Forecasting characteristic earthquakes in a minimalist model

Abstract

Using error diagrams, we quantify the forecasting of characteristic-earthquake occurrence in a recently introduced minimalist model. Initially we connect the earthquake alarm at a fixed time after the ocurrence of a characteristic event. The evaluation of this strategy leads to a one-dimensional numerical exploration of the loss function. This first strategy is then refined by considering a classification of the seismic cycles of the model according to the presence, or not, of some factors related to the seismicity observed in the cycle. These factors, statistically speaking, enlarge or shorten the length of the cycles. The independent evaluation of the impact of these factors in the forecast process leads to two-dimensional numerical explorations. Finally, and as a third gradual step in the process of refinement, we combine these factors leading to a three-dimensional exploration. The final improvement in the loss function is about 8.5%.

Full text

Page 1

Nonlinear Processes in Geophysics (2003) 10: 565–571
NonlinearProcesses
in Geophysics
© European Geosciences Union 2003
Forecasting characteristic earthquakes in a minimalist model
M. V´ azquez-Prada1,´A. Gonz´ alez2, J. B. G´ omez2, and A. F. Pacheco1
1Departamento de F´isica Te´ orica and BIFI, Universidad de Zaragoza, Pedro Cerbuna 12, 50009 Zaragoza, Spain
2Departamento de Ciencias de la Tierra, Universidad de Zaragoza, Pedro Cerbuna 12, 50009 Zaragoza, Spain
Abstract. Using error diagrams, we quantify the forecast-
ing of characteristic-earthquake occurrence in a recently in-
troduced minimalist model. Initially we connect the earth-
quake alarm at a fixed time after the ocurrence of a charac-
teristic event. The evaluation of this strategy leads to a one-
dimensional numerical exploration of the loss function. This
first strategy is then refined by considering a classification of
the seismic cycles of the model according to the presence,
or not, of some factors related to the seismicity observed in
the cycle. These factors, statistically speaking, enlarge or
shorten the length of the cycles. The independent evaluation
of the impact of these factors in the forecast process leads
to two-dimensional numerical explorations. Finally, and as a
third gradual step in the process of refinement, we combine
these factors leading to a three-dimensional exploration. The
final improvement in the loss function is about 8.5%.
1Introduction
The earthquake process in seismic faults is a very complex
natural phenomenon that present geophysics, in spite of its
considerable efforts, has not yet been able to put into a sound
and satisfactory status. However, in the crucial field of earth-
quake prediction, recent years have witnessed significant ad-
vances. For recent thorough reviews dealing with this issue,
see Keilis-Borok (2002); Keilis-Borok and Soloviev (2003),
and references therein, in particular chapter four by Kos-
sobokov and Shebalin. See also Lomnitz (1994). The in-
troduction of new concepts coming from modern statistical
physics seems to add some light and put some order into
the intrinsic complexity of the lithosphere and its dynamics.
Thus, for example, references to critical phenomena, dynam-
ical systems, hierarchical systems, fractals, self-organized
criticality and self-organized complexity are now found very
frequently in geophysical literature (Turcotte, 2000; Sor-
nette, 2000; Gabrielov et al., 1999, 2000). Hopefully, this
Correspondence to: A. F. Pacheco (amalio@unizar.es)
conceptual framework will prove its usefulness sooner better
than later.
We have recently presented a simple statistical model of
the cellular-automaton type which produces an earthquake
spectrum similar to the characteristic earthquake behaviour
of some seismic faults (V´ azquez-Prada et al., 2002). The
largestearthquakesonafaultorfaultsegment(theeventsthat
break its complete length) are usually termed characteristic
(Schwartz and Coppersmith, 1984; Wesnousky, 1994; Dah-
men et al., 1998). For this reason, in the minimalist model
the event of maximum size is called the characteristic one.
Our model is inspired by the concept of asperity, i.e. by
the presence of a particularly strong element in the system
which actually controls its relaxation. This model presents
some notable properties, some of which will be reviewed in
Sect. 2. In Sect. 3, an algebraic procedure for the exact cal-
culation of the probability distribution of the time of return
of the characteristic earthquake is presented. The purpose of
this paper is to quantify the forecasting of the characteristic
earthquake occurrence in this model, using seismicity func-
tions, which are observable, but not stress functions (Ben-
Zion et al., 2003), which are not. In Sect. 4, we construct an
error diagram (Molchan, 1997; Newman and Turcotte, 2002)
based on the time elapsed since the occurrence of the last
characteristic event. This permits a first assessment of the
degree of predictability. In Sect. 5, we propose a general
strategy of classification of the seismic cycles which, ade-
quately exploited in this model, allows a refinement of the
forecasts. Finally, in Sect. 6 we present the conclusions.
2Some properties of the model
In the minimalist model (V´ azquez-Prada et al., 2002), a one-
dimensional vertical array of length N is considered. The
ordered levels of the array are labelled by an integer index
i that runs upwards from 1 to N. This system performs two
basic functions: it is loaded by receiving stress particles in its
various levels and unloaded by emitting groups of particles

