Geophysical Prospecting, 1997, 45, 745–762

New digital linear filters for Hankel J0 and J1

transforms1
D. Guptasarma2 and B. Singh2

Abstract
The numerical evaluation of certain integral transforms is required for the
interpretation of some geophysical exploration data. Digital linear filter operators
are widely used for carrying out such numerical integration.
It is known that the method of Wiener–Hopf minimization of the error can be used
to design very efficient, short digital linear filter operators for this purpose. We have
found that, with appropriate modifications, this method can also be used to design
longer filters. Two filters for the Hankel J0 transform (61-point and 120-point
operators), and two for the Hankel J1 transform (47-point and 140-point operators)
have been designed. For these transforms, the new filters give much lower errors
compared to all other known filters of comparable, or somewhat longer, size. The new
filter operators and some results of comparative performance tests with known
integral transforms are presented.
These filters would find widespread application in many numerical evaluation
problems in geophysics.

Introduction
The application of digital linear filters for computations needed for resistivity
sounding was introduced by Ghosh (1971). Improvements on the method were then
made by Koefoed (1972, 1976, 1979), Koefoed and Dirks (1979), Das and Ghosh
(1974), Das, Ghosh and Biewinga (1974), Kumar and Das (1977), Johansen and
Sorensen (1979), Anderson (1975, 1979, 1982, 1989), Guptasarma (1982a),
Christensen (1990), Mohsen and Hashish (1994), Sorensen and Christensen
(1994), and others.
Such a filter operator consists of a set of n abscissa values for sampling the input
function and n weights. The placement of this pattern of sampling points is
determined by the abscissa at which the output function is needed. The value of the
output function is computed as the sum of products of the n samples of the input
function with the corresponding filter weights.
It is convenient to have the input sampling points spaced evenly on the logarithmic

1
Received March 1996, revision accepted January 1997.
2
National Geophysical Research Institute, Uppal Road, Hyderabad 500 007, A.P., India.

q 1997 European Association of Geoscientists & Engineers 745

746 D. Guptasarma and B. Singh

scale. The advantage of even spacing is that after one set of n evenly spaced samples
has been computed, only one more sample at the same spacing needs to be computed
for getting another value of the output function at an adjacent point (Anderson 1982).
In all the work referred to above, the input samples are taken at an even spacing.
Guptasarma (1982a) gave a method of designing optimized linear filter operators in
which the placing of the set of the input abscissae, as well as the spacing between
them, is adjusted in order to obtain the best possible results with a given number of
filter weights. This method was demonstrated for the cases of 7-, 11- and 19-point
digital filters for the computation of the Schlumberger apparent resistivity from a known
resistivity transform. Guptasarma (1982b) used the same method to design an
optimum filter for the computation of the time-domain response of a polarizable
target with given Cole-Cole parameters of induced electrical polarization. Nissen and
Enmark (1986) used this method to design a 20-point filter for a Fourier sine
transform and Rijo (1993) used a 19-point filter for a Fourier cosine transform. In
principle, the method can be used for generating filters for any linear operation such
as an integral transform. However, due to the finite length n of the operators used in
practice, and the finite spacing between the samples, they work best when applied to
smooth functions similar to the functions used for designing them.
In general, the accuracy of the result of computation with digital linear filters
increases both with increasing length n, because the so-called truncation error (the
error due to the fact that a finite number of weights is used) is reduced, and with
increasing sampling density, because the sampling error is reduced (Mohsen and
Hashish 1994; Rijo 1996; Christensen 1991). But short filters designed by adjusting
both the placing and the spacing of the operator can give very much more accurate
results than filters of comparable length n designed by using a pre-assigned spacing, as
shown by Guptasarma (1982a). This makes it possible either to reduce substantially
the amount of computation needed for comparable accuracy or to increase the
accuracy for a comparable computation time.
It was believed (Guptasarma 1982a) that the advantage of this method of
optimization might be lost if the length of the filter is increased. However, the
improvement in accuracy continues to be dramatic for fairly long filters. We present
four new digital linear filters, two for the Hankel J0 transform and two for the Hankel
J1 transform.

Digital filters for the Hankel transform

Hankel integral transforms of the type
…∞
f ðrÞ ¼ K ðlÞJi ðrlÞ dl; ð1Þ
0

where Ji is the Bessel function of the first kind and order i and K(l) is the function
being transformed, appear in many problems involving cylindrical symmetry,
specially in the area of electrical and electromagnetic induction prospecting. For a

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 747

given function K(l) and a given value of r, the numerical value of an integral of the
above form may be conveniently computed by using a digital filter, designed for use
with a specific kernel function, such as J0, or J1.
A digital filter specifies a set of filter weights, and two more constants: the shift a
and the sampling interval or spacing s. The shift a determines the placing of the
starting point for sampling the input function and the spacing s specifies the interval at
which the remaining samples are to be taken. The logarithm of the abscissa, log l, is
used as the abscissa, so that a is the logarithm of the factor by which the starting point
is to be shifted from 1/r, and s is the logarithm of the factor by which successive sample
points shift on the l-scale. (Throughout this paper we use logarithms to the base 10.)
Thus, one first computes a set of n abscissae given by
li ¼ ð1=rÞ × 10½aþði¹1Þsÿ ; i ¼ 1; 2; …n: ð2Þ
The function K is then evaluated for these values of l and the value of f(r)
computed from
X
n
r f ðrÞ ¼ K ðli ÞWi ; ð3Þ
i¼1

where W i are the n filter weights, by dividing the sum on the right-hand side of (3)
by r.

