Вы находитесь на странице: 1из 9

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

GEOPHYSICS, VOL. 56, NO. 12 (DECEMBER 1991); P. 1999--2007,4FIGS.

Phase-smoothed robust M-estimation of magnetotelluric impedance functions

D. Sutarno* and K. Vozoff*

ABSTRACT

Accurate estimation of impedance functions is es- sential for the correct interpretation of magnetotelluric (MT) measurements. Noise is inevitably encountered when MT observations are conducted and, conse- quently, impedance estimates are usually based on least-squares (LS) regression. Least squares ulti- mately assumes simple Gaussian statistics. However, estimation procedures based on LS would not be statistically optimal, as outliers (abnormal data) are frequently superimposed on a normal ambient MT noise field which is approximately Gaussian. In this situation, the estimation can be seriously misleading. An alternative method for making unbiased robust estimates of MT impedance functions is based on regression M-estimation and the Hilbert Transform, operating on minimum-phase MT impedance func- tions. In the resulting regression estimates, outlier contamination is removed and other departures from Gauss-Markov optimality are not critical. Using MT data from the Columbia River Plateau and the EMSLAB Lincoln line, it is shown that the method can produce usable MT impedance functions even under conditions of severe noise contamination and in the absence of remote reference data.

INTRODUCTION

In the frequency domain, the equations governing the magnetotelluric (MT) relations between the signal compo- nents of the electric and magnetic fields at the earth's surface are given by:

i = x

or y.

(1)

From the viewpoint of linear system theory, the tensor impedances Zix and Ziy are transfer functions of a dual input, single output linear system through which the hori-

zontal magnetic field components H x and H y (input) are related deterministically to the horizontal electric field com- ponents E, (output).

A variety of methods have been proposed for the numer-

ical computation of impedance functions and their associ- ated errors in the presence of noise. Most ofthese are based on classical least squares (LS) regression. It is well known that the concept, as well as the computation, of LS estimates is quite simple. Furthermore, with the Gaussian error as-

sumption, LS procedures are statistically optimal in a pre- cise sense. Unfortunately, the presence of outliers (abnor- mal data) superimposed on a common Gaussian noise background which constitutes the normal ambient noise field, often makes such assumptions about the error unten- able and results in a distortion of the estimates. One appeal- ing approach to dealing with outliers is to make the estima- tion procedure robust, i.e. to modify it in such a manner that it is resistant to the effects of the outliers.

In the geophysical context, robust procedures were intro-

duced by Claerbout and Muir (1973). Egbert and Booker (1986) and Chave et al. (1987) proposed the application of a robust procedure which can be classified as an M-estimator to geomagnetic data. More recently, Jones et al. (1989), Chave and Thomson (1989), and Sutarno and Vozoff (1989) also implemented this procedure with MT data. This paper aims at improving the performance of the robust procedure. Our object is to obtain unbiased robust impedance estimates in the absence of remote reference (RR) data.

ROBUST IMPEDANCE ESTIMATION

In the real situation, equation (1) does not hold exactly due

to the presence of noise. Therefore, it is necessary to estimate the impedance functions from many noisy data, and the problem becomes statistical. Suppose n independent sets of evaluations at a given harmonic are available to estimate Zij. Then the following equation, which represents a linear regression problem, can be formed

x = V0il + r.

(2)

. *Centre for Geophysical Exploration Research, Macquane University, New South Wales, 2109, Austraha. © 1991 Society of Exploration Geophysicists. All rights reserved.

Manuscript received by the Editor December 18, 1990; r~vise~ ma~uscript received May 28, 1991.

1999

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2000

Sutarno and Vozoff

The vector x consists of n observations of the ith horizontal electric field component, the n x 2 matrix Vgives n values of the dual horizontal magnetic field components, the ele- ments of !JA are impedance tensor elements (the unknown parameters to be estimated), and the elements of rare unknown errors. The above equation assumes that noise is restricted to an output, while the input is assumed to be noise free. Therefore this assumption may be approximately valid if the magnetic field (input) may be measured much more accurately than the electric field (output). Except when the magnetic channel is contaminated by artificial noise, the assumption is normally used in practice. All of the quantities in (2) are complex. However, it will be simpler to consider the real case first, from which the complex analog can be deduced. The M-estimator (or Maximum likelihood type estimator of Huber, 1981) for !JA in (2) is the value of !JA which minimizes

pet) =

{

k t2

I

2

toltl-"2to

ItI S

to;

