Chem.

Educator 1999, 4, 8188

81

Beers Law Is Not a Straight Line: Amplification of Errors by

Transformation
Judith M. Bonicamp,* Kelly L. Martin, Gerald R. McBride and Roy W. Clark
Department of Chemistry, Middle Tennessee State University, Murfreesboro, TN 37132, bonicamp@mtsu.edu

Abstract: Beers law plots are commonly straight-line graphs in which absorbance is plotted versus
concentration. The errors in such a plot are accentuated in the high-absorbance region due to the logarithmic
transformation. This paper discusses the errors inherent in all Beers law plots and the choices the student has
available: the use of either linear or nonlinear plotting methods.
An experiment is described in which student Beers law data are treated in several ways to find out if
transformation errors are significant. The conclusion is that if the absorbance values obtained for the calibration
data exceed A = 1, then nonlinear regression or weighted linear regression is indicated.

When
using
radiation
absorption
instruments
(spectrophotometers) to determine concentrations, the straightline relationship between absorbance and concentration seems
a simple analytical tool. It is not that simple, however, for the
statement that absorbance is proportional to concentration is a
mathematical transformation of the exponential nature of the
transmittanceconcentration relationship. This transformation,
A = log T, has a striking influence on the errors that occur in
experiments, particularly in regions of high absorbance. We set
out to determine if these errors make much difference in the
typical student Beers law determination of an unknown
concentration.
Instruments used in chemistry to measure the concentration
of a species of interest fall into two categories. Either the
instrument response (called the signal) is directly proportional
to the concentration of the species, or it is not. If it is not, then
some known function relates the signal to the concentration.
When using instrumentation to determine an unknown
concentration, it is important to know which type your
instrument is because it affects the error analysis of the result.
Electrical conductance meters, fluorometers, and osmometers
are three examples of instruments of the first type. Their signal
is, at least approximately, proportional to concentration. Any
nonlinearities can be taken into account by suitably small
calibration ranges or by curve fits to the calibration points.
Many students are surprised to find out that Beers law
instruments, which means all absorption spectrophotometers,
fall into the second category of instruments. The origin of the
surprise is that students think that the Beers law signal is
absorbance, and we all know that absorbance is linearly related
to concentration. But the signal is transmittance, not
absorbance; therefore, because of the transformation to a
logarithmic function, error analysis for Beers law experiments
should not be done using the usual straight-line error analysis
equations [15].
The mathematical transformation, logT = A, is commonly
performed either by the analyst or by the instrument itself, and
A is proportional to concentration. Thus, the familiar A = abc
form of Beers law is better known than the true Beers law
equation, the exponential decay equation.
Beers law, more correctly called the BeerLambert law, is a
statement that the intensity of light, I, exiting from an optical

absorption cell is exponentially related to the intensity entering

the cell,
I = I 0 10 kC

(1)

Here, I0 is the intensity of the entering light, I is the intensity

of the exiting light, C is the concentration of absorbing species
in the solution, and k is a decay constant. This exponential
decay is far from a straight line (Figure 1a).
Most of us think of this law as a straight line because of a
transformation of the variables. The ratio I divided by I0 is
defined as T, the transmittance, and log T is defined as A, the
absorbance. Thus A= kC. Plot A versus C and a straight line
results, one that in theory passes through zero. By plotting the
data transformed, our eyes are able to envision the straight-line
data fit (Figure 1b) much easier than the decaying-exponential
one.
One more change is usually made to the equation. The
constant k depends upon the absorbing species (its optical
properties at the wavelength chosen) and on the cell path
length. It is usual to write
A = abC or A = bC

(2)

In these two forms b is the path length, usually in

centimeters, and a or are decay constants characteristic of the
absorbing species. When is used, the concentrations are in
moles per liter, and is called the molar absorptivity. In older
texts is called the molar extinction coefficient. If a is used
the concentration units can be anything convenient, including
1
mol L , and a is usually simply called the absorptivity. In our
linear plots the ab in equation 1 is combined into k, the slope
of the line as determined by regression.
To illustrate the errors that can arise in spectrophotometry,
we generated a set of perfect Beers law data with a
spreadsheet. Using the equation
T = T0 101k C

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

(3)

82

Chem. Educator, Vol. 4, No. 3, 1999

Bonicamp et. al.

Table 1. PERFBEER Data from T = T0 101kC with k = 15 L mol1,

T0 = 1
Concentration
mol L1
0
0.004
0.008
0.012
0.016
0.020
0.024
0.028
0.032
0.036
0.040
0.044
0.048
0.052
0.056
0.060
0.064
0.068
0.072
0.076
0.080
0.084
0.088
0.092
0.096
0.10
0.104
0.108
0.112
0.116
0.120
0.124
0.128
0.132

T
unitless ratio
1.000000
0.870964
0.758578
0.660693
0.57544
0.501187
0.436516
0.380189
0.331131
0.288403
0.251189
0.218776
0.190546
0.165959
0.144544
0.125893
0.109648
0.095499
0.083176
0.072444
0.063096
0.054954
0.047863
0.041687
0.036308
0.031623
0.027542
0.023988
0.020893
0.018197
0.015849
0.013804
0.012023
0.010471