The method used for design

We provide two digital filter operators for the Hankel J0 transform for n ¼ 61 and 120,
and two filters for the Hankel J1 transform for n ¼ 47 and 140. These filters were
designed with a modified scheme of optimization, based on the work by Guptasarma
(1982a). Using a known Hankel transform and a given value of n, a set of equations
for different values of r, with r spread over a large range, is formed from (3), using
assumed values of the shift a and the spacing s. The solution of such an
overdetermined set for the weights Wi produces a filter operator. Its accuracy can
then be tested by using the same Hankel transform, or another known Hankel
transform with the same kernel, for an appropriately chosen set of values of r, and
computing the error for each r.
The standard transformations which we used for designing the filters and for testing
their error performances are listed below. Transforms (6) and (7) were used for
designing the J0 and J1 filters, respectively. For the J0 filters, transform (6) was also
used for the computation of errors during design. For the design of J1 filters,
transform (9) was used for the computation of errors. For the overall evaluation of
performance, transforms (4) to (6) were used for the J0 filters and transforms (8) to
(10) for the J1 filters.
…∞
expð¹clÞJ0 ðrlÞdl ¼ p ;
1
ð4Þ
0 ðc þ r 2 Þ
2

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
748 D. Guptasarma and B. Singh

…∞ 1

r2

l expð¹cl2 ÞJ0 ðrlÞ dl ¼ exp ¹ ; ð5Þ
0 2c 4c
…∞ c
l expð¹clÞJ0 ðrlÞ dl ¼ ; ð6Þ
0 ðc2 þ r 2 Þ3=2
…∞h i r
2
re¹ðr =4cÞ
¹cl 2 ¹cl2
le þ al e J1 ðrlÞ dl ¼ 2 þ a ; ð7Þ
0 ðc þ r 2 Þ3=2 4c2
…∞ r
l expð¹clÞJ1 ðrlÞ dl ¼ ; ð8Þ
0 ðc2 þ r 2 Þ3=2
…∞ 2
re¹ðr =4cÞ
l2 expð¹cl2 ÞJ1 ðrlÞ dl ¼ ; ð9Þ
0 4c2
…∞ p

ðr 2 þ c2 Þ ¹ c
expð¹clÞJ1 ðrlÞ dl ¼ p : ð10Þ
2 2
0 r ðr þ c Þ
The value of c was taken as 1 for the design process, but different values of c were
tried for checking the error. It may be noted that transform (7) is a linear combination
of the transforms (8) and (9), the relative proportion of the two transforms being
determined by the factor a . Such a linear mixing was carried out in order to improve
the performance of the filters for the functions similar to those in transforms (8) to
(10). The value of a was taken as 0.066 for designing the 47-point J1 filter and as 0.01
for the 140-point J1 filter. These values were arrived at after some trials, on the basis
of the performance of the resulting filters.
In some earlier publications, the absolute error, i.e. the absolute value of the
difference between the correct theoretical value and the value obtained using the
filter, was taken for the purpose of optimization, as well as for estimating
performance. However, most physical problems require the relative error, which
may be defined as the absolute error divided by the correct theoretical value. This
gives a measure of the number of significant digits of the result which can be relied
upon, regardless of the absolute value of the result. Thus, if the relative error is low,
the absolute error is not important in most physical problems. In contrast, a low
absolute error may be misleading because, if the correct value is very small, the
computed value could even have all its significant digits wrong while the error appears
to be small. It turns out, therefore, that examining the relative error is more
appropriate. This is what we have done.
As (2) shows, samples at large values of l are needed to compute the transforms for
small values of r. But the input functions in all the above transforms vanish for large
values of l. Consequently, for such small values of r for which all samples of the input
function are vanishingly small, any linear filter operator of the type we are considering
would produce a zero or near zero output, and a relative error close to unity. Similarly,