Itl~to,

(6)

with tuning constant to = 1.5. This hybrid form, which is based on a density function with a Gaussian center and Laplacian tails, results in robustness of the M-estimator for data with a "least informative" distribution (Huber, 1981). Note that the tuning constant to together with the scale parameter s determine the transition point, above which the residuals are considered to be large. For p(t) given by (6) the corresponding Huber weight function is

Wet) =

{

I

ItI S

to;

tolltl

ItI ~ to·

(7)

The scale parameter s must also be estimated robustly, for which the median absolute deviation (MAD),

where pet) is a general function, called the loss function, and S is a scale parameter. Equivalently, the estimates may be defined as the solutions of equations

n

L l\I(rils)uij = 0,

i=

I

j

=

1, 2,

(4)

dp(t)ldt, is called the influence function.

Alternatively, by defining the weight function to be Wet) =

1\I(t)lt, the above equations become

where 1\I(t) =

(5)

where W is a diagonal matrix. This alternative form of M-estimator might more properly be called a W-e stimator, but M-estimators are more familiar and the difference is minor. The W comes from "weighted," since equation (5) can be viewed as the normal equation for a weighted least-squares (WLS) regression problem. However, unlike the ordinary WLS regression, the weights Wi are neither equal nor U-determined: they depend on the normalized residual (r il s).

In general, equation (5) is not linear unless

Wet)

=

1,

which corresponds to the LS estimator. Thus iterative methods are necessary. For this, an initial solution is chosen as the starting value, and both the residuals and a scale estimate are computed. The weights are calculated from these, and the solution to equation (5) is found. This proce- dure is repeated using the residuals and the scale estimate from the previous iteration at each stage until convergence is achieved. Because the weights are data-adaptive, the influ- ence of data corresponding to large residuals will be mini- mized. There are a great number of possible forms for the loss function p(t) and thus of the W(t) to be used in the regression M-estimation. The most widely used is a "stan- dard hybrid" (Huber, 1981),

med {Iri

-

med (rdl}

s=-------- (8)

U'MAD

is one suitable choice. In the above equation, U'MAD is the expected value of the MAD for the appropriate probability density function (PDF). For the MT problem, in which the data are complex rather than real [the vectors and matrices in equation (2) are complex], the determination of the robust scale as well as iterative reweighting may be performed either in a real or a complex framework. However, the use of a complex frame- work, within which the scale and weights are determined based on the magnitude of the residuals, is preferable because it is rotationally (i.e., phase) invariant (Zeger, 1985). In the complex framework, the normal equation (5) becomes

U*WU!JA = U*Wx

-

--

-

-'

(9)

where * denotes the Hermitian conjugate. The convexity of the Huber p function makes the iterative M-estimators less sensitive to the starting value, for which the LS estimator is still a possible choice. Furthermore, this characteristic also ensures that convergence to a unique estimate is guaranteed in the iterative WLS solution. How- ever, the Huber weights fall off slowly for large residuals and provide inadequate protection against severe residuals. Therefore, more severe types of weight function such as the one proposed by Thomson (1977), need to be applied for a few iterations after convergence with the Huber weights has been achieved. The Thomson weights have the form

Wet) = exp {_e"(lt!- Ill},

(10)

where ex is a parameter that determines the scale at which downweighting begins, and is analogous to to in equation (7). Empirically, it was found that the nth quantile of the appropriate probability distribution is a suitable choice for ex (Chave et al., 1987). It should be noticed that, if outliers have been eliminated, the magnitude of residuals are the square root of the sum of the squares of the two almost normally distributed variates, and hence are Rayleigh-distributed with pdf (Johnson and Kotz, 1970)

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

M-Estimation of MT Impedance Functions

2001

(11)

The MAD for the above PDF is 0.44845, whereas its quan- tiles are given by (Chave et al., 1987)

f(t) == t(exp (-t 2 /2))

(t ~ 0).

a.

-V

=. J 2 log (

n -

.n

I

+ 0.5

)

i == 1,

, n.

