.

I l •
J Geophys (1985) 57:191-206 . ,.Journ•• of.
· Ge~physlcs
'

\ I Construction of conductance bounds from magnetotelluric impedances

~. P. Weidelt
Institut ftir Geophysik und Meteorologie, Technische lJniversitiit, MendelssohnstraBe 3, D-3300 Braunschweig,
Federal Republic of Germany

Abstract. Whereas any finite set of impedance data class of conductivity models consistent with the data.
does not constrain the electrical conductivity o-(z) at a In principle, the latter problem can be approached
fixed level z in a 1D-model, the conductance function, either by exploring the whole space of feasible models
S(z 2 ) as the depth-integrated conductivity from the sur- by Monte Carlo techniques or by explicitly construct-
face to the depth z 2 will be constrained. Assuming pnly ing the model, which extremizes the geophysically in-
the non-negativity of o-(z), it is shown that for a given teresting property. Firm foundations for the use of the
( depth z 2 the models generating the lower and upper
bound of S (z 2 ) consist of a sequence of thin sheets'. The
Monte Carlo method as a tool for geophysical in-
verstion were laid by Anderssen and Seneta (1971,
\ determination of the positions of the thin sheets and 1972) and the method of parameter extremization in
\ their conductances leads to a system of nonlinear equa- geophysical inverse problems was pioneered by Parker
tions. As a limitation the present approach requires the (1972, 1974, 1975).
existence of a model, which exactly fits the dat'I. The Any set of magnetotelluric impedances for a finite
structure of the extermal models as a function 0£ z 2 is number of frequencies does not impose bounds on the
discussed in examples with a small number of frequen- conductivity o-(z) at any fixed depth z. At this depth,
cies. Moreover, it is shown that any set of complex ID either a thin insulating sheet or a sheet of unbounded
impedances for M frequencies can be represented by a conductivity, but of finite conductance (conductivity
partial fraction expansion involving not more than 2M thickness product), may exist. However, conductivity
(positive) constants. For exactly 2M constants there are averages over a finite depth range will, in general, be
two complementary representations related to the lower constrained by the data, provided the field penetrates
and upper bound of S(z 2 ). down to this depth. This has been exemplified in detail
For the simple one-frequency case, a more general by Oldenburg (1983), who constructed bounds on con-
extremal problem is briefly considered, where the ad- ductivity averages by linearizing the nonlinear problem.
mitted conductivities are constrained by a priori The existence of bounds for conductivity averages or
bounds o-_(z) and o-+(z) such that o-_(z);';o-(z);';o-+(z). for the simpler conductance function
In this case, the extremal models for S{z 2 ) consist of a
sequence of sections with alternating conductivities ,,
er_ (z) and er+ (z). The sharpening of conductance S(z 2 ) ~ Jo-(z)dz (1.l)
bounds by incorporating a priori information is illus- 0
trated by an example.
reflects the fact that the inverse problem for S(z 2 ) is
Key words: Electromagnetic induction - Inverse prob- well-posed (VJ. Dmitriev, private communication).
lem - Extremal models The present study is centered on the computation of
bounds for S(z 2 ), imposing apart from the non-nega-
tivity condition o-(z)~O no further constraints on the
conductivity. Any model consistent with the data must
I. Introduction lie within these bounds. The problem under consider-
ation resembles the problem of discovering extremal
The lD-magnetotelluric inverse problem is known to models of linear functionals of the density from a trun-
be ill-posed and thus allows the construction of a cated set of eigenfrequencies of an elastic string, as
whole set of satisfactory conductivity models from a treated by Barcilon (1979), Barcilon and Turchetti
given real data set. The COPROD study of Jones (1979), and Sabatier (1979). Briefly addressed is also the
(1980) provides a good illustration of this fact. At pres- more general problem of computing bounds for con-
ent there are two attempts to overcome the problem: ductivity averages rather than for S(z 2 ).
either the inversion is stabilized by incorporating We shall consider only for the simple one-frequency
known or assumed properties of the conductivity struc- case the construction of bounds for S(z 2 ) when o-(z) is
ture as a priori constraints, or one may try to extract constrained by a priori bounds o- _(z) and er +(z) such
geophysically useful properties pertaining to the whole · that er_ (z) ;<; o-(z) ;<;er+ (z).

I
 ~ - -'

192
·,
Contrary to the pragmatic approach of Oldenburg or simpler for a given z 2 >0 the conductance function (
(1983), who applies his approximate method to a large ,, L
number of frequencies, attention is confined in this very
preliminary study to the exact extremal models for a
S(z 2 ) = Ja(z)dz. (2.4b)
0
small number of impedances. We also have to assume
that there is a model that fits the data exactly. This These four cases are equivalent to the problem of mi- / '·
restriction, however, may be dropped in subsequent nimizing the objective function
work, thus extending the range of applicability to real 00
inconsistent data sets. Only in connection with the
COPROD study are approximate extremal models for
Q[a] = Jw(z)O'(z)dz (2.5)
0
many real data considered.
Although the present problem is one of the simplest with the weights
extremal pro bl ems in electromagnetic induction, the
nonlinearity introduces a great deal of complexity and 0, O~z<z 1
leaves the treatment still in an experimental stage. An- 1
, z1 ~Z~Zz for Q =O'mi"
other relatively simple extremal problem in electromag- · w(z)= Z2 - Z1 (2.6a)
netic induction can be solved for two-dimensional per- -1
fect conductors (Weidelt, 1981). , z 1 ~z~z 2 for Q =-a=max
The main part of the paper consists of three. sec- Z2-Z1

tions. The general structure of the extremal models is, or

l
derived in Sect. 2. Then a detailed discussion and illus- 1, O°"z°"z 2 for Q=Smin
tration of unconstrained extremal models is give,n in w(z)= -1, O°"z°"z 2 for Q=-Smax (2.6b)
Sect. 3, and the concluding Sect. 4 is devo.ted to simple Q, z>Z2
extremal 1nodels with a priori constraints on conduc-
tivity. The paper has two appendices, where in particu- Assuming that a(z) is only constrained by the non-
lar Appendix B contains the proof of an impedance negativity condition, we are faced with the nonlinear
representation theorem, to which we have to appeal in programming problem of minimizing Q [a] subject to
Sect. 3. the constraints
C;[<r]=C;, j=l,. .. ,M (2.7a)
2. Necessary conditions for extremal models
Assuming SJ units, a time factor eiroi, a lD-conduc- and
tivity structure cr(z), and neglecting displacement currents, a(z);;;O. (2.7b)
Maxwell's equations reduce for a quasi-uniform incident
magnetic field in the y-direction to The equality constraints (2.7 a) render the problem non-
linear. The Lagrange function is
E~(z, w)= -iwµ 0 H,(z, w), H;(z, w)= -u(z)E.«z, w),
(2.1) M oo
implying L[a]=Q[a]+Re I .<,{c;[<r]-c;J-J µ(z)O'(z)dz, (2.8)
i= 1 0
f"(z, w)=iwµ 0 u(z)f(z, w), (2.2)
where the M complex Lagrangian multipliers .l; enforce
where f (z)' = E_,(z, w). In the sequel we use the transfer the equality constraints, whereas the non-negative func-
function c introduced by Schmucker (1970). Its theoreti- tion µ(z) takes the inequality constraint into account
cal value for given cr(z) at a set of M frequencies m;, and satisfies
j = 1, .. ., M is defined as
µ(z)=O, a(z)>O; µ(z);;;O, a(z)=O (2.9)
f (0, w;)
(2.3) (e.g., Pearson, 1974, p. 1113). The formally introduced
f'(-0, w/ Lagrangian multipliers A; can be interpreted as the
where f(z) is a solution of Eq. (2.2), with f'(z)-->0 for sensitivity of the minimum value Q 0 of the objective
z--> oo. In Eq. (2.3), provision is made for a possible functional Q [a] to changes in the constraining data
discontinuity of H, or f' at z=O due to the presence of c;=g;-ih; (e.g., Pearson, 1974, p.1118):
a thin conducting sheet. The M complex data ci' which
arc assumed for the present to be exact, correspond to (2.10)
the functionals c;[cr]. Then the problem of interpre-
tation consists in finding at least one model a(z) such
that c;[a]=c;, j=l,. . .,M. When there is no risk of Hence, the .l; allow immediately a rough estimate of
confusmg functionals and data, [a] is omitted. the influence of data errors. Also Jt(z) can be inter-
Within the class of models fitting the finite data set preted as the sensitivity of Q0 by changing the lower
we are interested in those two models, which minimize bound of cr(z) in the neighborhood of z to a positive
and max1m1ze either for a given depth range value: let a(();;;iT_(() in z-Llz/2°"(°"z+Llz/2 and let
z 1 ~z~z 2 the conductivity average
1 ,, z+ LJz/2
iJ(z 1 , z 2 ) = - - J a(z)dz (2.4a) LI<= J a_(()d(.
Zz-Zt zi z- LJz/2
r:
)

('
Then we obtain in the limit LI z-> 0
oQo
models consist of a sequence of insulating layers (cr=O)
and thin conducting sheets, where cr(z) is positive only
at an isolated point.
193