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 749

for large values of r, samples of the input function are needed for very small values of l
and, for some of the transforms, the relative error can be large for large values of r as
well. Thus, all such filters are useful only over a certain range of values of r, depending
on how the function behaves for large and small values of r. Longer filters are expected
to produce smaller errors and to be effective over a larger range of r.
We find that for any given filter length n, there exist very many sets of a, s and W
which produce closely similar error performances. It becomes necessary to make an
appropriate choice from among many such possible filters on the basis of the variation
of relative error with r for different transforms.
In this context, it is interesting to examine the manner of variation of error for a
given filter length n and given design and test functions, with the variation of a and s.
Figure 1 shows a typical plot of such errors as a mesh surface. This ‘error surface’ is
for 61-point filters for the Hankel J0 transform. The height of the surface above the
a–s plane is the logarithm of the root-mean-square relative error computed over a
chosen range of r. The design function is transform (6), and a large set of equations,
formed with log10(r) varying over the range ¹6 > log10 ðrÞ >+6, was solved to get the
filter weights. The test function for determining the error for each combination of a
and s was the same transform as the design function, but the error was computed for
201 equally spaced values of log10(r) over the smaller range ¹4 > log10 ðrÞ >+3. The
spacing of s for the mesh in this diagram is 0.0003 and the spacing for a is 0.1.

Figure 1. Typical variation of error with changes in shift a and spacing s, plotted as a surface on
the a–s plane. There is a region in the a–s plane where the error is generally low, but there are
some deep ‘grooves’ on the surface, and some very deep local minima on these grooves, which
are like spikes pointing downwards. (See text for details.)

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
750 D. Guptasarma and B. Singh

It may be seen that the error surface consists of the intersection of a number of
individual surfaces, each of which is convex upwards. Each of these surfaces consists
of many minor, intersecting surfaces which are also convex upwards. Consequently,
the lowest error in the vicinity of an intersection of two such surfaces lies on the
intersection itself. On the a–s plane, these intersections are broadly like ‘grooves’
which are approximately linear. As Fig. 1 shows, the error surface rises in all
directions around a region of low errors. The figure also shows many downward
‘spikes’ in the surface. Each of these is close to a local minimum in error, but perhaps
not exactly on it, because the mesh is too coarse to be able to show all the minima. A
few additional points, at which the error becomes very low, were added to the mesh to
show that there are many local minima. Our objective is to find the actual lowest point
among all these local minima, and choose the best among them.
Because the error surface is so complex in its details, none of the conventional
minimization approaches for finding the lowest point seem to work. The problem is
complicated by the fact that the slope of the error surface is discontinuous everywhere
on the intersections.
The complexity of the error surface was found to be similar to the surface in Fig. 1
for all the transforms we used. Our approach was first to make a coarse search to find
the approximate location of some of the grooves on the error surface, then to follow
these grooves to find the lowest error points on each of these grooves, and finally to
make a selection of the best among the low error filters found in this way.
We found that it was necessary to make very fine adjustments of a and s in order to
arrive at the very lowest point of a local minimum. We succeeded in setting up
computer algorithms for speeding up the process of search for the lowest point near a
given location by finding the nearest groove and moving along it until the lowest point
on the groove was found. Then we used the fact that an adjacent groove could be
captured by changing a by an amount equal to s, and proceeding to search for the
groove nearest to this new location. Changing a by an amount equal to s is equivalent
to omitting the sample location at one end of the filter and adding another one at the
other end.
In order to find the filter weights by solving the Wiener equations, it is necessary to
use high arithmetic accuracy in the computation. While setting up the equations it is
necessary to take care of underflow and overflow of the exponentials. We used a
computer program capable of handling from exp(¹730) to exp(þ710) without going
into under- or overflow. The coefficients that went into overflow or underflow were
forced to zero. In order to bypass the problem of rank deficiency in the set of
equations, we avoided those areas of the a–s plane where the set of equations became
rank deficient. When some row in the matrix of coefficients for the equations had too
many zeros, we stipulated that a minimum number of non-zero coefficients should be
present in each equation. This minimum number was different for different filters and
was set to obtain the largest possible number of useful equations for solving for the
filter coefficients.
If the function under transformation is a constant, the transformed value should be

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 751

equal to the same constant for both the Hankel transforms. This requires that the sum
of all the filter weights be equal to unity. However, the solution of the equations gives a
set of filter weights, the sum of which may be slightly different from unity. This was
corrected by adjusting the last filter weight so as to make the sum of all the weights
equal to unity.
The computations were carried out on an IBM-486 personal computer.

The filter parameters

Tables 1 to 4 in the Appendix show the parameters of the filters presented in this
paper. The filter weights and the values of the shift s were designed to be restricted to
12 significant digits. We recommend that all the digits of the filter parameters be
retained in order to have good error performance. Truncating the parameters could
seriously affect the accuracy of these filters. Truncating and rounding-off could also
change the sum of all the filter weights, making it different from unity.
The error performance of these filters is dependent on the accuracy of computation
of the samples of the input function. This requires that, in addition to computing the
samples of the function with adequate arithmetic accuracy, the given values of the
shift a and the spacing s be used as provided, so that the samples of the input function
are evaluated at the correct abscissa values.

Figure 2. Relative errors of the new 61-point and 120-point J0 filters, compared with errors of
the 61-point J0 filter given by Koefoed et al. (1972), marked Koef 61, the 283-point, single-
precision J0 filter given by Anderson (1979), marked Ander 283, and the 801-point, double
precision J0 filter given by Anderson (1982), marked Ander 801, for the transform (4).