Page 2

566M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes
through the first level i = 1. These emissions that relax the
system are called earthquakes.
These two functions (loading and unloading) proceed us-
ing the following four rules:
i In each time unit, one particle arrives at the system.
ii All the positions in the array, from i = 1 to i = N,
have the same probability of receiving the new particle.
When a position receives a particle we say that it is oc-
cupied.
iii If a new particle comes to a level which is already occu-
pied, this particle has no effect on the system. Thus, a
given position i can only be either non-occupied when
noparticlehascometoit, oroccupiedwhenoneormore
particles have come to it.
iv The level i = 1 is special. When a particle goes to this
first position a relaxation event occurs. Then, if all the
successive levels from i = 1 up to i = k are occupied,
and the position k + 1 is empty, the effect of the re-
laxation (or earthquake) is to unload all the levels from
i = 1 up to i = k. Hence, the size of this relaxation is k,
andtheremaininglevelsi > k maintaintheiroccupancy
intact.
Therefore, the size of the earthquakes in this model range
from 1 up to N, being the event of k = N the characteristic
one. Note that the three first rules of this model are exactly
those of the forest-fires model (Drossel and Schwabl, 1992).
Our model has no parameter and, at a given time, the state of
the system is specified by stating which of the (i > 1)N −
1 ordered levels are occupied. Each one of these possible
occupation states corresponds to a stable configuration of the
system, and therefore the total number of configurations is
2(N−1). Thesementioned2(N−1)stableconfigurationscanbe
considered as the states of a finite, irreducible and aperiodic
Markov chain with a unique stationary distribution (Durrett,
1999).
The evolution rules of the model produce an earthquake
size-frequency relation, pk, that is shown in Fig. 1a, where
the results for N = 10, N = 100, and N = 1000 are su-
perimposed. Note that this spectrum has a distribution of the
characteristic-earthquake type: it exhibits a power-law rela-
tionship for small events, an excess of maximal (characteris-
tic) events, and very few of the intermediate size. Besides,
the three superimposed curves of probability are coincident.
The result for the probability of return of the characteristic
earthquake, P(n), is shown in Fig. 1b for N = 20. Here n
represents the time elapsed since the last characteristic event.
During an initial time interval 1 ≤ n < N, P(n) is null, then
it grows to a maximum and then finally declines asymptot-
ically to 0. (In Sect. 4, P(n) for N = 20, will be usually
denoted as curve a.) In Sect. 3, we explain a general alge-
braic method for the exact computation of P(n).
The configurations of the model are classified into groups
according to the number of levels, j, that are occupied
Fig. 1. (a) Probability of occurrence of earthquakes of size k. Note
that the simulations corresponding to N equal to 10, 100, and 1000
are superimposed. (b) For N = 20, the probability of return of the
characteristic earthquake as a function of the time elapsed since the
last event, n. (c) Time evolution of the state of occupation in a sys-
tem of size N = 100. Note that after each characteristic event that
completely depletes the system, there follows the corresponding re-
coveryuptoahighlevelofoccupancy, andthenthesystemtypically
presents a plateau previous to the next characteristic earthquake.

