RETAS: a restricted ETAS model inspired by Baths law

D. Gospodinov
Geophysical Institute of the Bulgarian Academy of Sciences, Centralna Posta PK 258, 4000 Plovdiv, Bulgaria drago pld@[Link]

R. Rotondi
C.N.R. - Istituto di Matematica Applicata e Tecnologie Informatiche Via Bassini 15, 20133 Milano, Italy reni@[Link]

Abstract A version of the restricted trigger model is used to analyse the temporal behaviour of some aftershock sequences. The conditional intensity function of the model is similar to that of the Epidemic Type Aftershock-Sequence (ETAS) model with the restriction that only the aftershocks of magnitude bigger than or equal to some threshold Mtr can trigger secondary events. For this reason we have named the model restricted epidemic type aftershocksequence (RETAS) model. Varying the triggering threshold we examine the variants of the RETAS model which range from the Modied Omori Formula (MOF) to the ETAS model, including such models as limit cases. 1. Introduction HE most widely used models to describe clusters of events are: the modied Omori law [5] and the ETAS model [2]. The rst assumes that all the events in an aftershock sequence are triggered by the stress eld change due to the main shock, are conditionally independent and follow a nonstationary Poisson process; the second extends the capability of generating secondary events to every aftershock of the sequence. Between these two extreme cases there is a range of different versions of trigger models which differ in the nature of the primary events. According to Baths law the difference between the main shock magnitude and the magnitude of the strongest aftershock is constant, on average 1.2 for Richter and 1.4 for Utsu [4]. By extending this principle to the subsequences generated by primary events as well, we get that the difference between the weakest primary event and the weakest event in the aftershock sequence must be at least 1.2. This suggests choosing the primary events among those of magnitude larger than a suitable threshold M tr . The aim of our work is to examine the class of trigger models obtained by varying Mtr between the cut-off M0 and the maximum magnitude Mmax. At rst we have simulated two data sets according to the ETAS and RETAS model respectively, and we checked that the true model was identied using the Akaike criterion. Then we have applied the RETAS model to the analysis of two aftershock sequences with different characteristics: the rst is formed by the events which followed the strong earthquake of M = 7.8 occurred in Kresna, SW Bulgaria, in 1904, the second includes three main shocks and a large swarm of minor shocks following the quake of 26 September 1997 in Umbria-Marche region, central Italy.

where t is the time elapsed from the occurrence of the main shock; whereas P is formed by the entire sequence in the ETAS model [2] (t|Ht) = +
ti<t

The analysis performed on the relaxation process indicates that we have a typical aftershock sequence in which the process is just controlled by the stress eld change caused by the main shock.
(a)
70 60 Cumulative number 50 Magnitude 40 30 20 10 0 0 500 Time [days] 1000 7.8 6.6 5.4 4.2 3.0 Cumulative number 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 7.8 6.6 5.4 4.2 3.0 Magnitude

Ki (t ti + c)p .

(b)

where the term Ki may have the explicit formulation Ki = K0 e(MiM0) (1) In intermediate situations P may be composed by events randomly distributed [6], or identied before estimation [2] and possibly, in their turn, triggered by events belonging to the same subset P, occurred previously [4]. In this work we have examined a trigger model in which the primary events are those having magnitude larger than or equal to a threshold Mtr , but, contrary to [4], the parameters Km obey the restriction (1); hence the conditional intensity function of the model turns out to be (t|Ht) = +
ti < t Mi Mtr

As the non-monotonic trend of the curve (Fig. 3(a)) of the AIC values is similar to that obtained by analysing a sequence simulated by a RETAS model, and as that curve has a local minimum in correspondence of Mtr = 4.6, we repeat the analysis on the subset of the aftershocks sequence with cut-off magnitude M 0 = 3.6. The minimum AIC value (Fig. 3(b)) is achieved for Mtr = 5.0; this means that, by raising the threshold magnitude, secondary clusters appear in the so reduced Umbria-Marche sequence, generated by the six strongest events. The RETAS model we have proposed can be applied for modelling compound seismic sequences in regions of complex tectonic structure with highly fractured earth crust.
(a)
80 70 Cumulative number Cumulative number 60 Magnitude 50 40 30 20 10 6 5 4 3 0 5 10 15 Time [days] 20 25 30 80 70 60 Magnitude 6 5 4 3 0 20 40 60 80 Frequencylinearized time 50 40 30 20 10 0