(12)

PHASE SMOOTHED ROBUST IMPEDANCE ESTIMATION

The superiority of the robust impedance estimation pro- cedures over the conventional procedures (which are based on LS) was demonstrated by Jones et al. (1989), Chave and Thomson (1989), and Sutarno and Vozoff (1989). Using synthetic and real data, Sutarno and Vozoff (1989) showed that the robust procedures always yield impedance estimates which are better than the conventional estimates. However, since the underlying procedures are essentially based on the sort of model assumption that is intrinsic to the transfer function approach (that noise is restricted to the output), the procedures still need to employ the remote reference (RR) technique of Gamble et al. (1979) in order to minimize the bias effect caused by noise in the input. Unfortunately, RR is more costly than single site observations, requiring more field equipment and personnel. Biasing can still occur if there is insufficient separation between the measuring and refer- ence sites or if there are effects of possible nonuniform fields that are coherent between the measuring and reference sites. Therefore, it remains a very desirable goal to obtain unbi- ased, robust impedance estimates using only one set of equipment. It is well known that the real and imaginary parts of a causal transfer function such as the MT impedance function are not independent bur are related by the Hilbert transform (HT). Furthermore, with a minimum-phase assumption, the relationship at radian frequency Wo may be written (Boehl et al., 1977):

and

In IZ(wo)1 == -

1 foo

11'

-00

cI>(w)

--

w

-

Wo

1 foo In IZ(w)1

<I>(wo) == --

11'

-00

Wo

-

w

dw

dw.

(13a)

(13b)

Boehl et al. (1977) and Fischer and Schnegg (1980) presented methods for calculating the amplitude of the impedance and thus the apparent resistivity from the phase or vice versa, for a one dimensional response or a two dimensional magneto- telluric response rotated to its principal axes. Recently, Yee and Paulson (1988) derived such a relation from the Cauer integral representation for the impedance tensor, which is constructed based on the imposition of passivity and sym- metry requirements on the impedance tensor. In this way, they suggested that the dispersion relations are shown to be valid for the general three dimensional (3-D) earth system. As pointed out, this is because the impedance function can be associated with a minimum-phase system. It is then possible to find a consistent estimate of the MT impedance function from the phase. The basic idea behind such a procedure, namely "phase smoothing," which was first

utilized by Boehl et al. (1977), stems from the fact that only the phase function can be estimated consistently using the standard (4-channel) analysis techniques (Vozoff, 1972), whereas estimates for the amplitude are biased. The new robust procedure proposed here, incorporates phase smoothing in the regression M -estimation, The proce- dure is essentially an iteratively reweighted least-squares approach, which is similar to the one proposed by Sutarno and Vozoff (1989). Likewise, the procedure uses a complex framework. The main steps of the procedure are as follows:

(a)

Use the LS estimator for the preliminary estimator

oo(m)(m = 0).

(b)

Compute the predicted outputs

 

i~m) == u

 

i==I, ••• ,n

I

lJ t-'J

'

and residuals

r~m) = x. - i~m)

i==l, ••• ,n.

I

I

I'

(c)

Use the magnitude of residuals to obtain the scale estimator (8).

(d)

The Huber weights (7) are computed using the scaled magnitude of residuals and applied to matrix Wof the weighted normal equation (9).

(e)

The weighted normal equation (9) is solved and the entire process is repeated until convergence is achieved. The iteration formula is given by

(f)

Use the Thomson weights [equation (10)] and a fixed scale estimate derived from the above final iteration, again solving the weighted normal equation (9) and terminating when convergence is achieved.

(g)

Apply the Hilbert transform operation to compute the amplitude of the impedance from the phase and obtain the phase-smoothed estimate of the impedance at each frequency.

(h)

Replace the original estimate by the phase smoothed

estimate and perform one-step iteration toward the Thomson estimator. U) Repeat step (g) in order to find the final phase- smoothed estimate of the impedance. Note that this robust regression procedure can be used with the standard (4-channel) or the RR (6-channel) technique. To employ the RR technique, elements of matrix U* in equation (14) should be replaced by the remote referen& field values.

,

-1

FIELD EXAMPLES

MT data from the Columbia River Plateau

This MT data set was taken in an extremely noisy area in the Columbia River Plateau, by Phoenix Geophysics Inc., using a remote reference MT system. The data are very highly contaminated with common noise at the two loca- tions. It was indicated that a hydroelectric dam and alumi- num refinery which exist in this area were responsible for the noise (Sutarno and Vozoff, 1989). The impedance function example was computed for site 101 using site 102 horizontal magnetic field as a reference. The distance between the two