µ(z)=fh, (2.11)
The problem of model construction consists in find-
) i.e. Q 0 is not affected, if the constraint is not binding
ing for a given impedance set a sequence of thin sheets
and a set of complex Lagrangian multipliers such that
) (cr(z)>O, µ(z)=O) and Q 0 will not decrease for a bind-
ing constraint (cr(z) = 0, µ(z);;; 0). These results, of
the model fits the data and the function D(z) is non-
course, were expected.
negative everywhere and in particular vanishes at the
(\ As a necessary condition for an extremum of Q, the
positions of the thin sheets, i.e.,
first variation of L with respect to er has to vanish. This D(z);;;O; D(z)=O for cr(z)>O. (2.18)
/ \
yields
M This is clearly a nonlinear problem, since both the data
w(z)+Re L A;F;(z)-µ(z)=O, (2.12) functionals and the Frechet derivatives depend non-
linearly on the positions and conductances of the thin
i= 1 , sheets. Also, it is not yet clear from the outset how
where F;(z) is the Frechet derivative of the functional many sheets have to be considered. (Semi-empirical ru-
,i cJcr], defined by les' derived from experience with a small number of
w frequencies are listed in Sect. 3.6.) Concerning the non-
lic;[crJ = J F;(z)ilcr(z)dz (2.13) negativity of D(z), closer examination in Sect. 3.3 shows
0
that for z<z 2 we have D(z)>O, for cr(z)=O, and D(z)
with =0 for cr(z)>O, whereas D(z) vanishes identically below
F;(z)= -iwµ 0 Jj 2 (z) (2.14) the first sheet occurring in z > z 2 .
In a more general approach, the incorporation of a
(e.g., Parker, 1977), where f;(z)•= f(z, w;) is the solution priori bounds er_ (z) and er+ (z) such that
of Eq. (2.2) with Ji'(-0)=1 and fj(z)->0 for z'-+w. cr_(z);;:;cr(z);;:;cr+(z) can be achieved by replacing the
The lowest-order differential equations satisfied by F;(z) last term in the Lagrange function Eq. (2.8) by
are w

(2.15a) -J [µ+(z){cr+(z)-cr(z)}+µ_(z){cr(z)-cr_(z)}]dz,
, 0
(2.15b) where µ+(z);;;O. Then the definition of D(z) in analogy
On using Eq. (2.12), a function D(z) is defined as to Eq. (2~16) is
M D(z)=µ_(z)-µ+(z). (2.19)
D(z)•=w(z)+Re L A;F;(z)=µ(z). (2.16) D(z) vanishes, if cr(z) attains neither its lower nor its
i= 1
upper bound. For the one-frequency case it was shown
Now assume that in some interval a<z<b, completely above that D(z) cannot vanish in a finite interval.
inside an interval where w(z) is constant, the conduc- Therefore, in this case the extremal models consist of a
tivity cr(z) is positive, i.e., cr(z)>O for zE(a,b). Then stack of sections with alternating conductivities er_ (z)
from Eq. (2.9) µ(z)=O for zE(a, b) and Eq. (2.12) reads and er+ (z), satisfying
D(z) =0, zE(a, b). (2.17)
cr(z) =er+ (z)} :s:o
for D(z) { - .
In Appendix A it is shown that for downward diffusing (2.20)
cr(z)= er_ (z) ;;;o
fields F;(z) this equation has no solution, except for the
trivial solution A;=O if w(z)=O for z>a. For M=l this Jumps between er_ and er+ occur, where D(z) changes
can be verified easily: since Eq. (2.17) holds in a whole sign. The generalization to M frequencies must still be
interval, it can be differentiated an arbitrary number of done, but it can be anticipated from the approximate
times. In particular, the first two derivatives at zE(a, b) extremal models presented by Oldenburg (1983) in his
read (skipping the subscript 1) Fig. 7 that also in this case only the conductivities er_
and O" + will occur.
D'= -2Re{2F/c}=0,
D" =" +Re {AF(iW/loCT + 1/c 2 )} =0
3. Extremal models for the unconstrained
with c=c(z)= -f(z)/f'(z)=g-ih, g>O, h>O for a conductance function
downward diffusing field. The above equations can be
considered as two linear homogeneous equations for 3.1 Formulas for a series of thin sheets
Re (J.F) and Im (J.F), admitting a nontrivial solution
only for a vanishing system determinant LI. However, The discussion is started with the unconstrained ex-
tremal models requiring only the assumption cr(z);;;O,
ILI I= 4(h + W !lo erg lcl 2 )/lcl 4 > 0. and attention is confined to the conductance function
The linear independence of the Frechet derivatives has in Eq. (1.1)
the important consequence that the extremal models
cannot comprise conductivity sections, where a(z) is S(z 2 ) = "Jcr(z)dz (3.1)
positive over a finite interval. Therefore, the extremal 0
 '.,,,. i~' ~; ' ] ·, ·-

194
~
rather than considering conductivity averages. In the ro a(Jc)d?c (
unconstrained case the extremal models reduce to a c[rr] =a 0 + J-,-.-,
0 A+l(l)
a0 ;o=O, a(J.);o=O, (3.8) \ '

series of thin sheets. Models of this kind have been

identified previously by Parker (1980) as giving the where a(Jc) is a generalized function to include both the
best fit to any real data set. discrete and continuous part of the spectrum (Weidelt,
Formulas for the treatment of a series of thin sheets 1972; Parker, 1980; Parker and Whaler, 1981). As an
are briefly summarized. Assume a stack of K sheets example, the stack of thin sheets considered in the
with conductance r, at depth (,, i.e., previOus section has a finite discrete spectrum and
K
leads to the representation
/
a(z)= L r,c5(z-(,). K-1 a a
k=l c[rr]=( 1+ L - b '. +:-"-;
k=l k+iw iw
The inter-sheet separations are d, = (, -(, _ 1, k =
2, ... , K. Then the solution f (z) of (2.2) varies linearly (3.9)
between two sheets, is continuous at z=·(k, but shoyvs Let the theoretical impedances C;[rr] for M distinct
the discontinuous slope ' . frequencies w; be given by ,
•' N
f'((,+O)-f'((,-O)=iwµ 0 rJ((,), k=1, ... ,K (3.2)
c;[rr]=a 0 + "'
L.., au. (3.10)
f(z) is constant in z;o:(K, i.e. f'((K+O)=O. The II= 1 bll + l(l)j
theoretical transfer function c[rr]= -f(O)/f'(-0) is ob- with a 1"b11 >0, n=l, ... , N-1; a 0 , bN~O, where the b 11 ,
tained as c[rr]=c1 +( 1 with c,,= -f((;)ff'((,-0) re-' n = 1, ... , N are distinct. [This form rather than Eq.
cursively from (3.8) is chosen for ease of presentation only, for equiva-
ck+t +dk+1 lent integral analogs based on (3.8) exist for all for-
c,= , k=K-1, ... ,l, (3.3) mulas involving the summation over n.] Then for N ;o: M
1 +iwµ 0 r,(c,+ 1 +~,+ 1 ) c;[rr] allows the two representations
starting with cK=l/(iwµ 0 rK). {In this section the sub- M A
script k on c denotes the value of cat z=(,-0, where- I: ci[a]= L 111
• , (3.11 a)
as in a different context the subscript j specifies the m=l Bm+iwi
particular frequency ro; of c at z = -0.) The value of f - M-1 A A
at z = (,, normalized to f' ( - 0) = 1 is , II: c;[rr]=A 0 + L "'.
+~. (3.11 b)
m= i Bm + iwi iwi
f((1)= -c1,
k where all constants Am, B111 , A0 , A111 , B111 , and AM are
f(C,l= nc1l TI2 Cll ~d11 , k> 1, (3.4) positive. For N < M there 1s no representation other
II= than Eq. (3.10). [See also Parker (1980).] The proofs are
given in Appendix B.
and the values of f'((,-O)=f'((,_ 1 +0) are deter- Model I is a series of M + 1 thin sheets: the first
mined from sheet is at z = 0 and the last sheet has an infinite con-
f'((1-0)=l, ductance. Complementary mod"-! II consists of M thin
k-1 sheets: the first sheet lies at z=A 0 >0 and the last sheet
f'((,-0)= TI (1-iroµ 0 r,,c,,), k>I. (3.5) has a bounded conductance. Between models I and II is
II= 1 a duality relationship in the sense that they can be
The partial derivatives of c with respect to the model transformed into each other by replacing a sheet of
parameters r, and (, are obtained by means of Eqs. finite (infinite) conductance .by an insulating layer of
(2.13) and (2.14): finite (infinite) thickness, and vice versa.
Assuming the ordering b11 >b 11 + 1 , B111 >B111 +i• we in-
oc[rr] . 2 fer from (B-19, 20), (B-24), and (B-26) the inequalities
-
0
-=F((,)= -iwµof {(,), (3.6a) M N
ur,
L (A,,JB,,,);::; a0 + L (a,Jb,,), (3.12a)
111=1 II= 1
oc[a]-=rJ'((,) = [f'((,-0)]'
a(, M N
-[f'((,+O)]', k=l, ... ,K-1 (3.6b) L Am;£ L a11 (ao=O), (3.12b)
111= l II= 1

O~~:J =irKF'((K-0)= [f'((K-0)]2, (3.6c) A 0 ~ao, (3.12c)

AM;o=aN (bN=O). (3.12d)
where
These inequalities express the extremal properties of
F((,)=i [F'((,-0) + F'((, + O)]. (3.7) models I and II, with the following physical interpre-
tation:
a) The depth of a perfect conductor (if present) is
3.2 An impedance representation theorem given by
The theoretical transfer function of a ID conductivity zro = lim c(ro). (3.13a)
distribution admits the spectral expansion w~o
 195

""Hence Eq. (3.12a) implies that I is the model with the The above considerations refer to theoretical or syn-
shallowest perfect conductor consistent with the data. thetic data, for which the existence of representations

~
Its position marks "the limiting depth below which (3.8) or (3.10) is granted. As mentioned above, real data
nothing can be learnt about the conductivity" (Parker, can be approximated by an expansion (3.10) by qua-
1982). dratic programming. However, the construction of ex-
b) The conductance of a surface sheet, the presence tremal models for the "cleaned" data is meaningful
of which requires a 0 = 0, is given by only for N"?, M, since for N < M those modified data
can be interpreted by only one model.
1

<
I
)
I
S( +0)= Jim
w-w iwµ 0 c(w)
.

