See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328949750

Joint Maximum Likelihood Estimators for Gutenberg-Richter Parameters λ 0

and β using Sub-Catalogs

Article in Earthquake Spectra · November 2018

DOI: 10.1193/071718EQS179D

CITATIONS READS
2 192

2 authors:

Petrus Johannes Vermeulen Andrzej Kijko

University of Pretoria Natural Hazard Centre, University of Pretoria
5 PUBLICATIONS 29 CITATIONS 191 PUBLICATIONS 4,051 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Thesis: Parameter estimation in PSHA. Current focus: estimating maximum possible magnitude View project

Reliability of earthquake hazard assessment in Iceland: improved models, uncertainties and sensitivities (SENSHAZ) View project

All content following this page was uploaded by Petrus Johannes Vermeulen on 05 April 2019.

The user has requested enhancement of the downloaded file.

 1 DISCUSSION OF

2 Joint Maximum Likelihood Estimators for

3 Gutenberg-Richter Parameters 𝝀𝟎 and 𝜷 using
4 Sub-Catalogs
5 Manuscript Reference: M. Ordaz and S. Giraldo, Earthquake Spectra, vol. 34, no. 1 (February 2018):
6 301-312.

7 Petrus J. Vermeulen,a) and Andrzej Kijkoa)

8 The maximum likelihood estimator solved by by M. Ordaz and S. Giraldo is

9 shown to be a special case of the maximum likelihood estimator that results from
10 application of the Kijko-Sellevoll procedure. The equation to obtain the maximum
11 likelihood estimator given by Ordaz and Giraldo is therefore the same as that which
12 would result from the application of the Kijko-Sellevoll procedure. It is good to
13 take cognizance of both the general solution of Kijko and Sellevoll and the specific
14 case solved by Ordaz and Giraldo.

15 INTRODUCTION

16 Ordaz and Giraldo (2017), further on referred to as O-G, attempt to improve upon the
17 extended Aki-Utsu maximum likelihood estimator, and that given by Kijko and Smit (2012)
18 for 𝛽 and 𝜆. Given a seismic catalog that comprises more than one sub-catalog with different
19 levels of completeness (Figure 1), the extended Aki-Utsu estimators are given by
(1)
−1
𝑟1 𝑟2 𝑟𝑠
20 𝛽̂ = ( + + ⋯ + )
𝛽̂1 𝛽̂2 𝛽̂𝑠
𝑛
21 𝜆̂ = ∑𝑠 𝑖 ,
𝑖=1 𝑇𝑖 exp⁡(−𝛽(𝑚𝑚𝑖𝑛 −𝑚𝑚𝑖𝑛 )

22 where 𝛽̂𝑖 denotes the classic Aki-Utsu estimator for each sub-catalog, 𝑖
1⁄(𝑚𝑚𝑖𝑛 − 〈𝑚〉𝑖 ),
𝑖
23 with 𝑚𝑚𝑖𝑛 the level of completeness of the 𝑖 𝑡ℎ sub-catalog and⁡𝑚𝑚𝑖𝑛 is the overall chosen
24 minimum value of magnitude taken into consideration. The number 𝑟𝑖 is defined as the ratio of
25 the number 𝑛𝑖 of events in the 𝑖 𝑡ℎ sub-catalog to the total number of events 𝑛 in the entire

a)
Natural Hazard Center, University of Pretoria; Email: u28038020@tuks.co.za (P. J. V.)
 𝑛𝑖
26 catalog, that is 𝑟𝑖 = . The estimator 𝜆̂ denotes the rate of seismicity of the whole catalog, and
𝑛

27 𝑇𝑖 denotes the time span of the 𝑖 𝑡ℎ sub-catalog.

