Documente Academic
Documente Profesional
Documente Cultură
1785/0120050039
Short Notes
Three Theorems of Earthquake Location
by Cinna Lomnitz
Abstract I prove three theorems on earthquake location in the laterally homogeneous, radially concentric layered earth showing that (1) epicenters of earthquakes
are locatable without using numerical travel-time information, (2) focal-depth residuals change sign at some epicentral distance that is a function of the focal depth but
not of the distance residuals, and (3) the distance D between two random points on
the surface of a sphere is distributed as sinD. Theorem (1) implies that the epicenter
can be located without simultaneously solving for hypocentral depth or origin time.
Theorem (2) suggests that the focal depth should never be determined by joint regression on all four hypocentral parameters. Theorem (3) suggests a likely reason
why late readings tend to overpower accurate readings when a standard least-squares
technique is used. Together these theorems suggest that the problem of earthquake
location is ill posed.
Joint least-square regression will lead to the true hypocenter only if well-posedness
is restored by restricting the class of admissible solutions using a priori knowledge,
such as a travel-time table. But if the travel-time table is derived from the hypocentral
solutions there is a feedback between location errors and errors in the travel-time
tables, and local minima cannot be eliminated. Gutenberg instructively attempted to
correct for bias due to azimuthal clustering, by assuming that residuals were nonnegative vectors.
Introduction
Large mislocations of earthquake hypocenters are not
uncommon. The published confidence bounds of earthquake
location errors are widely believed to be unrealistic. However, an objective comparison between different location
techniques, or a calibration of errors arising from various
approaches, is hampered by the scarcity of standard measurements from controlled sources.
Jeffreys (1970, p. 490) pointed out that the standard
error is a minimum estimate of uncertainty. Mislocations
are common even in large, well-recorded events. For example, the 1960 Chile earthquake (Mw 9.5) was found to be
mislocated by about 75 km landward after its relocation by
seismic-imaging techniques (Krawczyk et al., 2003). Menke
and Schaff (2004) tested routine location procedures on simulated data and concluded that the errors were as large as or
larger than those found from relative locations for pairs of
earthquakes.
Current procedures of earthquake location may be selfconsistent without being necessarily accurate, because they
are not checked against data from controlled sources. In this
article we re-examine some of the fundamental assumptions
that underlie location procedures. We show that the problem
of locating a point inside a radially symmetrical sphere from
307
Short Notes
(1)
308
Short Notes
Ti Ti1. But even this simple idea involves serious problems. The travel time is a positive increasing function of the
epicentral distance D, but we dont know how to estimate
the intercept at the origin.
Corollary 2. In principle, one might estimate a provisional
travel-time curve T(D) by fitting the arrival-time differences
to the epicentral distance D. Next we may try to find the
velocity structure of the interior by inverting T(D). In this
way, one should be able to obtain increasingly accurate estimates of the focal depth h. But such estimates will be affected by errors in measuring the arrival times and in estimating the travel-time curveparticularly, the zero-distance
interceptin the presence of lateral heterogeneity. The estimation of the epicentral parameters (, k) might remain
reasonably free of these errors as shown by theorem (1), as
long as we keep it separate from the procedure of focal-depth
estimation.
Figure 2.
309
Short Notes
estimate of h). On the other hand, where there are many nearfield stations, as in California, the focal depths may be systematically overestimated.
(2)
Figure 3.
and
F(D x) 0.5 (1 cosx) ,
0 x p,
(3)
where F is the cumulative probability distribution of the angular distance x. By differentiation, the probability density
is found to be f(x) 0.5 sinx.
Corollary 1. The epicentral location of earthquakes ought
not to depend on the location of seismic observations
(though a case might be made for the converse to be true!).
Yet the actual sample of arrival-time measurements is always likely to be dominated by distant observations.
For any given earthquake of epicenter O the probability
of a station P being located at some distance D looks like
Figure 4. This may be easily confirmed by plotting the actual
distribution of epicentral distances in a seismic bulletin.
Thus we should expect the observations to increase in number with increasing distance, up to a distance of 90 degrees,
and then decay back to zero at 180 degrees. Actually small
earthquakes are much more frequent than are large ones, and
small shocks are not recorded at large distances. Thus the
actual distribution begins to fall off well before 90 degrees:
it looks like the dotted line in Figure 4. The peak in the
distribution depends on the magnitude of the earthquake.