Page 3

M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes567
(0 ≤ j ≤ N − 1). Using the Markov-chain theory or pro-
ducing simulations (V´ azquez-Prada et al., 2002), one easily
observes that in this model the system resides often in the
configurations of maximum occupancy, i.e. in j = N − 2
and j = N − 1.
This last property can be observed in Fig. 1c, where we
have represented, for N = 100, the time evolution of the
level of occupancy, j, in an interval long enough to observe
theoccurrenceofseveralcharacteristicearthquakes. Thetyp-
ical pattern after a total depletion is a gradual recovery of j
up to a new high level of occupancy. Once there, the sys-
tem typically presents a plateau before the next characteristic
earthquake. Especially during the ascending recoveries, the
level of occupancy j suffers small falls corresponding to the
occurrence of rather small earthquakes, that in this model are
abundant. Of course, one also observes that occasionally j
falls in a significant way corresponding to the occurrence of
a N > k ≥ N/2 intermediate earthquake.
Due to the fact that this model is not critical, it is reason-
able to consider it as an example of self-organized complex-
ity (Gabrielov et al., 1999).
3Algebraic approach to P(n)
The function P(n), for a minimalist system of size N, is ob-
tained from the Markov matrix of the system, M, following
the following three steps: (i) The element of the last row and
first column of M is changed by a 0. After this pruning, the
matrixwillbecalledM?. (ii)ThenewmatrixM?ismultiplied
by itself n−1 times to obtain M?(n−1)and the element of the
first row, last column of this matrix is identified. (iii) P(n) is
the product of this selected matrix element times 1/N.
The whys of this recipe are explained in V´ azquez-Prada et
al. (2002), where P(n) for N = 2 is explicitly obtained by
mere inspection. The result is
P(n) =n − 1
2n
,N = 2.
(1)
The explicit form of P(n) for larger values of N, can be
achieved by exploiting the Jordan decomposition of M?,
M?= Q J Q−1,
and hence,
(2)
M?n−1= Q Jn−1Q−1.
The matrix J is formed by “Jordan blocks” in the diag-
onal positions, i.e. by square matrices whose elements are
zero except for those on the principal diagonal, which are all
equal, and those on the first superdiagonal, which are equal
to unity. Thus, the task of obtaining an arbitrary power of J is
simple because, as said, each Jordan block is the sum of two
conmuting matrices: one is a constant times the unity ma-
trix, and the other is nilpotent. Therefore, in the computation
of any arbitrary power of J, each block is independent and
the corresponding Newton bynomial formula can be applied.
(3)
As an example, we now present the calculation of the case
N = 3. In this case,


which is decomposed as
3M?=


1 1 1 0
0 2 0 1
1 0 1 1
0 0 0 2



,
(4)


Therefore,


1 1 1 0
0 2 0 1
1 0 1 1
0 0 0 2

=






−1 1 1
0 0 2 −1
1 1 0
0 0 0
2
0
2








0 0 0 0
0 2 1 0
0 0 2 1
0 0 0 2








−1/2
1/2 −1/4 1/2
0
0
1/4 1/2 −3/8
3/8
1/4
1/2
1/20
00

.(5)