28
29 The improvement O-G propose is a joint maximum likelihood estimation of the pair (𝛽, 𝜆).
30 This is indeed an improvement, as it uses not the marginal likelihoods of 𝛽 and 𝜆 but also their
31 joint (simultaneous) likelihood. In addition, O-G show numerically that this estimation is
32 superior to the separate, marginal maximum likelihood estimators as applied by Kijko and Smit
33 (2012). However, it is interesting to note, as we have discovered, that the equations given by
34 O-G turns out simply to be a special instance in the scheme developed by Kijko and Sellevoll
35 (1989) (further on referred to as K-S). If no extreme part of the catalog were used, and one
36 supposed 𝑚𝑚𝑎𝑥 = ∞ , the O-G equations looked exactly the same as those from the scheme
37 developed by K-S. The derivation of the likelihood functions differ slightly from K-S in that
38 O-G use every single interval between consecutive earthquakes, whereas K-S use the total time
39 span of each complete sub-catalog. It turns out that these differences in the derivation lead to
40 equivalent likelihood functions. Specifically, this is because the construction of the likelihood
41 function involves the multiplication of exponential terms having time as a factor in the
42 exponent, i.e., it is of the form ∏𝑖(𝜆𝑡𝑖 )exp⁡(−𝜆𝑡𝑖 ) = (𝜆𝑡𝑖 )𝑛 exp(𝜆 ∑𝑖 𝑡𝑖 ) = (𝜆𝑡𝑖 )𝑛 exp(𝜆𝑇),
43 where 𝑇 is the time span of a catalog or sub-catalog, 𝜆 is the Poissonian rate, and 𝑛 is the
44 number of events. This is to be expected, as the likelihood functions are derived from the same
45 initial distributions of magnitude and inter-event time distribution. The attentive reader might
46 note that K-S make use of the likelihood of observing 𝑛𝑖 earthquakes in a time 𝑇𝑖 (K-S, equation
47 8), and O-G use the the inter-event time distribution. However, from a formal point of view,
48 the equivalence of the descriptions of the Poisson process (as a counting process, or a
49 distribution of inter-event times) tells us that the outcomes should be equivalent. Recall that
50 the Poisson process as a counting process is characterized by the distribution
(𝜆𝑡)𝑛 (1)
𝑃[𝑁(𝑡) = 𝑛] = 𝑒 −𝜆𝑡 .
𝑛!

51
52 On the other hand, the distribution of inter-event times is given by
𝑃[𝑖𝑛𝑡𝑒𝑟𝑎𝑟𝑟𝑖𝑣𝑎𝑙⁡𝑡𝑖𝑚𝑒 > 𝑡] = 𝑒 −𝜆𝑡 (2)
53 THE LIKELIHOOD EQUATIONS

54 In this section, the derivation of the likelihood equations of O-D and K-S will be reviewed
55 and discussed in more detail. K-S derive the likelihood equation quite briefly and compactly
56 and do not state the end result, as the reader is expected to be acquainted well with the
57 procedures the authors are following. These equations will be derived in added detail here and
58 the result will be given explicitly. O-G give the full derivation up to the likelihood equation,
59 but we will restate it here for the sake of completeness and comparison. To simplify and
60 facilitate easy comprehension of the derivations, some notation borrowed from O-G is
61 introduced here
̂
62 𝑇𝑗∗ = 𝑇𝑗 𝑒 −𝛽(𝑚0𝑗−𝑚0 )
𝐿 (4)
𝑇 = ∗
∑ 𝑇𝑗∗
𝑗=1

63
𝑠
(5)
𝑄 = ∑ 𝑛𝑗 (𝑚0𝑗 − 𝑚0 )
𝑗=1

64
𝑠
(6)
𝑆 = ∑ 𝑆𝑗
𝑗=1

65
𝑛𝑗
(7)
𝑆𝑗 = ∑(𝑚𝑖,𝑗 − 𝑚0 )
𝑖=1
66

67 APPROACH FOLLOWED BY KIJKO AND SELLEVOLL (1989)

68 Following K-S, let us assume that earthquake magnitudes follow the doubly truncated
69 Gutenberg-Richter distribution. The likelihood function obtained from this distribution gives a
70 likelihood in terms of 𝛽, and for a specific sub-catalog this is
71
𝑛
−𝛽 ∑𝑗 𝑖 𝑚𝑖𝑗 (8)
𝐾−𝑆 𝑛𝑖
𝑒 𝑛
−𝛽 ∑𝑗 𝑖 (𝑚𝑖𝑗 −𝑚𝑖 )
ℒ𝑖𝛽 =𝛽 = 𝛽 𝑛𝑖 𝑒 𝑛𝑖 −𝛽𝑆𝑖
=𝛽 𝑒
𝑒 −𝛽𝑚𝑖