In conclusion, the probability of an observation is not
uniformly distributed in distance. Therefore the sample of
observations always tends to be dominated by distant observations. But the amplitude and the dominant frequency
of the signal decrease with distance while the noise remains
independent of distance, as it depends only on local conditions. Hence the signal/noise ratio tends to decrease with
Figure 4.
distance. If distant observations also tend to be more numerous, we have a serious bias. The location procedure is
likely to be dominated by low-accuracy observations at large
distances.
Corollary 2. The joint-regression procedure may tend to
introduce a systematic error into the epicentral solution when
310
Short Notes
An Ill-Posed Problem
A problem is well posed when a solution exists, is
unique, and depends continuously on the initial data. It is ill
posed when it fails to satisfy at least one of these criteria
(Hadamard, 1923). A well-known example of an ill-posed
problem is scene analysis, where one attempts to extract 3D
information from a 2D image.
The above theorems, and others like them, suggest that
the earthquake-location problem may be ill posed. For example, a least-square location procedure minimizes R(q2),
where q is the travel-time residual. This implies that
Prob{q} Prob{q}, as assumed by C. F. Gauss (1823)
in his derivation of the bell-shaped curve named after him.
But when a seismic signal arrives at a station it gradually
emerges from the local noise. The arrival time is picked with
some delay q, which reflects inversely the signal/noise
ratio. On the other hand, picking a signal in advance of its
actual arrival time is a purely mental phenomenon, since
there is nothing there to be picked. Thus a negative residual
is not the same kind of object as a positive residual. In one
case there is a signal, in the other the signal does not yet
exist. The bias is a result of causality.
Somewhat similarly, in joint-regression procedures an
error in focal depth does not affect the solution in the same
way as an error in latitude or in longitude. The bias is due
to spherical geometry. All our observations are on the surface of the sphere, and negative depths do not exist.
Well-posedness may be restored by restricting the class
of admissible solutions using a priori knowledge, such as a
travel-time table. In this case the solution space is assumed
Discussion
I realize that I have not provided a solid theory of earthquake location, or a remedial approach supported by real
data. The reason is clear. The location problem is ill posed
and there are not enough observations of controlled sources.
Controlled explosions are rarely large enough to calibrate a regional network. The early results using nuclear tests
for amending travel-time tables have been disappointing,
precisely because of serious discrepancies with tables in
common use. In the end, seismologists were tempted to
merge explosion data with earthquake data. Travel times
from earthquakes have even been used to correct explosion
data!
Statisticians are apt to suggest that travel-time residuals
should be transformed into distance vectors on the surface
of the sphere. But statistics on the sphere is an arcane subject
(Mardia and Gadsden, 1977). In theory there is no obvious
way of translating a least-square algorithm in the time domain to the space domain. Should we minimize the vectorial
sum of the distance residuals? This is not the correct answer
either.
Beno Gutenberg used a graphical location method that
was routinely utilized in Pasadena for years, particularly by
Charles F. Richter in the classical joint research work Seismicity of the Earth (Gutenberg and Richter, 1954). Many
solutions are on file at the Caltech Seismological Laboratory
where Gutenbergs famed yellow pads are still kept. While
the criteria used by Gutenberg to weigh stations were never
set down in writing, they could be inferred from the solutions, as the residuals were normally labeled.
Gutenberg realized, even before plate tectonics, that
seismic observations were not scattered at random about the
epicenter. They were clustered in certain azimuths and distance ranges. Stations in the near field or in the southern
continents were rare and might easily be overwhelmed by
distant stations, for example, in the United States or in Europe, during regression. More importantly, Gutenberg never
attempted to include the focal depth in his location procedure. Focal depth was estimated separately from reflected
phases such as pP, sP, and PcPnever by joint regression.
This was because Gutenberg was keenly aware of the fact
that the sign of the focal-depth residual changes over small
distances in the near field.
Consider an earthquake epicenter E(k,), where k is the
311
Short Notes
Conclusion
Theorem (1) shows that the epicentral parameters k and
may be determined on the sphere without using travel
times or travel-time tables. This suggests a location method
that assumes only radial homogeneity and is free of any bias
introduced by travel-time measurements on the sphere. The
location can be made as accurate as desired. It could be
adopted as the gold standard for earthquake locations.