Jn−1=




0
0 2n−1(n − 1)2n−21/2(n − 1)(n − 2)2n−3
002n−1
000
000
(n − 1)2n−2
2n−1




?1
3
?n−1
,(6)
and from Eq. (3)
M?n−1
1,4
=2n
32(n − 2)(n + 5)
?1
3
?n−1
.
(7)
Thus, finally,
P(n) =
?2
3
?n(n − 2)(n + 5)
32
,N = 3.
(8)
One could optimistically guess that P(n), for an arbitrary
N, can be deduced from the systematics observed in the pre-
vious low-N cases. This is disproved by the following for-
mula, which is the result of P(n) for N = 4.
?1
?
P(n) =
4
?n?
−3
−13
16+7n
4
−n2
2
+n3
32
+3n
16+
7n
324+
n2
108+
n3
2599
??
,N = 4.
(9)
Although it is not apparent, this formula, as it should, van-
ishes for n = 3. As in Eqs. (1) and (8), P(n) in Eq. (9) is
adequately normalized:
∞
?
n=N
P(n) = 1.
(10)
4Error diagram for the forecasting of the characteris-
tic earthquake
In the following paragraphs, we will stick to a model of size
N = 20 to make the pertinent comparisons. This size is big
enough for our purposes here, and small enough to obtain
good statistics in the simulations.
For n = 20 the mean value of P(n) is
∞
?
?n? =
i=20
P(i)i = 121.05,
(11)

Page 4

568M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes
the standard deviation is
?
i=20
and the skew of the distribution is
σ =
∞
?
P(i)(i − ?n?)2
?1/2
= 55.21,
(12)
γ =
1
σ3
∞
?
i=20
P(i) · (i − ?n?)3= −0.10.
(13)
Now we enter into the matter of forecasting. As in any
optimization strategy, we will try to achieve simultaneously
the most in a property called A and the least in a property
called B, these two purposes being contradictory in them-
selves. Here A is the (successful) forecast of the character-
istic earthquakes produced in the system. Our desire is to
forecast as many as possible, or ideally, all of them. B is the
total amount of time that the earthquake alarm is switched
on during the forecasting process. As is obvious, our desire
would be that this time were a minimum. The maximization
of A is equivalent to the minimization of an A?that represents
the fraction of unsuccessful forecasts.
Thus, in practice, our goal in this paper is to obtain simul-
taneously a minimum value for the two following functions,
feand fa. The first represents the fraction of unsuccessful
forecasts, or fraction of failures; the second represents the
fraction of alarm time. These two functions, in this first one-
dimensional strategy of forecasting, are dependent only on
the value of n, that is, the time elapsed since the last main
event, and to which the alarm is connected. Using the func-
tion P(n) previously defined, they read as follows:
fe(n) =
n
?
?∞
n?=1
P(n?),
(14)
fa(n) =
n?=nP(n?) (n?− n)
?∞
These two functions are plotted in Fig. 2a. By eliminat-
ing n between fe(n) and fa(n), we obtain Fig. 2b, which
is the standard form of representing the so-called error dia-
gram. The diagonal straight line would represent the result
of a random forecasting strategy. The curved line is the result
of this model.
Error diagrams were introduced in earthquake forecast-
ing by Molchan who contributed with rigorous mathemati-
cal analysis to the optimization of the earthquake prediction
strategies (Molchan, 1997). In his papers Molchan used τ
and n to represent the alarm fraction and the error fraction
respectively; and put τ in the horizontal axis.
To fix ideas, it is convenient to define a so-called loss
function, L, which expresses the trade-off between costs and
benefits in the forecasting (Keilis-Borok, 2002). Among all
the possible loss functions, we will choose the simple linear
function
n?=0P(n?) n?
.
(15)
L = fa+ fe.
(16)
Fig. 2. For N = 20, (a) Fraction of failures to predict, fe, fraction
of alarm time, fa, and loss function L = fa+feas a function of n.
(b) Error diagram for characteristic event forecasts based on n. The
diagonal line would correspond to a random strategy.
L(n) is also drawn if Fig. 2a. The position na= 66 provides
the minimum value of L(n). L(na) = 0.578. Note that na
does not coincide either with the n that maximizes P(n), or
with ?n?.
5 Improving the forecasts
In Sect. 4 we adopted the strategy of connecting the alarm at
a fixed time, n, after the occurrence of a characteristic event.
The evaluation of this strategy leads to the conclusion that
for n = na = 66, the loss function has a minimum value
L(na) = 0.578. The question now is: Can we think up other
strategies that render better results? To answer this question,
we now return to our previous comments on Fig. 1.