72
73 Not providing much motivation, as it is assumed that the reader is familiar with the Poisson
74 process, the likelihood function involving 𝜆 and 𝛽, which is, in fact, merely the probability
75 mass function for a time period 𝑇𝑖 takes the form
76

𝐾−𝑆 𝑇𝑖 𝑛𝑖 (−𝜆𝑒 −𝛽Δ𝑖 𝑇 ) −𝛽Δ𝑖 𝑛𝑖

𝑇𝑖 𝑛𝑖 (−𝜆𝑒 −𝛽Δ𝑖 𝑇 ) 𝑛 −𝛽𝑛 Δ (9)
ℒ𝑖𝜆 = ∙𝑒 𝑖 (𝜆𝑒 𝑇𝑖 ) = ∙𝑒 𝑖 𝜆 𝑖𝑒 𝑖 𝑖
𝑛! 𝑛!
77
78 where
79
𝜆𝑒 −𝛽𝑚𝑖 (10)
𝜆𝑖 = 𝜆(1 − 𝐹(𝑚𝑖 )) = = 𝜆𝑒 −𝛽Δ𝑖
𝑒 −𝛽𝑚𝑚𝑖𝑛

80
𝑒 −𝛽𝑚𝑚𝑖𝑛 −𝑒 −𝛽𝑥
81 𝐹(𝑥) = and Δ𝑖 = 𝑚𝑖 − ⁡ 𝑚𝑚𝑖𝑛 .
𝑒 −𝛽𝑚𝑚𝑖𝑛

82
83 The joint likelihood function is obtained by combining functions (8) and (9) over the s complete
84 sub-catalogs
85
𝐾−𝑆 𝐾−𝑆 𝑇 −𝛽Δ𝑖 𝑇 ) 𝑇 ∗
(11)
ℒ𝑖𝛽 ℒ𝑖𝜆 = ∏𝑠𝑖 𝛽 𝑛𝑖 𝑒 −𝛽𝑆𝑖 𝑛!𝑖 ∙ 𝑒 (−𝜆𝑒 𝑖 𝜆𝑛𝑖 𝑒 −𝛽𝑛𝑖 Δ𝑖 = 𝑛! ∙ 𝛽 𝑛 𝜆𝑛 𝑒 −𝛽𝑆 𝑒 −𝛽𝑄 𝑒 −𝜆𝑇 ,
 𝑇
86 Note that coefficient does not have any effect on the values of 𝛽 and 𝜆; therefore, the
𝑛!

87 likelihood function might well be written as

88
∗
ℒ𝛽 ℒ𝜆 = 𝛽 𝑛 𝜆𝑛 𝑒 −𝛽𝑆 𝑒 −𝛽𝑄 𝑒 −𝜆𝑇 (12)

89 where 𝑇 ∗ is as defined in the previous section.

90

91 APPROACH FOLLOWED BY ORDAZ AND GIRALDO (2017)

92 Consider the probability density function of inter-arrival times

93
𝑓𝑡 = 𝜆𝑒 −𝜆𝑡 (13)
94
95 Therefore, the likelihood of having 𝑛𝑖 intervals between the events in a sub-catalog of duration
96 𝑇𝑖 is given by
97
𝑛𝑖
𝑖 𝑡 𝑛 (14)
𝑂−𝐺 −𝜆𝑖 ∑𝑗=0
ℒ𝑖𝜆 = ∏ 𝜆𝑖 𝑒 −𝜆𝑖 𝑡𝑗,𝑖 = (𝜆𝑖 )𝑛𝑖 𝑒 𝑗,𝑖
= (𝜆𝑖 )𝑛𝑖 𝑒 −𝜆𝑖 𝑇𝑖
𝑗=0

98
99 where 𝑡𝑗,𝑖 denotes the 𝑗 𝑡ℎ inter-event time interval in the 𝑖 𝑡ℎ sub-catalog. Note that we are
100 looking at the likelihood of observing 𝑛 time intervals. This is, in fact, a counting process and
101 turns out to be the counting process characterizing the Poisson process. This will reveal
102 similarity (or stronger even, equivalence) of the likelihood function derived by O-G and K-S.
103 Note that
104
(𝑖)
𝜆𝑖 = 𝜆[1 − 𝐹𝑀 (𝑚𝑚𝑖𝑛𝑖 )] = 𝜆𝑒 −𝛽(𝑚𝑚𝑖𝑛 −𝑚0 ) (15)