The focal depth remains indeterminate and must be determined separately. This is because all the measurements
are on the surface. In joint regression the errors in the focal
depth will contaminate the epicentral solution and vice versa.
Theorem (2) shows that the estimation of focal depth
from travel-time residuals is not unique, because the rays
that exit the focal sphere at the equator divide the surface of
the sphere in two regions. Residuals will have opposite effects on focal-depth estimation depending on which of these
two regions contains the observation. Since the areas of the
two regions are very unequal there is a strong bias in focaldepth estimation.
Theorem (3) is intended to show that epicentral location
is biased too, but in a different way. In a random field of
stations the number of observations increases as the sine of
the distance. This means that most samples are dominated
by distant observations. But at distant stations the frequency
content of the initial P wave becomes increasingly similar
Figure 5. Gutenbergs method of earthquake location. Station residuals, black dots; residuals at
good stations, used for constraining the epicentral
correction curve, open dots. After the final iteration
the correction curve should become a straight horizontal line.
to the frequency of the background noise. This means that
detection of early arrivals becomes increasingly difficult,
and residuals are increasingly positive. As stations tend to
cluster in certain areas, such as California or Europe, epicentral solutions tend to be pulled in the direction of these
areas.
These results are intended to illustrate the fact that the
location of earthquakes from observations on the Earths surface is an ill-posed problem. We do not know how to locate
earthquakes. Therefore we dont know the structure of the
earths interior either, as it is derived from the inversion of
travel-time tables obtained from earthquake locations. Some
scientists have been aware of this situation for years, and
have bravely attempted to restore well-posedness by inventing new ways of locating earthquakes, or by restricting the
class of admissible solutions using a priori information.
Unfortunately, the location problem is the fundamental
problem of seismology. If we do not know how to locate
earthquakes, we cannot be certain of our information on the
Earths interiornot so much because it matters whether
some earthquake is mislocated a few miles seaward or landward but because locations have been used since Zoeppritz
(1907) to estimate travel-time tables. Location errors creep
into travel-time tables and from there into models of Earth
structure. There does not seem to be any good way of preventing this, since we are dealing with the effects of bias.
Earthquakes are not homogeneously distributed in space and
neither are stations. Plate tectonics make sure that observations will be skewed by the position of the major continents.
This bias is systematically built into the travel-time tables.
312
Short Notes
References
Anderson, K. R. (1978). Automatic processing of local earthquake data,
Ph.D. Thesis, Massachusetts Institute of Technology, 173 pp.
Gauss, C. F. (1823). Theoria Combinationis Observationum Erroribus Minimis Obnoxiae, Henricus Dieterich, Gottingen.
Geiger, L. (1910). Herdbestimmung bei Erdbeben aus den Ankunftszeiten,
Kgl. Gesell. Wiss. Gott., Nachrichten, Mathematisch-Physikalische
Klasse, 331349.
Gutenberg, B., and C. F. Richter (1954). Seismicity of the Earth and Associated Phenomena, Princeton University Press, Princeton, NewJersey.
Hadamard, J. (1923). Lectures on the Cauchy Problem in Linear Partial
Differential Equations, Yale University Press, New Haven, Connecticut.
Jeffreys, H. (1970). The Earth, Fifth Ed., Cambridge University Press,
Cambridge, Massachusetts.
Krawczyk, C., and the SPOC Team (2003). Amphibious seismic survey
images plate interface at the 1960 Chile earthquake, EOS 84, 301,
304305.
Lomnitz, C. (1982). An uncertainty principle of earthquake locations, Bull.
Seism. Soc. Am. 72, 669670.
Mardia, K. V., and R. J. Gadsden (1977). A small circle of best fit for
spherical data and areas of vulcanism, J. Roy. Stat. Soc. C 26, 238
245.
Menke, W., and D. Schaff (2004). Absolute earthquake locations with differential data, Bull. Seism. Soc. Am. 94, 22542264.
ber Erdbebenwellen II. Laufzeitkurven, Kgl. Gesell.
Zoeppritz, K. (1907). U
Wiss. Gott., Nachrichten, Mathematisch-Physikalische Klasse, 529
549.
Institute of Geophysics
UNAM
Mexico City, Mexico
Manuscript received 4 March 2005.