Page 5

M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes569
If we define a medium-size earthquake as an event with a
size between N/2 and N −1, i.e. N > k ≥ N/2, by observ-
ing the graphs in Fig. 1, one is led to the conclusion that in
this model the occurrence of a medium-size earthquake is not
frequent but when it actually takes place, the time of return
of the characteristic quake in that cycle is increased.
This qualitative perception can be substantiated by numer-
ically obtaining the probability of having cycles where no
medium-size earthquake occurs, i.e. k < N/2. This infor-
mation is completed by the distribution of cycles where the
condition N > k ≥ N/2 does occur. These two distribu-
tions are shown in Fig. 3a as lines b and c respectively. Here,
line a represents the total distribution of the times of return
of the characteristic earthquake in this model (the same as
plotted in Fig. 1b). Note that, as it should, the distribution
a covers both distributions b and c. The mean time ?n? for
the three distributions is ?n?a= 121.05, ?n?b= 107.57 and
?n?c= 166.84. The fraction of cycles under b is 0.77 and the
fraction under c is 0.23. A splitting of this type, in which the
a distribution separates into b and c, will be denoted hence-
forth as a = b ⊕ c.
Tocheckifa = b⊕c ispotentiallyusefulforourpurposes,
we will now analyze independently these two sets of cycles,
b and c, with the method used in Sect. 4 for curve a. The
result is the following: the best working n for dealing with
the cycles under distribution b is nb= 60. And with respect
to the cycles belonging to the distribution under c, the best n
is nc= 124.
Therefore, we will now study again the whole set of cy-
cles, i.e. those under a, by means of a retarding strategy,
which is based on the splitting a = b ⊕ c. We will adopt the
following steps: in any cycle, we will wait until an n, named
nret1(which is near to nb), before taking any decision. If
no medium earthquake has occurred so far, then the alarm
is connected at nret1. If, on the contrary, a medium quake
has occurred before nret1, then we move the alarm to nret2
(which is close to nc). This notation nret1and nret2comes
from the retarding strategy that we are exploring now. This
two-dimensional strategy is implemented by varying nret1
and nret2looking for the best value of L. This is illustrated
in Fig. 4a. The best option is nret1= 61 and nret2= 101,
with L(nret1,nret2) = 0.549.
Now we look for a similar property that can classify the
cycles from another point of view. This new property con-
sists in identifying the cycles where the sum of the sizes of
all the earthquakes before the characteristic one is less than
N/2. This condition will be represented by SUM < N/2.
ThereasonforthischoiceisthatifSUM < N/2, thesystem,
statistically speaking, tends to reach more rapidly the config-
urations of maximum occupancy, j = N −2 and j = N −1,
and the time of return of the characteristic quake in that cycle
tends to be smaller (see Fig. 1c). In Fig. 3b, line a represents,
as in Fig. 3a, the distribution of return intervals of the char-
acteristic earthquake for all the cycles of the model. And
lines f and g represent, respectively, the separation of line
a according to the fulfilment, or not, of the SUM < N/2
condition, a = f ⊕ g. The mean value of the f and g dis-
Fig. 3. (a) For N = 20. Line a is the distribution of return times of
the characteristic earthquake as a function of the time elapsed since
the last event, n. Line b corresponds to the distribution of cycles
where no medium-size earthquake occurs. Line c corresponds to
cycles with medium-size earthquakes. Curves b and c constitute
the splitting of curve a according to whether this retarding effect is
fulfilled or not. (b) Lines f and g, represent the separation of the a
distribution according to whether the advancing effect is fulfilled or
not.
tributions is ?n?f = 88.78 and ?n?g = 151.69 respectively.
The fraction of events under the f and g lines is 0.37 .and
0.63 respectively.
This second splitting of the whole set of cycles in the
model, a = f ⊕ g, can be used as an advancing strategy
in parallel to what we did with the retarding strategy. Thus
the independent analysis of curve f leads to nf = 60, and
the similar analysis of curve g leads to ng= 90.
We will now study again the whole set of cycles (under
a) by means of the advancing strategy, which is based on
the splitting a = f ⊕ g. Therefore, we proceed as follows:
In any cycle, we wait until nadv1(which is close to nf) be-
fore taking any decision. If the condition SUM < N/2 has
been fulfilled, then the alarm is connected at n = nadv1. If,
on the contrary, this condition has not been fulfilled, then
we move the alarm to nadv2(which is close to ng). This