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
752 D. Guptasarma and B. Singh

Performance of the filters

The performance of the new filters in relation to some other widely used filter
operators for the transforms (4), (5), (6), (8), (9) and (10) may be seen in Figs 2–7.
Each of these figures shows the relative errors made by the new filters and the errors
made by some older filters.
The new filters are seen to produce much smaller relative errors than the errors
made by the filters with which we have made comparisons. In most cases the
reduction in error is seen to be a few orders of magnitude over a large range of r.
We have plotted relative errors in all the figures except in Fig. 8 where, for the sake
of appropriate comparison with an earlier publication, the magnitudes of absolute
errors are plotted.
For the most part, the error plots are composed of small, continuous segments
going through a maximum. Actually the sign of the error frequently changes from one
segment to the next adjacent segment. This happens because, with the new filters, the
computed value stays close to the correct value, going slightly above and below it, as
the value of r is changed. Therefore, the error vanishes at every crossing of the plots of
the computed value and the correct value. The sharp downward excursions seen in
many parts of the error plots represent reduction of error due to this reason. The
performance of a filter should therefore be judged only by considering something like
a smooth upper envelope of the error plots, grazing the error maxima.
Figure 2 relates to transform (4). The new 61-point J0 filter is seen to produce an
error 5 orders of magnitude lower compared to that obtained with the 61-point filter

Figure 3. Relative errors of the same filters as in Fig. 2, but for transform (5).

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 753

of Koefoed, Ghosh and Polman (1972). It can be seen that this filter produces an
error which is lower by more than an order of magnitude compared to that obtained
using the 283-point J0 filter given by Anderson (1979), over a large range of r. The
new 120-point J0 filter is seen to be superior even to the 801-point, double precision,
J0 filter given by Anderson (1982), over a similar range of r. Although this 120-point
filter has less than half the number of coefficients compared to Anderson’s (1979)
283-point filter , its errors are 3 to 4 orders of magnitude lower over the entire range
of r.
Figures 3 and 4 show similar comparisons of the performance of the same filters for
transforms (5) and (6) respectively. For transform (5), the new 61-point filter gives
much lower errors compared to the filter of Koefoed et al. (1972) and the new 120-
point filter gives lower errors than the filter of Anderson (1979). However, for this
transform, the 801-point, double precision filter of Anderson is superior to the new
120-point filter for most values of r.
Figure 4 shows that the improvements achieved by the new 61-point and 120-point
filters for transform (6) are similar to those for transform (4), as shown in Fig. 2. The
error performance of the new 61-point filter far exceeds that of the filter of
Koefoed et al. (1972), and the performance of the new 120-point filter exceeds the
performance of even the 801-point, double precision filter of Anderson (1982).
Figures 5–7 show the performance of the new 47-point and 140-point J1 filters,
compared with older filters.
Figures 5 and 7 show that the new 47-point J1 filter gives a better than 8-digit
accuracy with transforms (8) and (10) for r varying from 0.001 to 10. This is about 3

Figure 4. Relative errors of the same filters as in Fig. 2, but for transform (6).

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
754 D. Guptasarma and B. Singh

Figure 5. Relative errors of the new 47-point and 140-point J1 filters, compared with errors of
the 47-point J1 filter given by Koefoed et al. (1972), marked Koef 61, the 283-point, single-
precision J1 filter given by Anderson (1979), marked Ander 283, and the 801-point, double
precision J1 filter given by Anderson (1982), marked Ander 801, for the transform (8).

Figure 6. Relative errors of the same filters as in Fig. 5, but for transform (9).

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 755

Figure 7. Relative errors of the same filters as in Fig. 5, but for transform (10).

to 4 orders of magnitude better than the performance of Koefoed’s (1972) J1 filter for
these transforms over the same range, and about 2 or more orders of magnitude better
than Anderson’s (1979) 283-point J1 filter. This is in spite of the fact that the sampling
density of the new 47-point filter (s ¼ 0.1106) is much less than that of Anderson’s
(1979) 283-point filter (s ¼ 0.0868) for the J1 transform.
It may be seen from the plots of errors in Figs 5 and 7, with transforms (8) and (10)
using the new 140-point J1 filter, that 10-digit (or better, up to about 14-digit)
accuracies are achieved over a very wide range of r. For these transforms, the new 140-
point filter performs much better than Anderson’s (1982) 801-point, double
precision filter.
Figure 6 shows the performance of the filters for transform (9). The new 140-point
filter is seen to perform much better than Anderson’s (1979) 283-point filter,
although it is somewhat inferior to the 801-point, double precision filter of Anderson
(1982).
Mohsen and Hashish (1994) have published a modified-shift FFT method of filter
design, and 9-significant-digit results of computation with transforms (4), (8) and
(10), using J0 and J1 filters having 256 weights. From the plots of results presented
here, our 120-point J0 filter and 140-point J1 filter can be seen to give much better
than 9-digit accuracy for the range of r (0.01 to 100) used by them for these
transforms (Mohsen and Hashish 1994, p. 135).
Figure 8 is similar to Fig. 1 of Christensen’s (1991) reply to Anderson’s (1991)
comment on the paper by Christensen (1990). In this figure we have shown a
comparison between absolute errors of our 120-point optimized J0 filter and