Frequencylinearized time

Figure 2: Observed (circles) and estimated (continuous line) cumulative number of events versus (a) the ordinary and (b) transformed time. The plug-in estimate corresponds to the RETAS = MOF model (Mtr = 7.8). The dotted lines in (b) are the twofold standard deviations of the residual point process.

(b)

K0 e(MiM0) (t ti + c)p

4. Case study: Umbria-Marche region, central Italy . (2)

The idea of a gap between the magnitude Mtr of the triggering event and that of the largest event generated is borrowed by Baths law, but we have not xed the size of this gap, even better by varying Mtr we have examined all the models between the MOF and the ETAS model on the basis of the Akaike criterion. 3. Case study: Kresna region, SW Bulgaria N April 4, 1904 an earthquake of M = 7.8 struck the region and triggered an aftershock sequence which continued for a long period. The sequence considered is constituted by 60 earthquakes of magnitude M 4.0 covering a time period of 1230 days (Fig. 1(a)). Fig. 1(b) shows the AIC values obtained by tting the class of the RETAS models to the sequence. The best model corresponds to Mtr = 7.8, that is, to the MOF. The AIC value is not sensitive to the Mtr variations because of the very strong dependence of the aftershock frequency on the magnitude of the triggering event as indicated by the values of the parameter estimates, high for and small for K.

HE sequence is characterized by the occurrence of six earthquakes of magnitude larger than 5 in a period of 20 days; it contains 508 events with M 2.9, covering the period from September 26 up to November 3, 1997. First we have tted the class of RETAS models to this data set. The AIC values obtained (Fig. 3(a)) show that the ETAS model is the best; according to Ogata [3] the small value of the parameter estimate, = 0.89, classies the sequence as a swarm.
(a)
1600 1650 1700 AIC AIC 1750 1800 1850 1900 1950 3.0 3.5 4.0 4.5 5.0 5.5 3.5 4.0 4.5 5.0 5.5 6.0 Triggering magnitude level Triggering magnitude level 40 30 20 10 0 10 20 30

Figure 5: Cumulative number and magnitude of the 1997 aftershocks in Umbria-Marche region, central Italy, for cut-off magnitude M0 = 3.6; observed (blue) and estimated (red) cumulative number of events versus (a) the ordinary and (b) transformed time. The plug-in estimate corresponds to the RETAS model with Mtr = 5.0. The dotted lines in (b) are the twofold standard deviations of the residual point process.

(b)

The algorithm
The algorithm we have applied is composed of two stages. The rst consists in the following steps: i. order the set D = {Mi}N in increasing way and build the set of different i=1 magnitude levels {M agk }K obtained by removing from D the repeated valk=1 ues, so that M agk < M agk+1, k = 1, . . . , K 1, k M agk D, and i k: Mi = M agk ; ii. set k = 0; iii. k k + 1, Mtr = M agk ; iv. given the conditional intensity function (2), maximize log () and evaluate AICk ; v. if k = 1, then set kbest = 1 and AICbest = AIC1; otherwise, if AICk < AICbest, then set kbest = k and AICbest = AICk ; vi. if k < K then go to (iii), otherwise stop. The nal output, expressed by the value k , identies the best model: ETAS best if kbest = 1, MOF if kbest = K and RETAS with triggering magnitude M agkbest if 1 < kbest < K. If the ETAS model comes out to be the best model, it is interesting to examine whether particular correlations between the strongest aftershocks and the remaining seismicity emerge by raising the threshold M0. In this second stage the analysis can be carried out throught the following steps: i. plot the set of pairs {(M agk , AICk )}K ; k=1 ii. if this set shows a monotonic trend, then stop; otherwise nd the index of the local minimum, lets say j, set Mtr = M agj and M0 Mtr 1, and go back to the rst stage considering the subset formed by the events of size exceeding M0.

