0 0 00 00 o DDD 0 o 909000 0 0 o 0 90 9 0 DDOOO 90
~ nnon Dno o ooo Doooaooo on& o a o o 0 o o 0 D 00 0 0 Onn D OD

of a fire is proportional to T(s) eel ee(f/p) ' which

D
0 9 Qo 000 000 &0 09 0 oo 9D*n ao 0 0* 00 9 09
DOO 90
0
DODD 90 O 0
on
0
9 * a on O 0 oo ooo 04
00
90OooOo
9 Don* 990
0 0
009 o o » * 9 9
DDODO
o 09
90900
*99*9
0 *99

leads to the dynamical critical exponent z = l.

o d
09 o DDDQ no 090o99
009 ooo o * 0 '
* O 0
9 00 0 0
900 o Do * 9oo 00
no o
9 0 o9
0
D DD
O Don ODD
9
0 *D
00 OOOOO
0 90 ODO
o oo oo o oo
Dn
0
9
o o
o D 0 090 D 9 0 9 900
0 OOOO 0
0
90 00 0 D 0 00
n
00 09 00 Dnn OOOO OOOO 0 Oo ~ 9
oo an * * 90 0 00 9
On OO 0 00 0 n * 9 0 00 OODO
OOOD 0 OQ 9

Knowing that the number of burning trees in a forest

o&n 0 ~ J&0 ooo D OOOOOOO 0 000 0
0 * 0 OOOOOOOO 000 D 0 9n*
ooo o
9 090
0 on o o
0 * I 0
o no o o o oo
DO OOO
0 OOOO D 0
DOOOD
DDD
OOOO
00 DODO 0
* 00000
0 on 09 o
909
0 o o
DO
0
9 9
9 000 0 00 09
9
0 990 00009
9000 ooo 0
0 009 90 9 9 9
aon
00* 0 9 o o oo 9 * 00 009 0 09
* *09090 9
0 0
00
o o
0
o ODO
00 Oo o Do 90
DOO
0OOOO0 90
00 000 DQD
00 0
0099
Dn
0 ODOOO o9 0

cluster t time steps after a lightning stroke is proportional

0 0 09
o Don
0o 090
o 9
on Doo *Dono oo o o &
0 Dn 9
Doonoo 9000*
00 OOOo aoo D 0 ooo0 0
onnn 0
D
oq
*0o099 0* 0 009 00 Dn 09 000 OOOD oOo OOo
Dn 0
0noon
O O * 9 0 OOOO D 000 *0000
o on o on o 99 ooooDo oo no
DDDOD
9o 099
* OOOO 9 00 Onn
»oo OOD On 0
000 OOOOD 0 o
0 9 OOOO
09 9 90 0 0 oD 0
000 noo o
0 0 9
DD

to t ', the Fourier transform of the fire-fire correlation

Oo 0

the.
00 0 0999000 ODODO 9 9 D
00 On *90
90 0 09 O DOOODD 0 o 0 0 0 09
*od
090 o 90 aoo noo o
Doa 000 00
o 0 o o *9
09 0
0
oo o 0 o & a*

f
*o o 000904 O* DJ
On Don
900 DDD9 o 0 09 oo *
0 0 9 0 OOOO 09 9
DODD
90 0090
0 9 00 0 0 d
o 9o0 ' ooo 0 *'
* o 000* *
* 9O 9
function can be shown to obey a power law ee
9
* 09999Onn 900o Dn o 0 ODOO 0 o 0
~ & '
ono DDDDo oo 0 0 0 090
a»»nno
O
000 000 90 9o 00
Jna 0 oo 090 0 o 0 0 * 00 00
0 0 o 0 o 'J 09

f
Q

f
o 90 0 0 00 o o o 0 9 0
o DD *o
9 099Dno O oo o o oao o 0 ooo an ono 0 00

+O(f +', . . . ,
0' 0 *a 09 o o 9 0 * 00 0* 00 099 000* Dn
90 90o 9 '& 9 DOD
9900 0 099 0 Da 9
OOOO
9 o o 00
0000 0009
009 00 0 9
9 9 0
0 0

) for small frequencies

0

In
DDO
09 oo 9000090 9 0 9 oOOo O 09 0 0 ooo 9 nd
. 0 . o 9 0 * oo 9 00 9
90 9
Jnnon
0 0 009
00 00 9QD
oo
o 000 00
OOOO
9 00 o 0 D D DD
90 Doo ODO o Ood
990000 0990 0 D 0000DD OO 0 o OOOOo
90 9 90 D d
0909 000 009 09 0 0 00
~ o 00
0 D 0 OOOO
D o 0 00
000 9 0090 o4
0 D
00 Ooo OO 09 0409 0 ~ 9 0
9 9 oon 900 * DD

sandpile model where the number of active sites is as-

JJJJJ 9 J J JJ JJ 0 Joa 4 a 4 aaa aa aa aa aaa aaa a a aaaa. a 9 Qo Dali

FIG. 1. Steady state of the forest-fire model for f/p =O. l sumed to be constant for the duration of an avalanche,
and d=2. Trees are represented by circles, and burning trees the corresponding power law is ~
mensions [6].
f in one and two di-
by crosses.