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
756 D. Guptasarma and B. Singh

Figure 8. Comparison of the absolute errors of the new 120-point J0 filter with some older
„ ∞ transform (6). G(v) ¼ r/[1 þ r ] , which is r times the theoretical value of the
2 3/2
filters, for
integral 0 l expð¹lÞJ0 ðrlÞ dl. G*(v) is the value of the same quantity as obtained by using one
of the filters. The plot of the errors obtained with Christensen’s (1990) method is obtained from
Christensen (1991). Ander 283 shows the errors with Anderson’s (1979) 283-point filter and
Ander 801 shows errors obtained with Anderson’s (1982) 801-point filter for the J0 transform,
using all the filter weights in each case.

Anderson’s (1982) 801-point, double precision filter, Anderson’s (1979) 283-point

filter (using all the 283 points) and Christensen’s (1990) filter, which we obtained
from Christensen (1991). The quantity plotted corresponds to transform (6), and is
„ ∞ of the absolute value of the difference between r/[1 þ r ] and the
2 3/2
the logarithm
value of r 0 l expð¹lÞJ0 ðrlÞ dl obtained by using the filters, against log10(r), just as in
Christensen (1991).
As pointed out by Christensen (1991), the comparison between absolute errors of
Christensen’s filter, which has a sampling interval s ¼ 0.1 (corresponding to 0.23 on
the natural log scale), and Anderson’s (1979) single precision filter having a sampling
interval of 0.0868 (corresponding to 0.2 on the natural log scale) is not quite exact.
But our new 120-point filter, which has a sampling interval of 0.0904 (corresponding
to 0.208 on the natural log scale), may be seen to give an error between 2 to 5 orders of
magnitude lower compared to the error of Anderson’s (1979) filter, over the entire
range of r, and much lower errors compared to Anderson’s (1982) filter, over a large
range of r. It would have been interesting to see what errors would actually be
obtained with a filter designed according to Christensen (1990), with the same
spacing of 0.0904 as in our 120-point filter, and whether such a filter of length similar
to our new filter and 12-digit coefficients, would perform as well, or better.

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 757

Conclusions
It has been found that, with some modifications, the method given by Guptasarma
(1982a) for the design of optimized short digital linear filters is also useful for
designing some longer filter operators. The new filters for the Hankel transforms with
J0 and J1 kernels presented in this paper show a better error performance than all other
older filters available for these transforms, some of which are much longer, have a
higher sampling density, and use 16-digit weights compared to the 12-digit weights
used in our filters.
It could be an interesting exercise to see if any of the other existing methods of filter
design produces better Hankel filters of the same length. It would also be useful to
design more such filters for various other useful linear transforms.

Acknowledgements
The authors are grateful to the Director, NGRI, for his kind permission to publish
this paper. Thanks are due to Dr. C. P. Gupta and Mr. Y.M. Ramachandra for making
an IBM PC available for some time, and to Ms. M. Kousalya for assistance in the
preparation of the manuscript.
The work was carried out as part of the research programme of one of the authors,
(D.G.), as an Emeritus Scientist of the Council of Scientific and Industrial Research,
New Delhi.

Appendix

The filter parameters

The parameters of the four filters are given below (Tables 1–4). The numbers should
be used as given, without truncation. The arithmetic accuracy of the computer
program used for numerical evaluation should be 12 digits, or more. Equations (1) to
(3) in the main text show how to use the filter parameters to compute the value of a
given function at a given value of r.
The filter weights are arranged row by row. The first weight is at the beginning of
the first row, the next weight is the next number in the same row, and so on, to the end
of the row; the following weight appears at the beginning of the next row, and so on, to
the end.

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
758 D. Guptasarma and B. Singh

Table 1. Parameters of the 61-point J0 filter.

a ¼ ¹5.08250000000eþ00 s ¼ 1.16638303862e¹01

Filter weights 1 to 61:

3.30220475766e¹04 ¹1.18223623458e¹03 2.01879495264e¹03 ¹2.13218719891e¹03

1.60839063172e¹03 ¹9.09156346708e¹04 4.37889252738e¹04 ¹1.55298878782e¹04
7.98411962729e¹05 4.37268394072e¹06 3.94253441247e¹05 4.02675924344e¹05
5.66053344653e¹05 7.25774926389e¹05 9.55412535465e¹05 1.24699163157e¹04
1.63262166579e¹04 2.13477133718e¹04 2.79304232173e¹04 3.65312787897e¹04
4.77899413107e¹04 6.25100170825e¹04 8.17726956451e¹04 1.06961339341e¹03
1.39920928148e¹03 1.83020380399e¹03 2.39417015791e¹03 3.13158560774e¹03
4.09654426763e¹03 5.35807925630e¹03 7.00889482693e¹03 9.16637526490e¹03
1.19891721272e¹02 1.56755740646e¹02 2.04953856060e¹02 2.67778388247e¹02
3.49719672729e¹02 4.55975312615e¹02 5.93498881451e¹02 7.69179091244e¹02
9.91094769804e¹02 1.26166963993e¹01 1.57616825575e¹01 1.89707800260e¹01
2.13804195282e¹01 2.08669340316e¹01 1.40250562745e¹01 ¹3.65385242807e¹02
¹2.98004010732e¹01 ¹4.21898149249e¹01 5.94373771266e¹02 5.29621428353e¹01
¹4.41362405166e¹01 1.90355040550e¹01 ¹6.19966386785e¹02 1.87255115744e¹02
¹5.68736766738e¹03 1.68263510609e¹03 ¹4.38587145792e¹04 8.59117336292e¹05
¹9.15853765160e¹06

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 759

Table 2. Parameters of the 120 ¹ point J0 filter.

a ¼ ¹8.38850000000e þ 00 s ¼ 9.04226468670e ¹ 02

Filter weights 1 to 120:

9.62801364263e¹07 ¹5.02069203805e¹06 1.25268783953e¹05 ¹1.99324417376e¹05

2.29149033546e¹05 ¹2.04737583809e¹05 1.49952002937e¹05 ¹9.37502840980e¹06
5.20156955323e¹06 ¹2.62939890538e¹06 1.26550848081e¹06 ¹5.73156151923e¹07
2.76281274155e¹07 ¹1.09963734387e¹07 7.38038330280e¹08 ¹9.31614600001e¹09
3.87247135578e¹08 2.10303178461e¹08 4.10556513877e¹08 4.13077946246e¹08
5.68828741789e¹08 6.59543638130e¹08 8.40811858728e¹08 1.01532550003e¹07
1.26437360082e¹07 1.54733678097e¹07 1.91218582499e¹07 2.35008851918e¹07
2.89750329490e¹07 3.56550504341e¹07 4.39299297826e¹07 5.40794544880e¹07
6.66136379541e¹07 8.20175040653e¹07 1.01015545059e¹06 1.24384500153e¹06
1.53187399787e¹06 1.88633707689e¹06 2.32307100992e¹06 2.86067883258e¹06
3.52293208580e¹06 4.33827546442e¹06 5.34253613351e¹06 6.57906223200e¹06
8.10198829111e¹06 9.97723263578e¹06 1.22867312381e¹05 1.51305855976e¹05
1.86329431672e¹05 2.29456891669e¹05 2.82570465155e¹05 3.47973610445e¹05
4.28521099371e¹05 5.27705217882e¹05 6.49856943660e¹05 8.00269662180e¹05
9.85515408752e¹05 1.21361571831e¹04 1.49454562334e¹04 1.84045784500e¹04
2.26649641428e¹04 2.79106748890e¹04 3.43716968725e¹04 4.23267056591e¹04
5.21251001943e¹04 6.41886194381e¹04 7.90483105615e¹04 9.73420647376e¹04
1.19877439042e¹03 1.47618560844e¹03 1.81794224454e¹03 2.23860214971e¹03
2.75687537633e¹03 3.39471308297e¹03 4.18062141752e¹03 5.14762977308e¹03
6.33918155348e¹03 7.80480111772e¹03 9.61064602702e¹03 1.18304971234e¹02
1.45647517743e¹02 1.79219149417e¹02 2.20527911163e¹02 2.71124775541e¹02
3.33214363101e¹02 4.08864842127e¹02 5.01074356716e¹02 6.12084049407e¹02
7.45146949048e¹02 9.00780900611e¹02 1.07940155413e¹01 1.27267746478e¹01
1.46676027814e¹01 1.62254276550e¹01 1.68045766353e¹01 1.52383204788e¹01
1.01214136498e¹01 ¹2.44389126667e¹03 ¹1.54078468398e¹01 ¹3.03214415655e¹01
¹2.97674373379e¹01 7.93541259524e¹03 4.26273267393e¹01 1.00032384844e¹01
¹4.94117404043e¹01 3.92604878741e¹01 ¹1.90111691178e¹01 7.43654896362e¹02
¹2.78508428343e¹02 1.09992061155e¹02 ¹4.69798719697e¹03 2.12587632706e¹03
¹9.81986734159e¹04 4.44992546836e¹04 ¹1.89983519162e¹04 7.31024164292e¹05
¹2.40057837293e¹05 6.23096824846e¹06 ¹1.12363896552e¹06 1.04470606055e¹07

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
760 D. Guptasarma and B. Singh

Table 3. Parameters of the 47-point J1 filter.