Hence Eq. (3.12b) says that model I has the greatest

surface conductance.
(3.13b)
3.3 General properties of the extremal models for S(z 2 )

In the absence of a priori constraints, the extremal

c) The depth to the first conductor is models for S(z 2 ) [Eq. (3.1)] consist of a sequence of
thin sheets. The necessary extremal conditions (2.18)
z 0 = Jim c(w). (3.13 cl
{))-> 00 provide no information about the required number of
sheets as a function of both z 2 and the number M of
Therefore according to Eq. (3.12c), model JI is the con- frequencies. However, the extremal properties of
ductivity structure with the deepest top. models I and II considered in the previous section give
d) The total conductance in the absence of a perfect the first hint. Let Sm,.(z 2 ) denote the upper bound of
conductor (bN=O) is , S(z 2 ) and let z00 =min zw = L; (A,,,/B,,,). Then the ex-
tremal models for both Sm,.( +O) and Sm,.(z00 ) are given
1 by the (M + 1)-sheet model I [extremal properties a)
S(w)= Jim . (3.13 d)
ro-o iwp 0 c(w) and b)]. Similarly, let Sm;,,(zJ signify the lower bound
of S(z 2 ) and z0 =maxz 0 =A 0 • Then Sm;,,(z 2 )=0 for
Hence, Eq. (3.12d) implies that model II has the least z 2 ;o;z0 and the M-sheet model II is the extremal for
total conductance of all models fitting the data. both S,,,;"(z0 ) and Sm;,,(oo) [extremal properties c) and
The presentation of Eqs. (3.11 a, b) also clarifies a d)]. Thus, S,,,"'(z 2 ) and Sm;,,(z 2) have models I and II,
situation described by Parker (1980): the best fit to a respectively, both at the beginning and end. Al-
set of M complex measured impedances by an ex'.pan- though it has not been proved generally, numerical
sion (3.10) is obtained when using quadratic program- experiments have shown that Smax(Smin) returns to
ming. This yields a representation with at most N = 2M model I (II) also at intermediate depths z 2 = (k + 0 (z 2
terms. In the case N > M (which may occur at least for =(,-0), where (, is the position of a thin sheet in
small M), the expansion (3.10) would require more con- model I (II). Hence these models form the backbone in
stants than available independent data. The above rep- the evolution of the extremal models for smax and smin
resentation theorem overcomes this unsatisfactory situ- as a function of z 2 .
ation by compressing Eqs. (3.10) to (3.lla) or (3.11 b) To study the continuous deformation of the ex-
with exactly 2M constants. It the quadratic program- tremal models for varying z 2 , the necessary condition
ming procedure requires N"?, M terms, the best fitting (2.18) is discussed in more detail. In the unconstrained
model is nonunique. case the sections with u(z)>O, implying D(z)=O, shrink
For a given representation (3.10), the condensed ver- to one point. This means that D(z) must be either
sions (3.11 a, b) can be obtained as follows: for model I, identically zero in the neighborhood of this point or it
one first either has to solve the M-dimensional non- must show a "double zero" at the position z = (k of a
linear system (B-13) for the M constants B,,., m = thin sheet not coinciding with z=z 2 • Since D(z) may
1, ... , M, or one determines B111 alternatively as the M have a discontinuous slope at z =ck, the necessary con-
roots of the polynomial (B-9) with K = M, qr>"' q., and ditions demand
qM=l, where the coefficients q., k=l, .. .,M-1 are the
solution of the linear system (B-8b). With the knowl- D((k)=O, D'((k-O)+D'((,+0)=0, (3.14a, b)
edge of B111 the constants A 111 (or G J are immediately
11 or using (2.16) and (3.7)
obtained from_(B-16) or (B-18). Similarly, for model II M
the constants B,11 , m=1, .. .,M-1, are obtained, for in-
stance, as the solution of the first M -1 equations of
w((k)+Re L: ,!iF;((,)=0, (3.15a)
i= 1
the system (B-22) (note b 1 ?:,B,,,?:,bN); A0 then follows M
from the M-th equation of (B-22) or from (B-24), and Re L ,!ili;((k)=O (3.15b)
finally A,,, (or G,,,) is given by (B-25, 26). For the so- i=1
lution of the various systems of nonlinear equations with w(z) defined in (2.6b). Multiplying (3.!Sb) by '"
considered in this paper, Brown's algorithm turned out the conditions (3.15a, b) are, on account of (3.6a-c),
to be extremely useful, as it does not require the pro- equivalent to
vision of partial derivatives. A FORTRAN program is
published in Brown (1973). After the determination of M 8c1[a]_
the constants in Eqs. (3.lla, b), the parameters of the Re L: ,!i- - - -w((,),
8 tk
(3.16a)
j=1
conductivity models are found on using the stable Rut-
M 8cj[a]
ishauser algorithm as described by Parker and Whaler Re L ,!i~=O. (3.16b)
(1981). i= 1 °"k
 196
/
Assuming a model with K thin sheets, the condition where lcl is related to the apparent resistivity p, and I
(3.16b) also holds for an ultimate perfect conductor 'K apparent conductivity a, by '
=OO at z=(K, where F;((K)=O=Fj((K-0), but
p,=wµ 0 lc12, a,=1/p,, (3.21)
. le2<K F'(Y
1tm 0)- ocj[a] _[f'(Y 0)]2 *o. and i/l is the complement to the phase <fl between the
j ,K- -~- j ,K-
fK-> 00 ":.K electric and magnetic. field, <P+i/1=90°. The data is
consistent with a lD model if g;;;O, h;;;O, or equiva-
Now it is shown that the extremal model for Sm,,{z 2) lently O;,;i/f;,;90°. The results of the last section have
always has a conducting sheet at z 2 -0, which is just already shed some light on the structure of the ex- /
(
included in the range of integration, whereas smin(z2) tremal models. The associated models I and II [Eq.
has a conducting sheet at z 2 +0, which is just excluded. (3.11 a, b)] using the notation of Sects. 3.1 and 3.2 are
It is assumed that z 2 lies between sheet p-1 and sheet
p, i.e., (v_ 1<z 2 <(r In between, sheets D(z) can vary at . A1 (2
I: c[a] . (3.22a)
most as a second degree polynomial since J;(z) is a B 1 +1w 1+iwµ 0 -r 1 C2 '
linear function in z. The discontinuity of w(z) between . - /1 1 l
(,_ 1 and Cv introduces in this range in addition a II,: c[a]=A 0 +~='1+. (3.22b)
discontinuity in D(z), but not in D'(z): rw zwµ 0 -r 1
Equating with Eq. (3.20), it is found that model I con-
+1 for Sm,,(z 2) sists of a thin sheet with conductance r 1=lcla,sini/l at
D(z 2 +0)-D(z 2 -0)= (3.17),
{ -1 for smin(z2)' z=O and a second sheet with <2 =oo at z 00 =lclsectf!,
D'(z 2 +0)-D'(z 2 -0)=0. \3.18) whereas model II consists of a single sheet <1=
lcla,/sin t/! at z 0 =lei cost/!. Hence,
First, it is shown that the conditions (3.14a, b) in con-
nection with the quadratic variation of D(z) lead (o the s..,,( +O)=lcla,sint/!,
conclusion that D(z)=O for z;;;(,. From f;(z)=const. s..,.Cz2)=00, z2;;;lclsect/!, (3.23a)
for z;;;(K and D((K)=O follows D(z)=O for z;;;(K. With
D'((K+O)=O it is inferred from Eq. (3.14b) that Sm;,,( oo) =lei a,Jsin i/l,
also D'((K-0)=0. Hence for p<K we have D(z) Sm;,,(z 2)=0, z2 ;,;lcl cost/!. (3.23b)
=AK(z-(K) 2 in (K- i ;,;z;,;(K, where D(z);;;O implies The models of the start and end are identical, with the
AK;;;o. However, AK=O, since D((K-1l=O on account of difference, however, that at the end the (lower) sheet is
Eq. (3.14a). Therefore, D(z) =O in (K- i ;,;z;,;(K. Repeat- included in S. Now the evolution of the models be-
ing the arguments, it is found that D(z)=O for z;;;(£. tween these limits has to be studied. Starting with Sm,,,
In (,>z;;;z 2 +0 we again have D(z)=AP(z-(,) 2 , AP;;;u. it is assumed that the extremal model for z 2 > 0 evolves
With reference to Eq. (3.18), D'(z) is a continuous linear from that of z 2 = +O by moving the surface sheet to
function in (,_ 1 ;,;z;,;(P with D'((p-0)=0, D"((P-0) ( 1= z 2 -0 > 0. Then also the second sheet with <2 = oo
;;;o, implying D'((p-l +O);,;O. On the other hand, will move. The two free parameters of the model are <1
if (p-l +0<z 2 <(,-0, then the two conditions D((P_ 1) and ( 2 , whereas ( 1 and <2 are preassigned. From Eqs.
=0 and D(z);;;O would require D'((p-l +o);;;O. This (2.7a) and (3.16a, b) results the system
does not contradict the preceding result only if
D'((p-l +0)=0, implying A,=0 and D(z)=O in c[a]=c, (3.24a)
(p-l ;,;z;,;( 1, [because of the continuity of D'(z)]. The
0
condition (3.17), however, requires that D(z)$0 in this
interval. The remedy is to take either z 2 =(p-l +0 or z 2
Re{}, ~;~]}=1, (3.24b)
= ( P -0. In this case a thin sheet lies at the preassigned
, depth z2 ±0, which is no longer a freely variable pa- Re{-< oc[a]}=O (3.24c)
rameter and hence is not to be included in the neces- 0(2
sary conditions (3.14b)-(3.16b). In view of Eq. (3.17), (.< 1=' A), where the partial derivatives of c [a] are de-
the case z2 =(P_ 1+O applies to s..,, with termined from (3.6a-c). The system (3.24a-c) consists of
D(z 2 -0)=D(( 1,_ 1)=0, D(z 2 +0)=1, (3.19a) four real equations for the four real unknowns T 1 , C2 ,
Rd, and Im.<. In practice, first Eq. (3.24a) is solved for
and the case Z2=Cp-O pertains to smin: <1 and ( 2 to obtain S,..,,(z 2)=< 1 , and then Eq. (3.34b, c)
is solved for the sensitivity measure A, which is re-
D(z 2 +O)=D((,)=0, D(z 2 -0)= 1. (3.19b) quired, for instance, in the discriminant function (2.16)
Since the evolution of the extremal models for varying D(z) = w(z) +Re {AF(z)}
z2 is quite involved in the general case, it will be studied
in detail only by two relatively simple examples for one with F(z),=F1(z), as defined in Eq. (2.14). This non-
and two frequencies. negative function has to vanish at z = ( 1 and z = ( 2 ,
where D(( 1)=0 is equivalent to Eq. (3.24b) and D(( 2 )
=0 is satisfied because of F(( 2)=0. With the present
3.4 Extremal models for one frequency two-sheet model the condition D(z);;;O can be satisfied
The data for the frequency w 1 =' w is c1=' c with only for sufficiently small z2 . This is a consequence of
D(( 1)=D(z 2 -0)=0 and the fact that D(z) is a second-
c=g-ih=lcle-••, (3.20) degree polynomial in o;,;z;,;z2> which for some limiting
 197
v
value z 2 = zn vanishes at z = 0 and thus signals the em- Again, Eq. (3.27.a) is sufficient to determine t 1 and t 2.
ergence of a third sheet at z=O. To determine z the The two-sheet model with the fixed surface sheet has to
system (3.24a-c) is augmented by the condition"'D(O) be modified, when ( 1 can also be considered as a vari-
= 0. This condition and (3.24 b, c) are linear in the two able, i.e. if, in addition to Eq. (3.27 a-c), the following
unknowns Re,\ and Im,\ and are compatible only if holds
linearly dependent. Using F(( 1)=8c/8t 1 , the compati-
bility condition is
(3.27 d)
lm{[F(( 1 )-F(O)] r~~:]}=O. The augmented system (3.27 a-d) allows the determi-
Expressing F and c[er] in terms of t 1 , ( 1 (=z 2 -0), and nation of the limiting value z 2 =z,, where the surface
( 2, this condition reads sheet starts moving. The condition of linear dependence
of (3.27b-d) leads to
2
((1 -C2J (1 + w µ~<f Cf)= CJ,
2