1630
VOLUME 69, NUMBER 11 PH YSICAL REVI EW LETTERS
Our computer simulations for d=1, 2, 3 confirm the
values of the critical exponents calculated above. For
d =2 and f/p =1/70, the cluster size distribution and the
cluster radius as functions of cluster size are shown in
I ig. 2. The slopes of the plots yield the exponents z =2
and v= 1/2.
We also determined the mean forest density p in the
steady state. For large values of f/p, the mean forest
density is small. It increases with decreasing f/p and ap-
proaches a constant value in the limit f/p 0. We
determined this value numerically and obtained p=0. 39
and p=0. 21 for 2 =2 and 3, respectively. In the follow-
ing, we show that the critical value p=0. 39 in two dimen-
sions is determined by an extremum principle: p assumes
its minimum possible value, i. e. , the fire destroys as much
forest as it can. Consider an area of the forest which is
very large compared to the lattice constant but very small
compared to the correlation length. The forest in this
area will grow until a fire passes through it which has
started somewhere in the forest around it (since the area
is small compared to the mean size of a forest cluster and
f/p is near 0, the lightning nearly never strikes it direct-
ly). Let hT denote the mean time interval between two
such fire fronts. Its value depends neither on the exact
size of the area nor on the precise value of f/p. The aver-
age forest density just before the fire enters the area is
pm, „, and is p;„just after the fire has passed. If the
value of hT were different from the actual value, the
values for p,
„and p;„and therefore the mean forest
density p would also be different. If hT were very large,
p, „would be near 1, and p;„near 0. With decreasing
hT, the value of p,
„would decrease, and the value of
p;„would therefore increase. For very small values of
hT, the difference between p~, „and p;„would also be
very small, and the value of p,
„would be just above the
percolation threshold 0. 59 for site percolation while the
value of p;„would be just below it. AT and p are related
top, „and p;„by the equations
pAT = 1n p=l—Pmax Pmin
Pmax
We still need a third relation between these four parame-
3.0
1.0
Z'.
CD
-1.0
-30 I I lksilii
0.0 1.0 2.0 3.0
log(s)
FIG. 2. Mean number of clusters and mean cluster radius as
functions of the cluster size for f/p = l/70 and d =2.
14 SEPTEMBER 1992
ters in order to determine all possible values of p and
therefore its minimum. Having in mind that the tree dis-
tribution just before the fire enters the area is stochastic
since trees grow stochastically, we determined p;„as a
function of p, „numerically by randomly filling a lattice
with trees up to the density p~, „ for diAerent values of
p, „) 0.59, and by counting the number of trees that be-
long to the largest forest cluster. This forest cluster con-
nects the edges of the lattice and would be destroyed by a
fire sweeping the lattice. From the number of trees that
do not belong to this cluster we obtained the density p;„.
Using the above relations, we plot p as a function of p, „
(Fig. 3). The minimum value of p is p=0. 39, and the
corresponding value of hT is AT=0. 91/p. The minimum
value of the mean forest density is just the value observed
in the critical state of our forest-fire model. We conclude
that the critical state is organized in such a way that the
number of growing trees and of burned trees is max-
imum, i.e. , that energy dissipation in the system is max-
imum. We expect that the extremum principle holds also
in higher dimensions.
Finally, we would like to comment on the physi-
cal significance of the two conditions f/p 0 and p
«(f/p)"', leading to the critical behavior of the forest-
fire model. Rewriting them in the form
(f/p) /
«p «f
we see that they describe a double separation of time
scales: The time in which a forest cluster burns down is
much shorter than the time in which a tree grows, which
again is much shorter than the time between two light-
ning occurrences at the same site. Separation of time
scales is quite frequent in nature, while the tuning of pa-
rameters to a certain finite value in nature only takes
place accidentally. Thus, the forest-fire model is critical
over a wide range of parameter values, and we expect
that its critical state should also be robust with respect to
slight modifications of the model rules, e.g. , another lat-
tice symmetry or fire spreading to next-nearest neighbors.
Let us compare the critical behavior of the forest-fire
model to the sandpile model: Here too, a separation of
time scales is required: Sand must be added slowly com-
0.60
i
0.55—
0.5 0.50—
33
V)
0.45—
-0.5
0.40
0.35 I s I
0.55 0.60 0.65 0.?0 0.?5 0.80
max
p
FIG. 3. Mean forest density as a function of maximum forest
density for d =2.
1631
VOLUME 69, NUMBER 11 PH YSICAL REV I EW LETTERS 14 SEPTEMBER 1992

pared to the lifetime of an avalanche. This corresponds [1] P. Bak, K. Chen, and C. Tang, Phys. Lett. A 147, 297
to our condition that a forest cluster burns down rapidly. (1990).
The power-law distribution of the avalanches in the sand- [2] P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. A 38,
364 (1988).
pile model is a consequence of the local conservation of
sand particles. In our forest-fire model, the power-law
[3] P. Grassberger and H. Kantz, J. Stat. Phys. 63, 685
(1991).
distribution of forest clusters is a consequence of a second
B. Drossel, and F. Schwabl (to be pub-
separation of time scales p '((f ', which guarantees [4] W. Mossner,
lished).
that a large amount of energy is deposited in the system [5] These critical exponents can also be derived from an in-
between two lightning occurrences and consequently a variance of the dynamic equations of motion for the forest
large number of trees is destroyed by a lightning stroke. density.
We thank W. Mossner and N. Knoblauch for writing [6] H. J. Jensen, K. Christensen, and H. Fogedby, Phys. Rev.
the simulation program. B 40, 7425 (1989).

1632