a ¼ ¹ 3.05078187595eþ00 s ¼ 1.10599010095e¹01

Filter weights 1 to 47:

3.17926147465e¹06 ¹9.73811660718e¹06 1.64866227408e¹05 ¹1.81501261160e¹05

1.87556556369e¹05 ¹1.46550406038e¹05 1.53799733803e¹05 ¹6.95628273934e¹06
1.41881555665e¹05 3.41445665537e¹06 2.13941715512e¹05 2.34962369042e¹05
4.84340283290e¹05 7.33732978590e¹05 1.27703784430e¹04 2.08120025730e¹04
3.49803898913e¹04 5.79107814687e¹04 9.65887918451e¹04 1.60401273703e¹03
2.66903777685e¹03 4.43111590040e¹03 7.35631696247e¹03 1.21782796293e¹02
2.01097829218e¹02 3.30096953061e¹02 5.37143591532e¹02 8.60516613299e¹02
1.34267607144e¹01 2.00125033067e¹01 2.74027505792e¹01 3.18168749246e¹01
2.41655667461e¹01 ¹5.40549161658e¹02 ¹4.46912952135e¹01 ¹1.92231885629e¹01
5.52376753950e¹01 ¹3.57429049025e¹01 1.41510519002e¹01 ¹4.61421935309e¹02
1.48273761923e¹02 ¹5.07479209193e¹03 1.83829713749e¹03 ¹6.67742804324e¹04
2.21277518118e¹04 ¹5.66248732755e¹05 7.88229202853e¹06

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
 Digital linear filters 761

Table 4. Parameters of the 140-point J1 filter.

a ¼ ¹ 7.91001919000eþ00 s ¼ 8.79671439570e¹02

Filter weights 1 to 140:

¹6.76671159511e¹14 3.39808396836e¹13 ¹7.43411889153e¹13 8.93613024469e¹13

¹5.47341591896e¹13 ¹5.84920181906e¹14 5.20780672883e¹13 ¹6.92656254606e¹13
6.88908045074e¹13 ¹6.39910528298e¹13 5.82098912530e¹13 ¹4.84912700478e¹13
3.54684337858e¹13 ¹2.10855291368e¹13 1.00452749275e¹13 5.58449957721e¹15
¹5.67206735175e¹14 1.09107856853e¹13 ¹6.04067500756e¹14 8.84512134731e¹14
2.22321981827e¹14 8.38072239207e¹14 1.23647835900e¹13 1.44351787234e¹13
2.94276480713e¹13 3.39965995918e¹13 6.17024672340e¹13 8.25310217692e¹13
1.32560792613e¹12 1.90949961267e¹12 2.93458179767e¹12 4.33454210095e¹12
6.55863288798e¹12 9.78324910827e¹12 1.47126365223e¹11 2.20240108708e¹11
3.30577485691e¹11 4.95377381480e¹11 7.43047574433e¹11 1.11400535181e¹10
1.67052734516e¹10 2.50470107577e¹10 3.75597211630e¹10 5.63165204681e¹10
8.44458166896e¹10 1.26621795331e¹09 1.89866561359e¹09 2.84693620927e¹09
4.26886170263e¹09 6.40104325574e¹09 9.59798498616e¹09 1.43918931885e¹08
2.15798696769e¹08 3.23584600810e¹08 4.85195105813e¹08 7.27538583183e¹08
1.09090191748e¹07 1.63577866557e¹07 2.45275193920e¹07 3.67784458730e¹07
5.51470341585e¹07 8.26916206192e¹07 1.23991037294e¹06 1.85921554669e¹06
2.78777669034e¹06 4.18019870272e¹06 6.26794044911e¹06 9.39858833064e¹06
1.40925408889e¹05 2.11312291505e¹05 3.16846342900e¹05 4.75093313246e¹05
7.12354794719e¹05 1.06810848460e¹04 1.60146590551e¹04 2.40110903628e¹04
3.59981158972e¹04 5.39658308918e¹04 8.08925141201e¹04 1.21234066243e¹03
1.81650387595e¹03 2.72068483151e¹03 4.07274689463e¹03 6.09135552241e¹03
9.09940027636e¹03 1.35660714813e¹02 2.01692550906e¹02 2.98534800308e¹02
4.39060697220e¹02 6.39211368217e¹02 9.16763946228e¹02 1.28368795114e¹01
1.73241920046e¹01 2.19830379079e¹01 2.51193131178e¹01 2.32380049895e¹01
1.17121080205e¹01 ¹1.17252913088e¹01 ¹3.52148528535e¹01 ¹2.71162871370e¹01
2.91134747110e¹01 3.17192840623e¹01 ¹4.93075681595e¹01 3.11223091821e¹01
¹1.36044122543e¹01 5.12141261934e¹02 ¹1.90806300761e¹02 7.57044398633e¹03
¹3.25432753751e¹03 1.49774676371e¹03 ¹7.24569558272e¹04 3.62792644965e¹04
¹1.85907973641e¹04 9.67201396593e¹05 ¹5.07744171678e¹05 2.67510121456e¹05
¹1.40667136728e¹05 7.33363699547e¹06 ¹3.75638767050e¹06 1.86344211280e¹06
¹8.71623576811e¹07 3.61028200288e¹07 ¹1.05847108097e¹ 07 ¹1.51569361490e¹08
6.67633241420e¹08 ¹8.33741579804e¹08 8.31065906136e¹ 08 ¹7.53457009758e¹08
6.48057680299e¹08 ¹5.37558016587e¹08 4.32436265303e¹08 ¹3.37262648712e¹08
2.53558687098e¹08 ¹1.81287021528e¹08 1.20228328586e¹08 ¹7.10898040664e¹09
3.53667004588e¹09 ¹1.36030600198e¹09 3.52544249042e¹10 ¹4.53719284366e¹11