rm{8c[er] ;ac[er]I }=o,

a~.d writin!l <1 and ( 2 in terms of the data, we end lip
with a cubic equation for z(l=( 1 as limiting value of.z 2 :'
. 8<2 ac1 ,,~o
or·
z.;'-4gz; + (2g 2+ 3 lcl 2)z,, -2g lcl 2 =0. (3.25)
+2tz=W µ~t 1 t~(~.
2
Tl
Of interest is the root z,, in O~z.. ~g=lcl cosijJ.
Renumbering the sheet parameters in the three- . The expression of t 1 and t 2 in terms of the data then
sheet problem, , the determination of Smax (z 2)=t 1+< 2 yields again a cubic equation for ( 2 =zi:
for z 2 > za requires the determination of t 1 , -r 2 , ancJ ( 3
)
from the set of equations 4g2 z~ -4g(lcl2 -g2)zt

c[er]=c, (3126a)
+ lcl 2(lcl 2 -4g 2)z, -lcl 4 g = 0. (3.28)
\.
2
Of interest is the solution z,~g with z,--+ lcl /(2g) for
R e {'A 8c[er]}= 1, (3.26b) Z;--+ w. For z 2> z, the system (3.27 a-c) is augmented by
8<1
8
8
~~:]}=1,
Re{,l c[er]}=O
Re{,\ (5.26c) 8(1
and the unknowns T1, t 2 , and ( 1 are determined from
Re{,\ 8c[er]}=O (3.26d) Eq. (3.27 a) and
ac, ·
In practice, first t 1 , -r 2 , and ( 3 are computed from rm{8c[er] /8c[er]}=o.
(3.26a) and the compatibility condition 8r 2 8( 1
The surface sheet, starting movi~g for z 2 =z 1, ap-
Im{[oc[er] _ ac[er]] ;ac[er]}=o, proaches z=g for z 2 ~oo and attains the conductance
a, 2 a, 1 8(, <1 = lcl er,,/sin ijJ. In this limit the lower sheet disappears
and ,\ is obtained in turn from two of the three equa- at infinity, and we return to the starting one-sheet
tions (3.26b-d). model.
With the emergence of the third sheet, the perfectly Figure 1 shows z,,(1/1), z,(1/1), and z 00 (ijJ). It also ske-
conducting ultimate sheet, which was first moving tches the type of extremal models for the various ran-
downwards, is moving upwards again and merges for ges of z 2 . The depth z* = g =-lei cos ijt is the "center of
z2--+z 00 =lclsecijJ with the sheet at z 2 -0, which is get- gravity" of the induced currents. Unless there is only
ting increasingly conductive. The conductance of the one sheet, conductors must be present both above and
surface sheet has meantime increased to Icier,, sin ijJ, and ?elow z*. The extremal models are not unique for Smax
we return for z 2 = z to the conductivity model for z 2 = Ill the range z 2 >zco (Smax=OO) and for Smin in the range
00

+o. 0 ~ z 2 < z 0 = g (Sm, 11 =0). The position of conducting

The extremal models for Smin(z 2 ) evolve similarly. sheets for ijJ = 45° is depicted in the left part of Fig. 6 as
As mentioned earlier, S,.. 1,,(z 2)=0 for z 2 ~g=lcl cosijJ. function of z 2 .
For ~ 2 > g we try a model consisting of a surface sheet Figure 2 displays the curves bounding S(z 2) for dif-
of conductance t 1 at z = ( 1 = 0 and a second sheet of ferent phases ijJ. S is normalized by the conductance
conductance t 2 at z=( 2 =z 2 +0. The unknowns are t 1 lcl er,,= l/(W/lo lei) [Eq. (3.21)]. All models compatible
and <2. The system of equations analogous to (3.24a-c) with the two-data problem must fall into the shaded
lS areas. These areas get narrow both for ijJ--+ 0° and
1/1 --+ 90°, reflecting the fact that these two limiting pha-
c[er] =c, (3.27 a) ses can be interpreted by one model only, consisting,
respechvely, of a perfect conductor at z = lcl and a thin
Re{,\ oc[er]}= -1, (3.27b) sheet of conductance lcl er,, at z=O. The limiting values
8r, of Smax and Smin as a function of i/l are given in Eq.
(3.23 a, b).
8c[er]}
Re { ,\~ =0. (3.27 c) As a particular feature of Fig. 2, note that the sha-
ded areas for ijJ and 90° - ijJ are mirror images obtained

1
198

/
,,- ~z­
\
l._. lcl ''° ¢
- 1.6

1.4
2
t'

..._ -..,,
o~
1
-·--··Z •
Zr~~-
..
I.'

1.0
o~
l Z2
TCior~~~~~~~~
2 3
3
0 2 3
'1
I

0
' ~
1"·30° "'·60'
. 0.8

',''J,:_:"''°'. O.o 1 "_ ,,, "''

o-1 '\\ 0-
''-=,.
2

\\ l
OA
I
\
''i:~ 0.2
3
1.

0
0° 30° 60° - t/i 90~ 0° 30° 60° - ; ¢ 90°
1"·45°
Fig. 1. The different types of extremal 1nodels in the uncon-
strained one-frequency case as function of the phase i/J. In I
Smax a surface sheet emerges for z 2 = za(t/t) (Lower solid curve at '
left) and in Smln the surface sheet is detached for z 2 =z;(l/t)
(upper curve at right)
2
I

by reflection at the main diagonal. This is an ex-

pression of the duality transformation (Weidelt, 1972), 3
o 2 3
which transforms Eq. (2.2) by means of Fig. 2. The bounds for S(z 2 ) in the unconstrained one-fre-
quency case for different phases tjt. The mirror symmetry of
the shaded domains for ifr and 90° -if! is explained in the text
z, = f <J(\)j<J 0 d(, a(z} = <Jgj<J(z), (3.29a, b)
0
3.5 Extremal models for two frequencies
](z)1=f'(z) (3.29c)
For more than one frequency various different si-
into tuations may occur in the evolution of the extremal
models, and the general discussion becomes cumber-
]"(z)= iwµ 0 a(z)](z) (3.30)
some (or even impossible) ..For this reason, attention is
with confined in this section to the exemplary study of a real
data set for M = 2 frequencies, consisting of the es-
c(w) = - J(0)/]' (0) = 1/[iw Jlo<f 0 c(w)], (3.31) timates of the transfer function c for the first and fourth
Sq harmonic for Europe as given by Schmucker (1984):
where a 0 >0 is an arbitrary reference conductivity.
1 cpd: c 1 =(575-i260) km
Equations (3.29a, b) and (3.31) imply
4cpd: c 2 =(290-i275)km.
S(z)=<J 0 z, S(z)=<J 0 z, i/!+<f/=90° By quadratic programming on using the program
NNLS of Lawson and Hanson (1974) [as suggested by
and admit the following interpretation. When determin- Parker (1980)], it is found that the data can be repre-
ing for given Zz and 1/1 the bounds S x(z2) and Smin(Zz), 1113
sented exactly by the series (3.10) with the maximum
one automatically also solves the. dual problem consist- number of N = 2M = 4 terms. The condensation of this
ing in determining_for <f/=90°_-1" and S2=<foZ2 the series according to Eq. (3.lla, b) then yields the two
depth bounds z,,,,,(S 2) and z,,, 1.. (S 2), to which a@ has to canonical models
be integrated to reach the given conductance S 2 under
the most adverse and favorable conditions. The cong- ( =0,
1
r 1 =4.09466·10 3 S
ruence of the shaded areas results from the particular I: ( 2 =5.23675·10 m, r 2 =3.77586·104 S
5