105
106 Then, substituting equation (15) in (14) gives
107
𝑛𝑖
(𝑖) −𝛽(𝑚
(𝑖)
−𝑚0 )
(16)
𝑂−𝐺
ℒ𝑖𝜆 = ∏ 𝜆𝑖 𝑒 −𝜆𝑖 𝑡𝑗,𝑖 = (𝜆0 )𝑛𝑖 𝑒 −𝛽(𝑚𝑚𝑖𝑛 −𝑚0 ) 𝑒 −𝜆0 𝑒 𝑚𝑖𝑛 𝑇𝑖

𝑗=0
 −𝛽Δi 𝑇
= (𝜆0 )𝑛𝑖 𝑒 −𝛽Δi 𝑒 −𝜆0 𝑒 𝑖

∗
= (𝜆0 )𝑛𝑖 𝑒 −𝛽Δi 𝑒 −𝜆0 𝑇𝑖
108
109 For a given 𝛽 value, the likelihood of observing 𝑛𝑖 events in a sub-catalog is given by
110
𝑛𝑖
(𝑖) 𝑛
𝑖 (𝑚 −𝑚 (𝑖) (17)
𝑂−𝐺 −𝛽(𝑚𝑗,𝑖 −𝑚𝑚𝑖𝑛 ) −𝛽 ∑𝑗=1 𝑚𝑖𝑛 )
ℒ𝑖𝛽 = ∏𝛽𝑒 = 𝛽 𝑛𝑖 𝑒 𝑗,𝑖

𝑗=1

= 𝛽 𝑛𝑖 𝑒 −𝛽𝑆𝑖
111
112 Combining the likelihood functions and considering the entire catalog
113
𝑠
∗
(18)
𝑂−𝐺 𝑂−𝐺
ℒ𝑖𝛽 ℒ𝑖𝜆 = ∏ 𝛽 𝑛𝑖 𝑒 −𝛽𝑆𝑖 (𝜆0 )𝑛𝑖 𝑒 −𝛽Δi 𝑒 −𝜆0 𝑇𝑖
𝑖
𝑛 (𝜆 𝑛 −𝛽𝑆 −𝛽𝑄 −𝜆𝑇 ∗
=𝛽 0) 𝑒 𝑒 𝑒

114

115 MAXIMUM LIKELIHOOD EQUATIONS

116 Thus we see that the two likelihood functions (12) and (18) are the same. Kijko and
117 Sellevoll (1989) give a general solution, which is rather cumbersome to work with, but they
118 note that the equations are derived from maximizing the log-likelihood functions by setting the
119 derivatives of the log-likelihood function with respect to 𝜆 and 𝛽, respectively, equal to zero.
120 Neither K-S nor O-G show the calculations explicitly. The derivation is not extensive and is
121 given here. Firstly, the log-likelihood function is
122
∗
log(ℒ𝛽 ℒ𝜆 ) = log(𝛽 𝑛 𝜆𝑛 𝑒 −𝛽𝑆 𝑒 −𝛽𝑄 𝑒 −𝜆𝑇 ) (19)
= 𝑛⁡𝑙𝑜𝑔(𝛽) + 𝑛⁡𝑙𝑜𝑔(𝜆) − 𝛽(𝑄 + 𝑆) − 𝜆𝑇 ∗ ,

123 where 𝑙𝑜𝑔(. ) denotes a natural logarithm.