Page 6

570M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes
Fig. 4. (a) For N = 20. Results of the two-dimensional strategy based on the splitting of curve a according to whether the retarding effect
is fulfilled or not. Values of L varying nret1and nret2. Minimum value of L = 0.549 for nret1= 61 and nret2= 101. (b) Results of the
two-dimensional strategy based on the splitting of curve a according to whether the advancing effect is fulfilled or not. Values of L varying
nadv1and nadv2. Minimum value of L = 0.537 for nadv1= 61 and nadv2= 90.
two-dimensional strategy is implemented by varying nadv1
and nadv2looking for the lowest L. This is illustrated in
Fig. 4b. The search for the best option leads to nadv1= 61
and nadv2= 90, with L(nadv1,nadv2) = 0.537. This value
of L is slightly better than that obtained using the retarding
strategy.
Inspired by these results, we will now analize the possibil-
ities of a mixed strategy which contains conceptual elements
of the two partial strategies discussed so far. Here we will
explore a 3-dimensional grid of points (n1,n2,n3) looking
for the minimization of L. The first coordinate, n1,will be
explored in the neighbourhood of nadv1, the second coordi-
nate n2in the neighbourhood of nadv2, and finally n3near
nret2. The two succesive key decisions to be taken are:
i In any cycle, we wait until n = n1. If SUM < N/2 IS
fulfilled, we connect the alarm at n1and leave it there.
If at n = n1, SUM < N/2 is NOT fulfilled, we move
the alarm to n2. And,
ii (We are now at n2). If no medium-size event has oc-
curred between n1and n2, we leave the alarm connected
at n2. If, on the contrary, one or more medium-size
events have occurred in this interval, then we move the
alarm to n3.
The search for the triplet (n1,n2,n3) that makes L min-
imum is illustrated in Fig. 5.
(n1= 61,n2= 84,n3= 104), and there, L = 0.528. This
is the best result obtained in this work.
Thus, the improvement obtained in L, when passing from
L(na) to L(n1,n2,n3) is around ? 8.5%.
The result corresponds to
Fig. 5. For N = 20. Illustration of the three-dimensional strategy.
For n1= 61, L-constant level-curves are plotted. The minimum
value of L is 0.528 for n1= 61, n2= 84 and n3= 104.
6Conclusions
In this paper, we have analyzed the behaviour of the min-
imalist model in relation to a quantitative assesment of the
forecasting of its successive characteristic earthquakes. We
havechosenasimplelossfunction, L = fa+fe. Ourfirsttry,
basedonaone-dimensionalsearchinn, producesaminimum
resultofLaround0.578. ThiswasillustratedinFig.2a. With
the aim of improving the forecasts, we then explored two
modes of a common strategy that divides the probability dis-
tribution of the time of return of the characteristic earthquake
into two distinct distributions. The first mode consists in us-
ing the occurrence of intermediate-magnitude earthquakes as

