Introduction to inverse problems

Practical sheet 1
Thursday 18th October 2007

Problem 1: Seismic profiling

A seismic profiling experiment is performed where the first arrival times of seismic energy
from a mid-crustal refractor are made at distances (km) of
 
6.0000
 10.1333 
 
 14.2667 
x=   (1)
 18.4000 

 22.5333 
26.6667

from the source, and are found to be (in seconds after the source origin time)
 
3.4935
 4.2853 
 
 5.1374 
t=  5.8181 
 (2)
 
 6.8632 
8.1841

A two layer flat Earth structure gives the mathematical model

ti = to + ssxi

where the intercept time to depends on the thickness and slowness of the upper layer, and s2
is the slowness of the lower layer. The estimated noise in first arrival time measurements is be-
lieved to be independent and normally distributed with expected value 0 and standard deviation
σ = 0.1s.

(a) Find the least squares solution for the two model parameters t0 and s2. Plot the data the
fitted model and the residuals.

(b) Calculate and comment on the parameter correlation matrix. How will the correlation
entries be reflected in the appearance of the error ellipsoid?

(c) Plot the error ellipsoid in the (t0, s2) plane and calculate conservative 95% confidence
intervals for t0 and s2. Hint: The following MATLAB code will plot a 2-dimensional
covariance ellipse, where covm is the covariance matrix and m is the 2-vector of model
parameters.

1
 %diagonalize the covariance matrix
[u,lam]=eig(inv(covm));
%generate a vector of angles from 0 to 2*pi
theta=(0:.01:2*pi);
%calculate the x component of the ellipsoid for all angles
r(:,1)=(delta/sqrt(lam(1,1)))*u(1,1)*cos(theta)+...
(delta/sqrt(lam(2,2)))*u(1,2)*sin(theta);
%calculate the y component of the ellipsoid for all angles
r(:,2)=(delta/sqrt(lam(1,1)))*u(2,1)*cos(theta))+...
(delta/sqrt(lam(2,2)))*u(2,2)*sin(theta);
%plot(x,y), adding in the model parameters
plot(m(1)+r(:,1),m(2)+r(:,2))

(d) Evaluate the p value for this model (you may find the MATLAB Statistics Toolbox func-
tion chi2cdf to be useful here).

(e) Evaluate the p value of χ2 for 1000 Monte Carlo simulations using the data prediction
from your model perturbed by noise that is consistent with the data assumptions. Compare
a histogram of these χ2 values with the theoretical χ2 distribution for the correct number
of degrees of freedom (you may find the MATLAB statistical toolbox function chi2cdf
to be useful here).

(f) Are your p value and Monte Carlo χ2 distribution consistent with the theoretical modeling
and the data set? If not, explain what is wrong.

(g) Use IRLS to evaluate 1norm estimates for t0 and s2. Plot the data predictions from your
model relative to the true data and compare with (a).

(h) Use Monte Carlo error propagation and IRLS to estimate symmetric 95% confidence
intervals on the 1norm solution for t0 and s2.

(i) Examining the contributions from each of the data points to the 1 norm misfit measure,
can you make a case that any of the data points are statistical outliers?

2
Problem 2: Earthquake location

An earthquake is often approximated as a point source and described by four hypocenter

(a) We will locate an earthquake using P-wave arrival times at stations of the Swiss national
seismic network. Assume that the P waves travel with a constant known velocity vp
along straight wave paths. Although a seismologist would not consider these to be valid
assumptions we will see that they are quite reasonable for obtaining a good idea of where
and when the earthquake occurred. Define te as the origin time of the earthquake and
(xe, ye, ze) as the spatial location. Set up the system of equations describing the relation
between the observed P-wave arrival times and the unknown earthquake parameters. Use
the notation (xi, yi, zi) for the location of the i-th station of the network.
(b) Is retrieving the hypocenter from P wave arrival times a linear or a non-linear inverse
problem? why ?
(c) Earthquake depth is often poorly constrained. In this exercise we fix the depth to be
ze = 10 km. For te and epicenter (xe, ye) we have to make an initial guess that we will
improve through inversion. Let’s use the notation t0 and (x0, y0) for the initial guess.
Now make a first order Taylor expansion of the earthquake-station distance D(xe, ye)
around (x0, y0). Here we approximate the epicenter by xe = x0 + δx and ye = y0 + δy.
Insert the Taylor expansion in your expression for the arrival time and obtain a linear
relationship between arrival times and hypocentre perturbations.
(d) Define the data di at station i as the difference between the original P wave arrival time
ti and the predicted arrival time determined from the initial guess original time t0 and
epicentre (x0, y0). Define the model as (m1, m2, m3)T = (δt, δx, δy)T . Does this pose a
linear or non-linear inverse problem?
(e) Write out an arbitrary row of matrix G of the earthquake location problem di = Gi,jmj.
(f) Now use MATLAB to locate the earthquake. The station coordinates and P wave arrival
times are stored in the data file loctim.txt. Start MATLAB and store the columns of the
data file in separate vectors. The first column contains latitude (in degrees), the second
longitude (in degrees), the third elevation (in km), and the fourth P wave arrival time (in
s after reference time 16:30) for 25 stations.
(g) To compute distances we need to know how many km correspond to degree in latitude
and longitude. For Switzerland one degree in latitude is approximately 111.19 km and
one degree in longitude is about 75.82 km. Store these values and the constant value
assumed for vp = 5.8 km/s in separate MATLAB parameters.
(h) Now make a first guess at the origin time and epicenter (t0, x0, y0). It might help to first
draw the station locations on a ”map”, using the commands plot(lon,lat,’*’) hold on Then

3
 guess t0 by dividing the estimated distance from guessed epicenter to a station by the
P wave velocity. Enter the guesses in MATLAB variables and plot the location on the
”map”. The file border.xy contains the boundary for Switzerland.

(i) Let MATLAB compute the vector D0 = D(x0, y0) with distances to all stations, the data
vector with arrival time corrected for the arrival times predicted from the first guess, and
the matrix G as defined in (8). (Make sure that the data vector is vertical.)

(j) Let MATLAB compute the generalized inverse matrix G−g for the least squares solution
mj = G−g
j,i di to the problem di = Gi,jmj.

(k) Using G−g let MATLAB compute the model vector m, which represents deviations to the
initial guess. Then update the initial guess with the values for vector m and plot the new
epicenter on the ”map”.

(l) Redefine the initial guess as the updated epicenter and origin time and repeat procedure
10 through 12 until the solution has converged. In this way we solve a non-linear inverse
problem by iteratively solving a linearized version of the inverse problem.

(m) Using 1s estimated standard deviation in the data. Calculate the data prediction error of
your solution. Then apply the goodness of fit test. What significance level (p value or %)
do you get for the solution. Interpret the result.

(m) Finally let MATLAB compute the covariance matrix for model vector m and estimate the
standard deviation in origin time and epicentral coordinates, presuming for simplicity that
the arrival time data have a standard deviation of 1 s.

(n) Compare your solution for this earthquake, which occurred on Nov 13, 2000, with the
SED (Swiss Seismological Service) solution. Is it consistent ?

D a t e Time (UTC) Location Dep Magni. T AGY R e g i o n HHMM

4
Do you get something like this ?

48
stations
initial guess
intermediate solutions
solution

47.5

47

46.5

46

45.5
6 7 8 9 10 11