scaling on exploiting JcJ<J"=lcl<J 0 , JcliJ'"=lcl<J 0 [Eq. 5

( 3 =7.83023·10 m, r 3 =oo
(3.31 )]. The symmetry between conductance and depth
breaks down, if a· priori constraints are imposed on II· ( 1 =L05947·10 5 m, r 1 =6.66233·10 3 S
conductivity (cf. Fig. 7). · ( 2 =6.99673·10 5 m, r 2 =L01835·10 5 S
 199

at z 2 "' 700 km. The position z of conducting sheets (ab-

scissa) as a function of z 2 (ordinate) is shown in Fig. 3
for Sm,. and Sm,n· The ticks mark perfectly conducting
sheets. This figure provides an idea of the nonlinearities
in the problem by showing emerging, coalescing, disap-
pearing, and reappearing sheets.
The continuous deformation of the start model to
an identical end model is performed, on the one hand,
by the sheets emerging at z=O (level B, E, I, and L)
and getting detached at D, K, and M, and on the other
by the lower sheets either coalescing at D and F with
the ultimate perfect conductor (Sm.xl, or moving at L
and below N to infinity (Sm,J.
Model I is realized at A, D, K, and below F, where-
as model II occurs at B, I, L, and below N for -2_-> oo.
The models for Sm,. in the ranges BC and DE are
identical With those of Smin in the same range Of Z2;
Sm,. and Sm•• differ only by the conductance of the
\ sheet at z,. In the missing range CD all necessary
conditions imposed on smax can be satisfied on taking
the same model as for smin• but it turns out that choos-
I ing the four-sheet model for Sm,. with the reappearing
perfect conductor produces in this range slightly greater
\ values. This underlines the fact that the conditions de-
K rived in Sect. 2 are necessary but not sufficient.
600 Figure 4 displays the resulting bounds for the two
km frequencies, along with the bounds obtained by each
z,I L frequency separately. In the present example, joint con-
sideration of the two frequencies improves spectacularly
800 M
the lower bound on S, whereas the upper bound curve
~~~~~~~~~~~~~~~~~~~~N deviates only slightly from the curve consisting of the
Fig. 3. The position z of conducting sheets as function of z 2 in smaller of the values of smox obtained by single-fre-
the extremal models for the unconstrained two-frequency quency interpretation. At level z 2 =( 2 of model II, the
problem (first and fourth Sq harmonic) values of smin for 1 cpd and 1 &4 cpd do not appear to
be strictly equal, but differ only by 0.01 %. The marks
at the right margin denote the depth of the shal-
(giving for a possible check of the algorithm three to lowest perfect conductor, moving from 551 km for
four more digits than warranted by the accuracy of the 4 cpd to 692 km for I cpd and to 783 km for 1 & 4 cpd.
data). Sm,.(z 2 ) has model I not only at the start and Finally, it is noted that at the distinguished levels
end, but returns to it also at the intermediate depth z 2 "' 106, 524, and 700 km, where the slope of the
Zz"' 524 km. Similarly, sm,,,(z,) also returns to model II bounding curves changes discontinuously, the Lag-

1cpd : c c ( 5 7 5 - i 260 l km
200
4cpd: c - (290 - i 275) km

400

600
4c d

1cpd
- -· Fig. 4. Bounds for S(z 2 ) in the
unconstrained two-frequency
' '
''
... problem. All models fitting these
data must fall into the shaded
'' area. The joint consideration of 1
800 • and 4 cpd considerably improves
km the bound for smln and shows that
1 &4 cpd
all models must have a good
i 4cpd 1cpd
conductor not deeper than 700-
'2 800km
0 20 40 60 80 -s 100 kS
200
~
rangian multipliers J. 1 and A2 are also discontinuous, ing z2 ) can be determined from the 2M data. However,
meaning that near these levels the relation (2.10) for an the multipliers Ai remain undefined, since Eqs. (3.16a, b)
estimate of the influence of data errors has to be ap- yield only 2M -1 real equations (z 2 fixed!) for the M
plied with caution (cf. also case Cl in Sect. 3.6). complex Ai. In fact, the Ai are discontinuous for these
special values of z,, yielding different sets for slightly
smaller and greater z 2 (examples: A, D, F, K).
3.6 Extremal models for many fi'equencies
C2: M sheets of finite conductance (=canonical
As a basic limitation of this study, it is assumed that model II), z2 coincides with one of the sheets (cf. Cl for
there is a 1D model, which exactly fits the data set. further discussion; examples: B, I, L).
This restriction may be dropped in subsequent work, The models Al and A2, Bl and B2, Cl and C2 are
where the equality constraints (2. 7 a) are replaced by dual in the sense of Sect. 3.2.
inequality constraints that demand only a fit of the The model type for varying z2 may have to be
data within a suitable multiple of the standard de- changed for the following reasons:
viation. With the present tools, however, there are two a) The system of nonlinear equations no longer has
approximate ways to handle inconsistent data: a solution (e.g., because the deepest perfectly conduct-
a) The ID information (3.10) is extracted by. qua- iµg sheet disappears or reappears at infinite depth or
dratic programming; the cleaned data allow a reason- merges with another sheet, or the conductance of the
able construction of extremal models only if the num- deepest sheet increases from finite values to infinity).
ber M of frequencies does not exceed the numb~r N of
terms, since for M > N only one model exists. b) A surface sheet emerges. According to Eqs.
(3.14a), (2.14), and (2.3), the condition is
b) Extremal models are constructed for different
M
consistent subsets of the data; and the bounds on S(z 2 )
for the whole set are estimated by taking the greatest of D(O)= w(O)-Re L iwiµ 0 JcicJ =0,
i= 1
the lower bounds for each z 2 and the smallest of the
upper bounds. In the simplest case a consistent' subset whereas D(O)>O in the absence of a surface sheet.
comprises the data for just one frequency, requiring c) The surface sheet becomes detached and moves
only Rec;;:o, Imc;;:;O. downwards. This happens when ( 1 becomes a freely
Generalizing the experience gained from a small varying parameter, implying according to (3.14b), (2.14),
number of frequencies, it appears that in the' M-fre- (2.3), and (3.2) that
quency case the extremal models belong to one of the M
six categories: D'(-O)+D'( +0)= Re L 2iwi11 0 ciJc/2-iwiµ 0 t 1 cj)=O.
A 1: The model consists of M + 1 thin sheets with i= l

t M + 1 = oo and with one sheet fixed at depth z 2 . There This section concludes with an example for approxi-
remain 2M free model parameters, which can be de- mate extremal models from an inconsistent data set.
termined from the 2M real data. If required, the M- We choose the COPROD data of Jones (1980), which
complex multipliers A. are obtained by solving the lin- have been inverted by different authors using a variety
ear system of order 2M resulting from the necessary of techniques. The shaded area in Fig. 5 was deter-
conditions (3.16a, b) for the 2M free model parameters mined by the approximate method b) from 11 con-
----
(examples in Fig. 3: AB, DE, KL). sistent subsets comprising a single frequency. The figure
A2: This case differs from Al only in preassigning also displays as S(z 2 )-curves the various results of in-
---- version. These models freely use the space allotted them
(, =0 rather than <M+i = oo (examples: BC, IK, LM). and occasionally even transgress it because of the ap-
Bl: M+2 sheets with preassigned ( 1 =0, <M+z=oo, proximate nature of bounds and models. This example
and one sheet at fixed depth z,. There remain 2M + 1 shows that the approximate method b), which for con-
free model parameters, for which there are 2 M +I sistent data tends to produce conservative bounds (cf.
equations of type (3.!6a, b), linear in the 2M real quan- Smin in Fig. 4), can even yield too narrow bounds for
tities Rdi and ImJci. As a compatibility condition, the inconsistent data, where the range of acceptable models
determinant of order 2M + 1 formed by the matrix of is broader because in fitting the data different frequen-
the linear system and the right-hand side of (3.16a, b) as cies (or frequency bands) can be emphasized. Also, in-
(2M +!)st column has to vanish. This equation in con- consistencies in the data may lead to unreasonable
nection with the 2M data furnishes 2M + 1 nonlinear restrictions in the feasible (shaded) area. For this reason
equations for the 2M + 1 free-model parameters. If re- the four longest periods contained in the COPROD
quired, the multipliers Ai can finally be obtained by 2M data have been omitted when constructing the bounds.
of the linear equations (omitting an equation--"'.ith the
right-hand side vanishing). Examples are CD, EF.