124 The derivative of the log-likelihood function with respect to 𝜆 is
125
 𝜕 𝜕 (20)
log(ℒ𝛽 ℒ𝜆 ) = (𝑛⁡𝑙𝑜𝑔(𝛽) + 𝑛⁡𝑙𝑜𝑔(𝜆) − 𝛽(𝑄 + 𝑆) − 𝜆𝑇 ∗ )
𝜕𝜆 𝜕𝜆
𝑛
= + 𝑇∗
𝜆
126
127 and the derivative of the log-likelihood function with respect to 𝛽 is
128
𝜕 𝜕 (21)
log(ℒ𝛽 ℒ𝜆 ) = (𝑛⁡𝑙𝑜𝑔(𝛽) + 𝑛⁡𝑙𝑜𝑔(𝜆) − 𝛽(𝑄 + 𝑆) − 𝜆𝑇 ∗ )
𝜕𝛽 𝜕𝛽
𝑛 𝜕
= − (𝑄 + 𝑆) − ⁡(𝜆𝑇 ∗ )
𝛽 𝜕𝛽
𝑠
𝑛
= − (𝑄 + 𝑆) − ⁡𝜆 ∑(𝑇𝑖 𝑒 −𝛽Δi )
𝛽
𝑖
𝑠
𝑛 𝜕
= − (𝑄 + 𝑆) − ⁡𝜆 ∑ (𝑇𝑖 𝑒 −𝛽Δ𝑖 (−Δi ))
𝛽 𝜕𝛽
𝑖
𝑠
𝑛
= − (𝑄 + 𝑆) + 𝜆 ∑ (𝑇𝑖 𝑒 −𝛽Δ𝑖 (Δi ))
𝛽
𝑖
𝑠
𝑛
= − (𝑄 + 𝑆) + ⁡𝜆 ∑(𝑇𝑖∗ Δi )
𝛽
𝑖

129
130 Then, equating these derivatives to zero, the following is obtained
131
𝜕 (22)
log(ℒ𝛽 ℒ𝜆 ) = 0
𝜕𝜆
𝑛
+ 𝑇∗ = 0
𝜆

132
133 and
134
𝜕 (23)
log(ℒ𝛽 ℒ𝜆 ) = 0
𝜕𝛽
 𝑠
𝑛
− (𝑄 + 𝑆) + ⁡𝜆 ∑(𝑇𝑖∗ Δi ) = 0
𝛽
𝑖

135
136 and substituting the estimate of 𝜆 obtained from equation (22) in equation (23), the nonlinear
137 system of equations given by O-G is obtained
138
𝑛 (24)
+ 𝑇∗ = 0
𝜆
𝑠
𝑛 𝑛
− (𝑄 + 𝑆) + ⁡ ∗ ∑(𝑇𝑖∗ Δi ) = 0
𝛽 𝑇
{ 𝑖

139
140 Considering that K-S also start from the same likelihood equation and use the same method of
141 maximizing the likelihood, the two solutions must be the same. This shows that the work of O-
142 G is a special case of the work of K-S.
143
144
145 CONCLUSIONS

146 O-G give an improved (joint) maximum likelihood estimator for the parameters 𝛽 and 𝜆,
147 compared with the extended Aki-Utsu estimator developed by Kijko and Smit (2012), equation
148 (1). It was shown here that the joint maximum likelihood estimator of O-G is simply a special
149 case of the joint maximum likelihood equation given by K-S.
150
151
152 ACKNOWLEDGEMENTS

153 NRF Incentive Funding: 103724:2017

154 NRF THRIP MMI Holdings 96412:2018
155 NRF Grant SFH13081828987
156
157 REFERENCES

158 Kijko, A., & Sellevoll, M. A., 1989. Estimation of earthquake hazard parameters from
159 incomplete data files. Part I. Utilization of extreme and complete catalogs with different
160 threshold magnitudes, Bulletin of the Seismological Society of America, 79 (3), 645-654.
161
162 Kijko, A. and Smit, A., 2012. Extension of the Aki‐Utsu b‐value estimator for incomplete
163 catalogs, Bulletin of the Seismological Society of America., 102 (3), 1283-1287.
164
165 Marzocchi, W. and Taroni, M., 2014. Some thoughts on declustering in probabilistic seismic‐
166 hazard analysis, Bulletin of the Seismological Society of America, 104 (4) , 1838-1845.
167
168 Ordaz, M. and Giraldo, S., 2018, Joint Maximum Likelihood Estimators for Gutenberg-Richter
169 Parameters λ 0 and β Using Subcatalogs, Earthquake Spectra, 34 (1), 301-312.
170

171
172 Figure 1. A schematic illustration of a seismic event catalog that can be used in the estimation
173 of the Gutenberg-Richter 𝛽-value (after Kijko and Smit, 2011).

174

View publication stats