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2002

Sutarno and Vozoff

sites is about 3 miles. (No site map was available). Prelimi- nary processing of the data was conducted using standard MT analysis. The mean and linear trends were subtracted from each segment of 32 points. The segments were multi- plied by a Hanning window and the 6th and 8th discrete Fourier transform (DFT) coefficients were calculated for each segment. To calculate the Fourier coefficients for the output frequencies of the low bands in the low range, the cascade decimation procedure of Wight and Bostick (1986) was used. The impedance and thus the apparent resistivity and phase values were then computed using the new robust method as well as the conventional standard and RR meth-

ods. In addition, a diagnostic procedure, namely, the "nor- mal scores test" (Hettmansperger, 1984), was done in order 10 examine the error distribution resulting from the prelimi- nary fit. For this purpose a real framework was used, so that the residuals may be regarded as having independent Gaus- sian real and imaginary parts. From the diagnostic test, it is found that for frequencies higher than 0.035 Hz, most of residuals deviate from normal behavior. Figure 1 shows normal scores plots of Ex and E; residuals for two frequency values, 0.281 and 0.375 Hz. All the plots show a typical long-tailed behavior caused by outliers. The apparent resis- tivity computed by the new method, along with the results

e

::)

-e

fIJ

CI

-

w

O.

-0

~o.

F-O.28130 Hz

I

••

r

-2.0

-0.1

Normal

01

Scorn

r·o,37500 Hz

N-16Z0

••

t

<P

f

,J

2.0

F"O.Z8'30 Hz

0

~

"0

I)

lfl " ~ -0 '" ~ 0' )
lfl
"
~
-0
'"
~
0'
)

·0

v

r

Normal

f a O.37500 Hz

Scores

N"1SZO

!

)

.~

"

*

,.J

~

~

N-1620

o

~

"0

,.

W

-0

~O

r

-t.o

Normal

07

Scores

2.0

3.4

-0

-0

·3

r

I

-'l.O

-0

r

Normal

01

Scores

J

f

FIG. 1. Normal scores plots of Ex (left) and E

y

(right) residuals for MT data from the Columbia River Plateau site

WI, for frequencies 0.281 Hz (top) and lU75 Hz (bottom). Number of samples, N == 1620.

if

~."

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

M-Estimation of MT Impedance Functions

Ilf

lOS

,,----------------------------,

--~-- RMT.

- - -)( - -

--

-

REF.

M::THOD

STO.

STO. ROBUST

---e-- RR ROBUST

(a)

I~

10 1

Ie!'

10-1 -t---;r--1-rTTTTTr--r-T"'T

10-

5

.: \

.,

.,

,

TTT1-

10- 1

,-

-rMrTT1.----.-r"TTTTT

Ie!'

10

FREQUENCY (Hertz)

--

\

.(

)('

>C

'"

1

--,~""T'1rTT1rr1

I~

Ilf ~--------------------------.

lOS

10 1

- . +

- -

RMT.

REF.

--* -- STO. M::THOO

----

-

STO. ROBUST

--e--

RR ROBUST

(b)

Ie!'

10- 1 ;---;r--1-rTTTTTT--r-T"'T

ur 5

TTT1-

10- 1

,-

-rMrTT1.----.-,-"TTTTT

Ie!'

FREQUENCY (Herlz)

--.~""T'1l"'TTrrI

FIG. 2. (a) A comparison of results for Pxy on MT data from the Columbia River Plateau site 101, by four different analyses: standard, RR, robust on four-channel data (Std. robust), and robust on six-channel data (RR robust). The two robust analyses provide very smooth and consistent results. (b) As Figure 2a, but for Pvx' In addition to -

smoothness, the two

robust results are much more consistent than the other pair.

2003

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2004

Sutarno and Vozoff

from the conventional standard and RR methods, are shown in Figures 2a and 2b, and the corresponding phases can be seen in Figures 2c and 2d. Based on smoothness, slope, and compatibility of Pa and <I>, it can be seen from the figures that the conventional analyses yield impedance estimates that are very poor in these qualities. The new method markedly improves the impedance estimates. Furthermore, the proce- dure also provides more consistent impedance estimates. As shown, application of the procedure to the 4-channel data gives almost the same results as applied to the 6-channel (RR) data. This is a very powerful test of its effectiveness.

for two frequencies, 0.0012 and 0.0016 Hz. Here also, the plots show a typical long-tailed behavior. The apparent resistivity computed using the new method, along with the results from the standard and RR methods are depicted in Figures 4a and 4b, whereas their corresponding phases are shown in Figures 4c and 4d. As can be seen, the figures again show the superiority of the new method over conventional methods.