4. Extremal models
B 2: M + 1 sheets of finite conductance with one for the constrained conductance function
sheet fixed at z = z,. There are again 2M + 1 free-model
parameters, which are determined in analogy to B 1 Section 3 was devoted to the unconstrained extremal
(example: MN). models with er_ (z)"' 0, er+ (z)"' oo. In this section we
Cl: M+l sheets with ( 1 =0, <M+• =OO (=canoni- shall consider as a simple example for constrained ex-
cal model I), z2 happens to coincide with the position tremal models the one-frequency problem of Sect. 3.4,
of one of the sheets. The 2M-model parameters (includ- where now the range of admitted conductivities is
 201

100

200

g
a 16 s
300 21 s

/
b
c 52 s
d 70 s
e 95 s
400 f 236 s
g 585 s
h 792 s
j 1449 s
k 2653 s
500~~~~~~~~"--'-~~~~~~~~~~~"-~~'-L-----"~~~"--'-~~

km 500 1000 1500 -s 2000s

Fig. 5. Approximate bounds on S(z 2 ) as derived from
'
the 11
shortest periods of the CO PROD data; only the 10 periods listed
have contributed to the bounds. In general, longer periods should be binding at greater depth. Due to inconsistent data the
longer period of 21 s (b) is bounding Smax at shallower depth than the shorter period of 16 s (a). Also displayed as S(z 2 ) curves
are the results of the various interpreters

=
bounded by "+ (z) "+' where "+ is a positive constant.
In particular, the structure of the extremal models is
D(z)=w(z), which is non-negative for Sm;,, and non-
positive for Smax· The other exception occurs for Smax in
studied when assigning to "+
a varying multiple a of a certain range below the point, where the two con-
the apparent conductivity u, [cf. (3.21)], i.e., "+=au,. ducting layers merge into a thick conductor having
In this particular case, the extremal models consist of a z=z 2 as interior point. In this case D(z 2 -0)< -1 and
sequence of uniform layers with u(z) = "+and insu- D(z 2 +0)=D(z 2 -0)+1<0, i.e., D(z) does not change
lators with u(z)=O. According to Eq. (2.20), the neces- sign at z=z 2 •
sary extremal conditions are D(z)~O for u(z)=O and For smal1cr values of a, the conducting layers nec-
D(z);";O for u(z)=u~,, with layer boundaries at po- essarily become thicker, and for IX= I the extremal
sitions where D(z) changes sign. models for smax are just uniform half-space models.
Before considering more details, it might be useful Therefore, the case of a=l.l, for which narrow non-
to discuss first the structure of the extremal models as conductive channels still occur, is considered. Due to
displayed in Fig. 6 for i/J = 45° and various values of a. the overshot phenomenon of'·the apparent resistivity
The position of the thin sheets (abscissa) as a function curve, it is possible to construct extremal models even
of z 2 (ordinate) in the unconstrained case a=co is for a< 1, meaning that the true conductivity is smaller
shown at the left. The ticks mark again perfect con- everywhere than the apparent conductivity. It is easily
ductors. The main structural units of these models are verified that for the considered phase i/J = 45° the smal-
clearly discernible in the following models of con- lest value of IX is am,.,=tanh 2 (n/2)=0.8412, which corre-
strained conductivity, where the layers with cr= a+ arc sponds to a surface layer of conductivity a 111 i 11 a(I, thick-
shaded. As expected from experience with the uncon- ness (n lcl/V'2l coth (n/2) = 2.4221 lei, and an insulator
strained case, smin generally shows a conducting layer below. For the smallest possible conductivity the mod-
starting at z = z,, which is just excluded from the range els for smin and smax coincide and are independent of
of integration, whereas smax in general has a conducting z,.
layer, which ends at z 2 and is just included. There are The resulting bounding curves of S(z 2 ) for selected
two notable exceptions (dashed diagonal). For suf- values of a (curve parameter) are shown in Fig. 7. A
ficiently small z,, the extremal models demand only bounded value of (J. mostly affects smax' whereas smin is
that a=O for Smin and a=a+ for Smax in O~z;'£z 2 , only influenced for relatively small IX. The dashed lines
which can be reached in different ways. When z 2 ex- refer to the smallest possible value of IX, for which only
ceeds a certain limit, however, there will only be one one model exists. For 0;";1/1<45° it consists of a slab at
model. This first unique model is then also shown for finite depth, and for 45°;";1/1;";90° it is a surface slab.
smaller z 2 . This situation is comprised in the extremal The actual construction of the extremal models is
condition (2.20) since, fn this range Of Z2, Smin and Smax similar to that in the unconstrained case. If "_ (z) and
do not depend on the data, i.e., 2=0 [Eq. (2.10)] and "+ (z) are independent of z, the conductivity models

1
 202
I~

Smai<
. ~+•u O'a
(
"
i

];.2 B
ICI ICI
~-~~~~-~-~-~3~~~~~~-~-~-~JLL-'-"~-'-'--'-'_1_~-~--"'

Fig.6. Structure of the unconstrained (left) and constrained (center and right) one-frequency extremal models for 1/1=45°. For a
given z 2 (ordinate) the unconstrained models are presented only by the position of the conducting thin sheets (without specifying
their conductance), whereas the constrained models (a+ <co) are completely characterized by specifing the position of the
··~
1
conducting layers (shaded)

consist of a sequence of uniform layers. The pertinent will be up to four unknown levels for the one-frequency
formulae are briefly summarized. Let K uniform layers case. The dimension of the resulting nonlinear system
exist, with the top at z=(, (( 1=0), and layer con- can again be reduced to four by eliminating .A. on re-
ductivities <J,, k=l,. . .,K. With y;,=iwµ 0 <J, and placing the four equations (4.3), which are linear in .A.,
d,,=(,+ 1-(, as thickness of layer k, k=l,. . .,K-1, by two compatibility conditions. These conditions de-
the theoretical response c [<J] = c1 is recursively deter- mand that any two of the four (3 x 3) subdeterminants
mined from of the augmented (4 x 3) matrix, i.e., w((,) as third col-
umn, have to vanish.
c,=1 y,c,+1+tanh~,ad,)' k=K-1, .. .,1 (4.1) All extremal models (except the half-space for Sm,,,
y, 1+r.c,+ 1 tan (r.d,) o/=45°, a=l) terminate with an insulator after a finite
starting with CK= 1/yK. [Note again that in this context number of layers. This pattern changes, when the a
the subscript k on c refers to the level (., whereas in priori bound <J _ >0 is imposed, because w(z) vanishes
other applications the subscript specifies the frequency.] for z>z 2 , and in this range D(z)=Re{.A.F(z)} will show
The Frechet derivative F(z)= -iwµ 0 f 2 (z) at z=(, is a kind of damped oscillations rather than tend to a
determined from constant. As a consequence, there will be an infinite
number of layers, which below a certain level, however,
'1•• f(( 1)=f(0)= -c[<J], will have an insignificant influence on the actual
bounds.
f((,)=f((i) 'ff y,.+l/c,. r'"""
11=1Yu+1/c11+1 '
(4.2)
5. Conclusion
The necessary conditions (2.20) require that at a discon-
tinuity z=(,, different from 0 and z 2 , D(z) changes The explicit examples in Sects. 3 and 4 have demon-
sign, i.e., strated the possibility of constructing rigorous bounds
D((,)= w((,) +Re {AF((,)} =0, (, *0, z 2, (4.3) for S(z 2 ). But the examples have also shown that the
consideration of only a few frequencies yields pessimis-
D'((,)*O. At z=z 2 , the conductivity is only discon- tic bounds, which might not be very useful in the
tinuous if D(z 2 -0)-D(z 2 +0) <0. The condition (4.3) for geophysical application. The situation will even be
each unknown discontinuity level, together with the data worse if bounds for spatial averages of the conductivity
c [<J] = c, obviously provide the correct number of equa- are constructed rather than for the conductance. There-
tions to compute the positions of the unknown discon- fore, the joint interpretatio·n of a larger number of
tinuities and .A.. The examples in Fig. 6 show that there frequencies appears to be mandatory. The handling of
 203

- S/Clcla8 ) exists "".here (2.17) is satisfied and that a< z < b is the deepest.
Then (2.17) reads
0
M
D(z)=w(z)+ Re L: ¢ 1(z)=O, zE(a, b) (A-1)
:1=1

with ¢ 1(z)•=A.F.(z). For a<z<b, the conductivity a(z) is

positive. Let a{P1ib), p ~ 0 be the first non vanishing derivative
of a(z) at z = b-0. In z > b, a set of K thin sheets of finite
2
conductance may exist, possibly terminated by an additional
perfect conductor. For ease of presentation the case K = O and
= K>O are considered separately.

K=O