Figure 3: AIC value of the RETAS model for different Mtr triggering magnitudes with cut-off magnitude (a) M0 = 2.9 and (b) M0 = 3.6. The best model is (a) the ETAS and (b) the RETAS model with Mtr = 5.0.

(a)
70 60 Cumulative number 50
AIC 273.5 273.0 272.5 272.0

(b)

RETAS models

The residual analysis reveals that the number of the observed events exceeds that of the expected ones nearly in the whole time interval (Fig. 4(a)).
(a) (b)

40 30 20 10 0 0 500 Time [days] 1000

7.8 6.6 5.4 4.2 3.0

2. Trigger models

Magnitude

271.5 271.0 270.5

Cumulative number

3.5

4.0

4.5

5.0

5.5

6.0

6.5

7.0

7.5

8.0

400 Magnitude 300 200 4.5 100 0 0 10 20 Time [days] 30 40 3.5 2.5 6.5 5.5

Cumulative number

(t|Ht) = +
{mP;tm<t}

Km (t tm + c)p

Triggering magnitude level

300 200

6.5 5.5 4.5

being P = {m = ik , k = 1, . . . , L} the subset of primary events triggering offspring. This set is the discriminant element among the models of the class. In particular, when P is constituted only by the main shock we have the inverse power law of the MOF [5] K (t) = (t + c)p ,

Figure 1: (a) Cumulative number and magnitude of the events in the sequence; (b) AIC value of the RETAS model for different Mtr triggering magnitudes. The minimum AIC value indicates that the best model is the RETAS model with Mtr = 7.8, equivalent to the MOF in which only the main shock can trigger descendants.

100 0 0 100 200 300 400 500

3.5 2.5

Magnitude

HE conditional intensity function of a trigger model is given by

500

500 400

References
[1] D. Gospodinov and R. Rotondi. Statistical analysis of triggered seismicity in Kresna region SW Bulgaria, 1904) and Umbria-Marche region (central Italy, 1997), to appear on Pageoph. [2] Y. Ogata. Statistical models for earthquake occurrences and residual analysis for point processes, J. Amer. Statist. Assoc., 83(401):927, 1988. [3] Y. Ogata. Seismicity analysis through point-process modeling: a review, In Seismicity Patterns, Their Statistical Signicance and Physical Meaning (eds. M. Wyss, K. Shimazaki and A. Ito) (Birkhuser, Basel 1999), Pure and Applied Geophysics, 155:471507. a [4] Y. Ogata. Exploratory analysis of earthquake clusters by likelihood-based trigger models, J. Appl. Probab., 38A:202121. [5] T. Utsu. A statistical study on the occurrence of aftershocks, Geophysical Magazine, 30:521605, 1961. [6] D. Vere-Jones and R.B. Davies. A statistical survey of earthquakes in the main seismic region of New Zealand, Part 2, Time Series Analyses, N. Z. J. Geol. Geophys., 9:251284, 1966.

Frequencylinearized time

The residual analysis (Fig. 2(b)) reveals that the observations exceed the expected number of events at the beginning of the sequence. We think that this is due to the series of foreshocks so strong as to inuence the start of the aftershock process.

Figure 4: Observed (blue) and estimated (red) cumulative number of events versus (a) the ordinary and (b) transformed time. The plug-in estimate corresponds to the RETAS = ETAS model (Mtr = 2.9). The dotted lines in (b) are the twofold standard deviations of the residual point process.

Statsei4 - The 4th International Workshop on Statistical Seismology - The Graduate University for Advanced Studies, Shonan Village campus, Kanagawa Prefecture, Japan, 9-13 January, 2006