References
Anderson W.L. 1975. Improved digital filters for evaluating Fourier and Hankel transform
integrals. U.S.G.S. rep. USGS-GD-75-012; avail. as NTIS rep. PB-242-800/1WC, p 223.
Anderson W.L. 1979. Computer Program. Numerical integration of related Hankel transforms
of orders 0 and 1 by adaptive digital filtering. Geophysics 44, 1287–1305.

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762
762 D. Guptasarma and B. Singh

Anderson W.L. 1982. Fast Hankel transforms using related and lagged convolutions. ACM
Transactions on Mathematical Software 8, 344–368.
Anderson W.L. 1989. A hybrid fast Hankel transform algorithm for electromagnetic modeling.
Geophysics 54, 263–266.
Anderson W.L. 1991. Comment on ‘‘Optimized fast Hankel transform filters’’ by Niels Boie
Christensen. Geophysical Prospecting 39, 445–447.
Christensen N.B. 1990. Optimized fast Hankel transform filters. Geophysical Prospecting 38,
545–558.
Christensen N.B. 1991. Reply to comments by Walter L. Anderson. Geophysical Prospecting 39,
449–450.
Das U.C. and Ghosh D.P. 1974. The determination of filter coefficients for the computation of
standard curves for dipole resistivity sounding over layered earth by digital linear filtering.
Geophysical Prospecting 22, 765–780.
Das U.C., Ghosh D.P. and Biewinga D.T. 1974. Transformation of dipole resistivity sounding
measurements over layered earth by linear digital filtering. Geophysical Prospecting 22, 476–489.
Ghosh D.P. 1971. The application of linear filter theory to the direct interpretation of
geoelectrical resistivity sounding measurements. Geophysical Prospecting 19, 192–217.
Guptasarma D. 1982a. Optimization of short digital linear filters for increased accuracy.
Geophysical Prospecting 30, 501–514.
Guptasarma D. 1982b. Computation of the time-domain response of a polarizable ground.
Geophysics 47, 1574–1576.
Johansen H.K. and Sorensen K. 1979. Fast Hankel transforms. Geophysical Prospecting 27,
876–901.
Koefoed O. 1972. A note on the linear filter method of interpreting resistivity sounding data.
Geophysical Prospecting 20, 403–405.
Koefoed O. 1976. Error propagation and uncertainty in the interpretation of resistivity
sounding data. Geophysical Prospecting 24, 31–48.
Koefoed O. 1979. Geosounding principles. 1. Resistivity sounding measurements. In: Methods
in Geochemistry and Geophysics 14A. Elsevier Science Publishing Co.
Koefoed O. and Dirks F.J. 1979. Determination of resistivity sounding filters by the Wiener-
Hopf least-squares method. Geophysical Prospecting 27, 245–250.
Koefoed O., Ghosh D.P. and Polman G.J. 1972. Computation of type curves for
electromagnetic depth sounding with a horizontal transmitting coil by means of a digital
linear filter. Geophysical Prospecting 20, 406–420.
Kumar R. and Das U.C. 1977. Transformation of dipole to Schlumberger sounding curves by
means of digital linear filters. Geophysical Prospecting 25, 780–789.
Mohsen A.A. and Hashish E.A. 1994. The fast Hankel transform. Geophysical Prospecting 42,
131–139.
Nissen J. and Enmark T. 1986. An optimized digital filter for the Fourier transform. Geophysical
Prospecting 34, 897–903.
Rijo L. 1993. An optimized digital filter for the cosine transform. Revista Brasileira de Geofisica
(Brazilian Geophysical Journal) 10, 7–20.
Rijo L. 1996. Comment on ‘The fast Hankel transform’ by A.A. Mohsen and E.A. Hashish.
Geophysical Prospecting 44, 473–477.
Sorensen K.I. and Christensen N.B. 1994. The fields from a finite electrical dipole—A new
computational approach. Geophysics 59, 864–880.

q 1997 European Association of Geoscientists & Engineers, Geophysical Prospecting, 45, 745–762