Let the possible perfect conductor be at z=b+H. Differenti-
ate (A-1) 2M (p+2) times and evaluate the derivatives at
z=b-0 on using </>/(b)= -(2/H) <f>;(b) and the differential
equflJions (2.15a, b). The result is a set of 2M homogeneous
linear equations for Re¢1(b) and Im<f>;(b)
t
' ) M
Re L: (iw)"</>1(b)=O, n=l, ... ,2M, (A-2)
I i=l

I where then-th equation is obtained from the n(p+2)-th order

derivative, observing that this derivative of ¢i(z) at z=b-0 is
of the type

<f>;(b) L: y,,,(iw)"'
111=0

with real frequency-independent coefficients y.,, and y11 :F 0. The

set (A-2) disintegrates into two uncoupled systems for
wi Im¢· and wJ Re¢., both having as system determinant the
I nonvanishing Vandeimonde determinant (e.g., Smirnov 1964,
p. 21)

Fig. 7. Bounds on S(z 2 ) in the one-frequency case constrained

by cr(z) ~ aa11 , where o: is the curve parameter. The limiting
n
l~r<s~M
(w;-w;)<Fo. (A-3)
case a= oo is already shown in Fig. 2. For the smallest possi-
ble value of rx there is only one model depicted by the dashed
curve Hence, (A-2) allows only the trivial solution <f>,(b)=O,
j=l, ... ,M. This solution is compatible only with w(b-0)=0
many consistent data will not pose serious problems, and either a perfect conductor at z = b or b = oo, because the
but an extension of the theory is still necessary to deal modulus of <Pi cannot increase with depth. In the first case
with the most interesting case of many real data that also </>j(b)=O, but </>j'(b)<FO. Repeating the above arguments
are generally inconsistent. for ¢'/(b) by considering the derivatives of order n(p+4),
n = 1, ... , 2M, it follows that also </>j'(b) =0, implying ¢ 1(z) °=O
in a<z<b. In the case b=oo differentiate (A-1) three times,
Acknowledgements. The author is greatly indebted- to one of
the referees for pointing out a substantial error in an earlier use (2.15b), divide by 4µ 0 y-;, and integrate to obtain
version of Appendix A. Moreover, the author wishes to thank M
Prof. H.-J. Dtirbaum for providing the opportunity to carry Re L: iw1ir.;:w ¢ 1(z)= y, z >a, (A-4)
out a major part of the present study at the Federal Institute i= 1
for Geosciences and Natural Resources (BGR) at Hannover. where y is a constant. By comparing with power laws a(z)
Many thanks also to Dr. A.G. Jones for inaking available the =Az"' leading to Bessel function solutions (or appealing to a
results of the CO PROD study prior to publication. bounded energy dissipation), it is always found that

Appendix A:
Lmear independence of the Frechet kernels
The simple structure of the extremal models is a result of the which implies that y=O. Dividing (A-4) by Vu
and applying
assertion that there is no set of M complex constants Ai, not the same operations another (2M -1) times, we end up again
all equal zero, so that (2.17) is satisfied in any interval where with the system (A-2), <f>;(b) being replaced by ¢ 1(z). Hence,
¢i(z)=:O in z~a. This shows that for K =0, no solution
w(z) is constant and a(z) is positive. The functions
Fi(z)-f,.2(z) have to present a downward diffusing field with
¢ 1(z) $0 of (A-1) can be found.
FJ(z)-'>' 6 for Z-'>' zmax• where zmax is either infinity or the depth
K>O
to a perfect conductor. This qualification will eliminate the
nontrivial solutions of (2.17), which may exist for a particular Below z = b let there exist K thin sheets of finite conductance
choice of a(z). r, at z=(,, k=l, ... ,K and define d,•=(,-b;;;O, d,•=(,
At the beginning it is assumed that at least one interval -(k_ 1>0, k=2, ... ,K. In addition, there may be a perfect
·,:.,'

204

conductor at z=(K+H. Separately considered are the cases Appendix B:

w(z)=O in z;;;b and w(z)$0 in z;;;b. Equivalent partial fraction expansions

a) w(z)=O in z;;;b Let for M distinct frequencies wi the theoretical transfer func-
tion ci[o] of a layered ground be represented by
In this case the necessary conditions (3.15a, b) have to be N
satisfied at all K sheets and lead to the 2K equations (B-1)
ci[a]=a 0 + L -a,,.b , j=l, ... ,M,
11=1 ,,+iwi = (3.10) I
M
Re I e 2 K_,,(w;)<P;((K)=O, n=O, ... ,2K-l. (A-5) where the b11 arc distinct (otherwise N can be reduced). As- r
i= 1 sume the ordering b11 >b1J+t· Then a0 and bN have to be non-
Working from z=(K upwards and starting with e 2 K=1, the negative, all other constants are strictly positive. The follow-
functions e2 K_ 11 (w) are, on account of Sect. 3.1 [in particular ing statement will be proved:
Eqs. (3.2), (3.4), and (3.5)], recursively defined as a) For N~M exist two (condensed) representations

ezk- t =ezk(iOJllo't'k-2/ck), (B-2)

(A-6)
ezk- 2 = ezk(l + dk/ck)
2 =(3.lla)

with k=K, ... , 1 and the ck are obtained by the recu·rrencc (B-3)
relation (3.3), starting with cK=H/(l +iwµ 0 -cKH). Examining =(3.llb)
the structure of e2 K_ 11 (w), it is easily found that
j= 1, .. ., M. ExcludiI_!g only the identity M = N with a 0 =bN
"
e 2 K_,Jw)= L Y.,,(iw)"', n=O, ... , 2K-1 (A-7)
=0, implying also A 0 =BN=O, the representations (B-2) and
111=0
(B-3) consist each of 2M positive constants corresponding to
the 2M data. There is no representation with less than 2M
where f 111 is again real and frequency independent with y11 =t= 0. constants, i.e., B 111 and .B"' are distinct.
Hence by linear combination (A-5) is equivalent to 1 b) For N < M the representation (B-1) is unique, i.e., no
M alternative partial fraction expansion exists.
Re L(iw;)"</J;((K)=O, n=0, ... ,2K-l. (A-8) The proof is concentrated on the representation (B-2), and
i= I at the end only the necessary modifications for (B-3) are
For K;,;M (A-8) implies already <fJ.((K)=O; for K<M the stated. At the outset it is assumed that there is an expansion
missing linearly independent equatio~s are obtained from the of type (B-2) with K ~ M complex terms. Then it has to be
higher derivatives of D(z) at z=b-0. First, let d1 =('1 -b>O. shown that the nonlinear systctn of 2M equations
Then the three quantities
~ A,,,
;_, 2
B,,, _ ~ _ci,J>,,__)
2-ao+;_, 2 2 (B-4a)
<P1(bl/<P;(\K), </Jj(b)f<P;(\K), and m=1B.,.+wi "=1b,,+wi '=1 M
(A-9) K N } , .. .,
</J;' (b )/{ </J;(b) </J;(( K)} L ~= L _a_,_,- (B-4b)
m=l B~1 +wJ 11=1 b,7+wJ
are polynomials in iciJ. of exact degree 2K. After evaluating
the derivatives of </>;(z)'of order n(p+2), n = 1, ... , 2M -2K -1 has 2K positive solutions A,,,, Bm, 1n = 1, ... , K. First, we recall
at z = b-0 on, using the preceding result in connection with the elementary partial fraction decomposition
(A-8) and (2.15a, b), the validity of (A-8) is extended up to rx
I
M M
powers 2M-l. As a consequence, </>i((K)=O, implying in fact
<fJ;(z)=O. For d 1 =0 the first two terms of (A-9) are only of
x2k TI (x2+wf)= L ~. k=O, ... ,M-1 (B-5)
i=I j=l X +wj
degree 2K-2 and 2K-1, but the same result is obtained by
considering the derivatives of order -p+n(p+2), 11= with
l, ... ,2M-2K.

PJ w(z)$0 in z;;;b

According to Sect. 3.3 the discontinuity of w(z) coincides with then multiply (B-4a, b) for k =0, ... , M -1 by ";" sum over},
the position of a sheet. If this is the sheet k= 1, then (cf. identify x on the left-hand side with B 111 , on the right-hand
Sect. 3.3) side with b11 , and obtain by (B-5) the new set of 2M equations
K N
D(\ 1 -0)=1, D'(\ 1 )=0 for smin•

D(\ 1 -0)=0, D'(( 1 )<0 for smax

(A-10) L G,,,B:,,=a
111=1
0 b1,zM·-l + L g,,b:,,
11=1
1=0, ... ,2M-1, (B-6)

However, D(b)=O, D'(b)=O implies D(z)=A(z-b) 2 in where (jik is the Kronecker symbol and
b~z~( 1 , which for d 1 ~O is incompatible with both cases of
M M
(A-10). If the discontinuity occurs at level k> 1, then
G111 ==A 111
Ifl (B,7 +wJ),
}= 1
1
i= 1
g11:=a11

The equations for odd and even I result from (B-4a) and
ITI (b,7+wJ). (B-7)

which again cannot be met by any choice of A for d 1 ~O. (B-4b), respectively. The quantities G111 rather than A,,, inay 110\V
Summarizing the above results, in the unconstrained case be considered as unknowns. In the derivation of the first tern1
no downward diffusing field solution satisfying (2.17) can be on the right-hand side it has been observed that by virtue of
found for a(z)>O in a<z<b. The assumption a(z)>O is es- (B-5)
sential, since for a(z)~O non-trivial diffusing solutions of
(2.17) exist, e.g., D(z)=O for z>z 2 in the case of Smin (cf.
Sect. 3.3) without </J;(z) vanishing identically.
 ---- ~

205
1
The 2M equations (B-6) are linear in the K unknowns G111 • In for this homogeneous linear system, the determinant has to
order for there to be a solution, any K + 1 equations of (B-6) vanish, implying
have to be linearly dependent. Taking, for instance, the K + 1
eqtfations from l=i to l=i+K, i=0, ... ,2M-K-1, then for B, ... BM