Page 7

M. V´ azquez-Prada et al.: Forecasting characteristic earthquakes571
a sign that the characteristic earthquake would likely return
at a time later than usual in that cycle. This is based on the
fact that medium-size events significantly deplete the load in
the system and its recovery induces a retardation. This effect
takes place in any system of the sand-pile type. The exploita-
tion of this idea leads to a two-dimensional search that finally
renders an L value around 0.549 (Fig. 4a). The second idea
consists in using the fact that a significant absence of small
earthquakes during a sizeable lapse of time in the cycle is a
sign of imminence of the next characteristic event, or at least
of a shortening of its period of return. This strategy is similar
to the old wisdom in seismology that links a steady absence
of earthquakes in a fault with the increase in the risk of oc-
currence of a big event. The exploration of this idea proceeds
similarly to what we did with the retarding strategy: this also
leads to a two-dimensional search. It renders a minimum L
around 0.537. (Fig. 4b).
Finally, a mixed strategy that tries to incorporate the in-
formation acquired is implemented by means of a three-
dimensional search, and provides a value of L = 0.528. The
identification of the three optimum parameters is illustrated
in Fig. 5.
It is important to remark that the information we have used
in our forecasts is based only in the observed systematics
of earthquake occurrence in the model, i.e. only seismicity
functions have been used. Thus, for example, in Sect. 5 we
have not used the state of occupancy of the system j, which
would have given much more accurate predictions. In real
life, the use of this information would be equivalent to know-
ing, in real time, the value of the stress level and the failure
threshold at any point in a fault.
Acknowledgements. We are grateful to the two Referees of the first
version of this paper for their thorough and helpful reviews. This
work was supported by the project BFM2002-01798 of the Spanish
Ministry of Science. Miguel V´ azquez-Prada and´Alvaro Gonz´ alez
are respectively supported by the PhD research grants B037/2001
(funded by the Autonomous Government of Arag´ on and the Euro-
pean Social Fund) and AP2002-1347 (funded by the Spanish Min-
istry of Education).
References
Ben-Zion, Y., Eneva, M., and Liu, Y.: Large earthquake cycles and
intermittent criticality on heterogeneous faults due to evolv-
ing stress and seismicity, J. Geophys. Res., 108 (B6), 2307, doi:
10.1029/2002JB002121, 2003.
Dahmen, K., Ertas, D., and Ben-Zion, Y.: Gutenberg-Richter and
characteristic behavior in simple mean-field models of hetero-
geneous faults, Phys. Rev., E58, 1444–1501, 1998.
Drossel, B. and Schwabl, F.: Self-organized critical forest-fire
model, Phys. Rev. Lett., 69, 1629–1632, 1992.
Durrett, R.: Essentials of Stochastic Processes, Chapter 1, Springer,
1999.
Gabrielov, A., Newman, W. I., and Turcotte, D. L.: Exactly soluble
hierarchical clustering model: inverse cascades, self-similarity
and scaling, Phys. Rev., E60, 5293–5300, 1999.
Gabrielov, A., Zaliapin, I., Newman, W., and Keilis-Borok, V. I.:
Colliding cascades model for earthquake prediction, Geophys. J.
Int., 143, 427–437, 2000.
Keilis-Borok, V.:Earthquake prediction: state-of-the-art and
emerging possibilities, Annu. Rev. Earth Planet. Sci., 30, 1–33,
2002.
Keilis-Borok, V. I. and Soloviev, A. A. (Eds): Nonlinear Dynamics
of the Lithosphere and Earthquake Prediction, Springer, 2003.
Lomnitz, C.: Fundamentals of Earthquake Prediction, J. Wiley &
Sons, 1994.
Molchan, G.M.: Earthquake predictionasa decision-making prob-
lem, Pure. Appl. Geophys., 149, 233–247, 1997.
Newman W. I. and Turcotte, D. L.: A simple model for the earth-
quake cycle combining self-organized complexity with critical
point behavior, Nonlin. Proces. Geophys., 9, 453–461, 2002.
Schwartz, D. P. and Coppersmith, K. J.: Fault behaviour and char-
acteristic earthquakes: Examples from the Wasatch and San An-
dreas Fault Zones, J. Geophys. Res., 89, 5681–5698, 1984.
Sornette, D.: CriticalPhenomena inNaturalSciences, SpringerVer-
lag, Berlin, Germany, 2000.
Turcotte, D. L.: Fractals and Chaos in Geology and Geophysics,
2ndEdit., Cambridge Univ. Press, 2000.
V´ azquez-Prada, M., Gonz´ alez, ´A., G´ omez, J. B., and Pacheco,
A. F.: A Minimalist model of characteristic earthquakes, Non-
lin. Proces. Geophys., 9, 513–519, 2002.
Wesnousky, S. G.: The Gutenberg-Richter or characteristic-
earthquake distribution, wich is it?, Bull. Seism. Soc. Am., 84,
1940–1959, 1994.