CONCLUSION

MT data from the EMSLAB Lincoln line

This field example was recorded at one of the EMSLAB Lincoln lines, by the Geological Survey of Canada. The impedance function example was computed for the "quiet period" of the MT record of the Lincoln line site 1, using site 13 (135 km distant) horizontal magnetic field as a reference. As reported (Jones et al., 1989), the MT data in this period are badly contaminated. A telluric noise spike of about 200 mVIkm and a polar substorm event dominate the data. As before, except for interpolating across a few short data gaps, no preliminary editing of data was deemed necessary on the basis of visual inspection, and the same process is conducted on the data. Figure 3 shows normal scores plots of residuals

Regression M-estimation which is robust to violations of the usual Gaussian error assumption has been applied in the processing of MT data to estimate the impedance tensors. The outlier rejection facility provided by the procedures improves the impedance estimates. When the impedance function is assumed to be a minimum-phase system, it is then possible to compute the impedance amplitude from the phase by applying the phase smoothing process, which is based on the Hilbert transform. Since the phase function can be estimated consistently from the noisy MT data, incorpo- ration of the process in the regression M-estimation solves the bias problem and results in robust, consistent impedance estimates.

~

l

;;.

i

135

---+-- RMT.

- - -)( - -

STD.

,

---------------------------.

REF.

M::THOD

I

I

,

\

Cc)

-4-

STD. ROBUST

--G-

RR ROBUST

80

I

\

\

45

O+---,r-r-rrTTTTr-r-r-rTT1rTTr-"-"T""'1r-T'1"TTT

Icr l

10'1

FREQUENCY

uP

(Hzl

'\

I

~

,I

\ /

--,--rTTTTT

---.r-r-nrTTm

Idl

-45

-80

·1

--0+-- RMT.

,

-------------------

REF.

--* -- STD. M::THOD

- -4- RR ROBUST

STD. ROBUST

-----

----------:----:---.

I

\

Cd)

,I

I 'J

\

"-,I

\ \

\

I

-/

,

\

/

-llIO-l--r"-rT1rrnr--'--'-"T'TTTT,.,--,r-r-rrTTnr-r-T""TTTrTTT--r-.,-",rrn1

10. 1

IcP

FREQUENCY (Hz I

FIG. 2. (c) (d) Phases for Figures 2a and 2b. The two robust results are very smooth and consistent.

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

F-O.OOI17 Hz

a

~

'0

.,#

-I.

M ••

LIJ

-5.

-5.

-Z.'

-1.7

-0.'

M-Estimation of MT Impedance Functions

0.'

N-168 15. • • •• ., • a • ~ ,/ '0 • • -2.
N-168
15.
••
.,
a
~
,/
'0
-2.

1.7

2.'

>.

LIJ

-7.

-Z.'

F-O.OOI17 Hz

-1.7

-0.'

0.'

N-I68

1.7

2005

2.'

Normal

Scores

Normal

Scores

F-O.OOI56 Hz

N-168

F-O.OOI56 Hz

N-168

a

~

'0

M

LIJ

-0.

-I.

-Z.

-Z.'

-1.7

'i

-0.'

Normal

0.1

Scores

,.

1.7

Z.

a

~

'0

>.

LIJ

-I.

-5.

-t.

-Z.'

-1.7

'

I

I

-0.'

0.'

NormaI Scores

0

7#'0·

1.7

2.'

FIG. 3. Normal scores plots of Ex (left) and E)' (right) residuals for the quiet period of the MT data from the EMSLAB Lincoln line site 1, for frequencies 0.0012 Hz (top) and 0.0016 Hz (bottom). Number of samples, N =

168.

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2006

Iff

Sutarno and Vozoff

--+-- RMT. REF. (0 ) - - ---- STD. ROBUST - - STD. t.£THOD ,
--+--
RMT.
REF.
(0 )
- -
---- STD. ROBUST
- -
STD. t.£THOD
,
~
--4- RR ROBUST
\
\
10'
\\
r
~
i:
,\
I:
,
e
,
~ \
J \
Irl
~.
4 > ~t I :".'~
-
'\
II
\ ,,:'\
~
.:'
\
II
\\
<,
, II
'\
\
\\ J
.rl
\
\
<,
"
10 1
10""
10"'
10'2
FREQUENCY (He,' z I
Iff
"-\i
--+-- RMT. REF.
(b)
\
STD.
t.£THOD
-
-
-)t
-
-
STD. ROBUST
RR ROBUST
\\
~
\
\
10'
\
I'
,
r
,
,
I
i
'\
e
i
Irl
"- <, --"

10"1

10' +-----.--r--r-!'"TTTrr--

10""

10"'

,.--r--r-!'"TTTrr--

FREQUENCY (He,' zI

,.--r--r-r-rTTr!