B, ... BM-I
I
qr
each i there has to be a set of coefficients 1, k=O, ... , K, not
aoOi,M-1
I

+I
N

glib:, =0,
all equal to zero, such that
11= I BM ... B''IM
K
BM ... s~-1
b,. ... b'\1
L qri n:.-:-k=o, 1n= 1, ... , K, (B-8a) "
k=O i=O, ... , M -1-. The above Vandermonde determinants are
easily evaluated by (B-12) and yield after much cancellation
(B-8b) since the B,,. are distinct
N M
Since we are searching for positive B 111 , (B-8a) may be divided
n:
by 11 to yield
a 0 0i,M-l+ L g,,b;, IT (b, -B 1 111 )=0, i=O, .. .,M-1, (B-13)
11= I m= I
K
'<'
L., qk '''B'm -0
- • 1n=l, ... ,K. (B-8c) or equivalently
U=0
N M
qr 1
is independent of m. Hence, each B111 satisfies the same aoYM-1 +L g,,pM_ 1(b 11 ) IT (b, -B,,,)=O,
1
(B-14)
algebraic K-th order equation, i.e., by the fundamental theo- II=} m= I
rem of algebra the existence of the set qrl implies the exis-
tence of solutions B,,, being the K roots z of ' where PM_ 1 (b,J is an arbitrary polynomial in b11 of degree not
higher than M -1 with YM- l as coefficient of b~1 - 1 . The M
K equations (B-13) form a nonlinear syste1n for B 111 • The exis-
L qr>zk=o. (B-9)
I
tence of a solution set Bm is granted from (B-8c) and (B-9). To
k= 0
show the positivity of the particular element Bk, k=l, ... ,M,
On the other hand, the roots define the coefficients 1 un- qr we take
iquely apart from a scaling factor. Therefore, the qrl al·e in M
fact independent of i, qiil=:qk. Normalizing by qK=l, the PM-l(h..)= IT' (b .. -B~). YM-1 =1,
remaining K coefficients qk have to be determined frotn the m=l
2M -K linear equations (B-8b). First, let K = M. Then the
where prime and asterisk denote, respectively, the omission of
system determinant
the factor 1n = k and the complex conjugate, and obtain from
(B-14)
Ll"=det{J g,.b~+'}.
1 i,k=O, ... ,M-1 (B-10)
a0 + L:
N
g,,(b,,-B,)
M
TI' lb,.-B,,,l'=O. (B-15)
can be expressed for N ~ M as 11= I 111= 1

This shows that Bk is real and positive.

Next it is proved that also A, [or G., cf. Eq. (B-7)] is
(B-11) positive. The solution of the first M equations of (B-6) for Gk
yields by Cramer's rule on exploiting, again the simple prop-
erties of the resulting Vandermonde determinants
where the summation extends over all M-tupels nk,
k=l, ... ,M with 1~n 1 < ... <nM-:£_N (e.g., Smirnov, 1964, G = ~ TIM' b/l-B/11 (B-16)
p. 28). The determinants in (B-11) are again Vandermonde k L, g,, B B .
II= I >II= l k- Ill
determinants, which are generally defined as V =<let {x:~- 1 },
j,k=I, ... ,Mand are given by (e.g., Smirnov, 1964, p.21) The definite sign of Gk is not yet obvious. However, expand-
ing the product by
V= TI (x,-x,,). (B-12) M M
1 &p<q;f;M
f1' (B,-B,,,)= f1' [(b,,-B,,,)-(b,,-B,)]
111= 1 Ill= l
All ter1ns in the sum (B-11) are positive by virtue of the
M
assumptions N ~ M and b11 distinct. Hence, LIM> 0 and the
coefficients qk can be detennined uniquely. In the case N < M, =(b,,-B,)PM-l(b,,)+ TI' (b,,-B,,,), (B-17)
111=1
however, LIM=O, as is seen by formally adding in (B-10)
M -N terms with gN+i = ... =gM=O and using (B-11). In this where PM- l is, in fact, a polynomial of degree M - 2, we
case there is no set of coefficients qk, implying that (B-1) obtain from (B-16) by means of (B-14) with Yu- r =0:
cannot be modified.
The case of K < M ~ N must still be discussed, in which (B-18)
the linear system (B-8b) consists of 2M-K=K+2(M-K)
equations for only K unknowns. A necessary condition for This proves the existence of the representation (B-2).
the existence of a solution is the linear dependence of the first Skipping details of derivation, we mention only two ad-
K + 1 equations, which is equivalent to the state1nent that ditional identities required in Sect. 3.2:
LIK+>=O. However, (B-11) with M=K+l shows that
LIK+ 1>0. Hence, for K<M the first K+l equations are
already linearly independent and there is no solution qk. (B-19)
Therefore, in the sequel only the case K=M, N~M has to be
M N
considered.
Having established the existence of the set Bm as the roots L: (A,,,/B,,,)=a 0 + L: (a,,/b,,)
m=l 11=1
of (B-9) with K =M, w~ have to show that B,,, is real and
positive. Without exploiting (B-9) any further, we return to
(B-8b, c). In order for there to be a solution qrl = qk, k = 0, ... , M
-TI (w/B)'·
M [
a0 + L:N (g,,/b,,) TI M
(b,,-B,,,)
2]
• (B-20)
j=1 11=1 111=1

I
 .d.

206 \
,f
The existence of the representation (B-3) is proved similarly. posed problems in differential equations, G. Anger, ed.:
The equivalents of (B-6), (B-13), and (B-14) are pp. 25-34, Berlin: Akademie-Verlag 1979
M-1 Brown, K.M.: Computer oriented algorithms for solving sys- ,
Ill=
L 1
G.,.B:,,+GMt5 10 tems of simultaneous nonlinear algebraic equations, G.D.
Byrne, C.A. Hall, ed.: pp. 281-348. NewYork and Lon-
I
N don: Academic Press 1973 J
=(a 0 -A 0 )0 1, 2 u_ 1 + L g b:,.11
l=0, .. .,2M-1, (B-21) Jones, A.G.: Report on the COPROD study. Paper read at
h
11= 1 the Fifth Workshop on Electromagnetic Induction in the
N M-1
Earth and Moon, Istanbul, August 18-23, 1980 i
(a 0·-A 0 )~ 1 M+ L g,.b;, TI (b,.-B,,,)=0, Lawson, C.L.,. Hanson, R.J.: Solving least squares problems. ).
ll=l m=l
Englewood Cliffs: Prentice-Hall 1974
i=l, ... ,M, (B-22)
Oldenburg, D.W.: Funnel functions in linear and non-linear
(ao-Ao)YM-1 +
N
L g11PM-1(bll)bll
M-1
n (bu-Bm)=O. (B-23)
appraisal. J. Geophys. Res., 88, 7387-7398, 1983 (
Parker, R.L.: Inverse theory with grossly inadequate data. '
11= 1 m= 1
Geophys. J. R. Astron. Soc., 29, 123-138, 1972
The positivity of Bk, k = 1, ... , M -1, is proved by taking Parker, R.L.: Best bounds on density and depth from gravity
M-1
data, Geophysics, 39, 644-649, 1974
PM-1(b11)= IT' (bll-B,;,), YM-1 =0, Parker, R.L.: The theory of ideal bodies for gravity interpre-
111= 1 tations. Geophys. J. R. Astron. Soc., 42, 315-334, 1975
Parker, R.L.: The Frech et derivative for the one-dimensional
implying electromagnetic induction problem. Geophys. J. R. Astron.
N M-1 Soc., 49, 543-547, 1977
L g,.(b,.-B,) TI' lb,.-B,,,1 2 =0 Parker, R.L.: The inverse problem of electromagnetic in-
II= 1 111= 1 duction: existence and construction of solutions based on
and b 1 ~jjk~bN. Choosing incomplete data. J. Geophys. Res., 85, 4421-4428, 1980
Parker, R.L.: The existence of a region inaccessible to magne-
M-t
to tell uric sounding. Geophys. J. R. Astron. Soc., 68, 165-
PM-1(b,,)= TI (b,.-B,,,), YM-1=1 170, 1982
Ill= 1
Parker, R.L., Whaler, K.A.: Numerical methods for establish-
we obtain . ing solutions to the inverse problem of electromagnetic
N Af-1 induction. J. Geophys. Res., 86, 9574-9584, 1981
A0 =a 0 + L: g,.b,. TI (b;,-B,,,) 2 >0. (B-24) Pearson, C.E. [ed.]: Handbook of applied mathematics.
New York: Van Nostrand Reinhold Co. 1974
11-l m=l
Sabaticr, P.C.: On extremal solutions of Sturm-Liouville in-
Finally, G,,, is found by solving the first M equations of (B-21) verse problems. In: Inverse and improperly posed prob-
for G111 and using analogues of the identity (B-17): lems in differential equations, G. Anger, ed.: pp. 223-231.
- - N M-,1 (b,.-B,,,)' Berlin: Akademie-Verlag 1979
G,B,= TI g,.b,. TI B- -B- >0, k'=l, ... ,M-1, (B-25) Schmucker, U.: ·Anomalies of geomagnetic variations in the
1 1 k southwestern United States. Bull. Scripps Inst. Oceanogr.
(b -B111 )' >0.
II= »I= Ill

_ N M -1 13, 1-165, 1970

GM= L g11 f1
11=1 lll=l
11
jj
m
(B-26)
Schmucker, U.: EM-Obertragungsfunktionen ftir Sq. In: Pro-
tokoll Kolloqu. ,,Erdmagnetische Tiefenforschung" in
Grafrath March 19-23, 1984, J. Homilius and V. Haak,
References ed., Hannover: Niederstichsisches Landesaint ftir Boden-
Anderssen, R.S., Seneta, E.: A simple statistical estimation forsch ung 1984
procedure for Monte Carlo inversion in geophysics. Pure
Appl. Geophys., 91, 5-13, 1971
Smirnov, V.I.: A course of higher mathematics, Vol. 111/1.
Oxford [etc.]: Pergamon Press 1964
i
Anderssen, R.S., Seneta, E: A simple statistical estimation Weidelt, P.: The inverse problem of geomagnetic induction. Z.
procedure for Monte Carlo inversion in geophysics. II: I. Geophys., 38, 257-289, 1972.
Efficiency and Hempel's paradox. Pure Appl. Geophys., Weidelt, P.: Extremal models for electromagnetic induCtion in
96, 5-14, 1972 two-dimensional perfect conductors. J. Geo phys., 49, 217-
Barcilon, V.: Ideal solution of an inverse nonnal mode prob- 225, 1981
lem with finite spectral data. Geophys. J. R. Astron. Soc.,
56, 399-408, 1979
Barcilon, V., Turchetti, G.: On an inverse eigenvalue problem Received October 1, 1984; Revised August 1, 1985
with truncated spectral data. In: Inverse and improperly Accepted August 1, 1985

" '