10"1

FIG. 4. (a) A comparison of results for Px on the quiet period of the MT data from the EMSLAB Lincoln line site 1, by four different analyses: standard, RR, robust on four-channel data (Std. robust), and robust on six-channel data (RR robust). Here also, the two robust results are very smooth and consistent. (b) As Figure 4a, but for Pyx' In addition to smoothness, again the two robust results are very consistent.

Downloaded 10/13/14 to 129.2.29.132. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

~Estimation of MT Impedance Functions

13II

--

,-

,----""""""T------------------------,

.

REF.

(c)

--->t"-- sro. t.£THOO

--JI.- slq. ROBUST

10

,

RR\~OBUST

\\

\

~~~

0

10-4 10"5 FREOlEHCY (Hz) ·45 --~•• RMT. REF. - - -)( - - STD. t.£THOO
10-4
10"5
FREOlEHCY (Hz)
·45
--~•• RMT. REF.
- - -)( - -
STD.
t.£THOO
STD.
ROBUST
-.A--
---lD--
RR ROBUST
·
-10
~
~
if -I
,.---
00--
-I
10. 5
10-4
FREQUENCY
(Hz)

10. 2

<,

10. 2

(d)

10"1

10'1

2007

FIG. 4. (c) (d) Phases for Figures 4a and 4b. Again, the two robust results are very smooth and consistent.

ACKNOWLEDGMENTS

We thank Bob Anderson, Alan Jones, Phoenix Geophys- ics Inc., and the Geological Survey of Canada for provision of the data and permission to release it. Don McNeil and Dave Kleinbaum of Macquarie University and Ken Paulsen of the University of Saskatchewan helped to improve the first version of the paper. EMSLAB was supported by the US National Science Foundation and the Geological Survey of Canada.

REFERENCES

Boehl, J. E., Bostick, F. X., and Smith, H. W., 1977,An application of the Hilbert transform to the magnetotelluric method: Tech. Rep. Electr. Geophys. Res. Lab., University of Texas, Austin.

D. J., and Ander, M. E., 1987, On robust

Chave, A. D., Thomson,

estimation of power spectra, coherences, and transfer functions:

J. Geophys. Res., 92, 633-648. Chave, A. D., and Thomson, D. J., 1989, Some comments on magnetotelluric response function estimation: J. Geophys. Res., 94, 14215-14225. Claerbout, J. F., and Muir, F., 1973, Robust modeling with erratic data: Geophysics, 38, 826-844.

Egbert, G. D., and Booker, J. R., 1986,Robust estimation of geomag- netic transfer functions: Geophys. J. Roy. Astr. Soc., 87, 175-194. Fischer, G., and Schnegg, P.·A., 1980, The dispersion relations of the magnetotelluric response and their incidence on the inversion problem: Geophys. J. Roy. Astr. Soc., 62, 661-673. Gamble, T. D., Goubau, W. M., and Clarke, J., 1979, Magnetotel- luries with a remote reference: Geophysics, 44, 53-68. Hettmansperger, T. P., 1984, Statistical inference based on ranks:

John Wiley and Sons, Inc.

Huber,

Johnson, N. L., and Kotz, S., 1970, Continuous univariate distri- bution -I: Houghton Mifflin Co. Jones, A. G., Chave, A. D., Egbert, G. D., Auld, D., and Bahr, K.,

1989, A comparison of techniques for magnetotelluric response estimationrJ. Geophys. Res., 94, 14201-14213. Sutarno, D., and Vozoff, K., 1989, Robust M-estimation of Magne- totelluric impedance tensors: Expl. Geophys., 20, 383-398. Thomson, D. J., 1977, Spectrum estimation techniques for charac- terization and development of WT4 waveguide 1: Bell Syst. Tech. J., 56,1769-1815. Vozoff, K., 1972, The magnetotelluric method in the exploration of sedimentary basins: Geophysics, 37, 98-141. Wight, D. E., and Bostick, F. X., 1986, Cascade decimation-A

P. J., 1981, Robust statistics: John Wiley and Sons, Inc.

technique for real time estimation of power spectra, in Vozoff, K., ed., Magnetotelluric methods: Geophysics Reprint Series no. 5:

Soc. Expl. Geophys., 215-218.

V., 1988, Concerning dispersion relations

for the magnetotelluric impedance tensor: Geophys. J., 95, 549-

559.

Yee, E., and Paulson, K.

Zeger, S. L., Exploring an ozone spatial time series in the frequency domain: 1. Amer. Stat. Assoc., 80, 323-331.