A
log T
0.000
0.060
0.120
0.180
0.240
0.300
0.360
0.420
0.480
0.540
0.600
0.660
0.720
0.780
0.840
0.900
0.960
1.020
1.080
1.140
1.200
1.260
1.320
1.380
1.440
1.500
1.560
1.620
1.680
1.740
1.800
1.860
1.920
1.980

Figure 1. PERFBEER data from T=T0 10-1kC with K=15 L

mol-1, T0=1 and C as shown. A). Plot of T vs. C in green. B). Plot of
-logT vs. C in red.
1

with k =15 L mol , T 0 = 1, and C values from 0.004 to 0.132

1
mol L , we generated the data table PERFBEER (Table 1). T 0
stands for an adjustable parameter, which in ideal data sets
will always equal 1, but in real data sets will not necessarily be

1, but a parameter adjusted by the nonlinear regression

program. See Figures 1a and 1b. The intercept, i, plotted
linearly will be log T 0. For plotting purposes we use the
symbol i as the intercept of a linear plot and k as the slope of
that plot. Again, the parameter k is in reality ab in the
commonly seen form of Beers law, A = abc, and the
parameter i is the y intercept of the linear plot. When we do
nonlinear plots we call the pre-exponential factor T0 and the
decay constant k because it is numerically the same as the line
slope in the linear form.
Errors in Exponential Decay Curves and Distortion of This
Error by Transformation
If instruments actually measured A rather than T, then a
constant error in measurement of A would mean a constant
error in deducing the concentration from that A reading.
Because T is what the instrument measures, a look at the
exponential decay curve (Figure 1a) shows that a constant
error in T, call it T, would not mean a constant error in
concentration. The varying slope of the line shows that the
resulting error in concentration, C, would be small at low
concentrations, high at large concentrations.
The inverse of the slope times the constant error in T would
be the resulting error in C, the concentration error. This error
would be small (but not zero) at low concentration values, and
with increasing concentration it would become very large.
The relative error in concentration would be this error
divided by the concentration (relative error = C/C). The
percent error is, of course, this times 100. The division by C
results in a relative error that is more favorable at some
concentration values than at others. This can be illustrated by
taking some perfect Beers law data and introducing a
constant transmittance error into each point.
Constant Error in T
As previously described, the plots in Figures 1a and 1b were
created by generating perfect Beers law data using a
spreadsheet. By simulating error in this data it is easy to see
the different ways that errors show up in the two kinds of
plots, exponential and linear. After seeing these plots with
constant simulated error, we became convinced that nonlinear
regression on the T versus C plot is to be preferred over the
usual linear regression on the A versus C plot. In the computer
age, nonlinear regression is no more difficult than beststraight-line fits (Please see Endnote 1 located in the
supporting material, 43jb1897.pdf).
We introduced a constant error of +0.01 into each data
point. It is important to realize that the error is introduced into
T, not A. Instruments measure T, not A. If an instrument gives
a read-out of A, it is because the instrument converted from
transmittance to absorbance by taking the negative logarithm
of T. The resulting plot (Figure 2a) looks much the same as the
perfect data plot (Figure 1a). After introduction of this error,
the parameters reported by nonlinear regression were T 0 =
1.0026 and k = 14.5. Note that this constant error made about a
3% difference in the k parameter.
Next, we plotted the same data in the transformed (linear)
way (Figure 2b). At high concentrations the small constant
error in A. The linear regression parameters for this set of

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

Beers Law Is Not a Straight Line

Chem. Educator, Vol. 4, No. 3, 1999

B
Figure 2. A). Plus 0.01 error in T. Data with 0.01 error,

Perfect line before error, Nonlinear regression line. B). Plus

83

Figure 3. A) Minus 0.01 error in T.

Data with 0.01 error,
Perfect line before error, Nonlinear regression line. B)

Data with 0.01 error, Perfect line

0.01 error in T.
before error, Linear regression line.

Data with 0.01 error,

Minus 0.01 error in T.
line before error, Linear regression line.

error in T gets transformed into a surprisingly large, negative

points are intercept, i, equal to 0.0471 and slope, k, equal to
13.12. The slope of the regression line is 12.5% too low.
Repeating the process nonlinearly with a 0.01 error results
in small changes (Figure 3a), but an even more striking
deviation from straightness is obtained when plotted with a
transformed variable (Figure 3b). Here the intercept is 0.1626
and the slope is 19.83. The slope of the regression line is
32.2% too high.
Plotting both T 0.01 data in one graph and doing nonlinear
regression on the combined set of data points yields the correct
result (Figure 4a). The result is correct because the errors have
canceled. But plotting combined plus and minus errors linearly
is a very different story. See Figure 4b. Here, because the
transformed errors do not cancel, the best straight line is not
good at all. In all fairness we should point out that such gross
errors do not result until the absorbance exceeds unity.

A Closer Look at the T Versus C Data Before

Transformation

Perfect

The T versus C plot does not exhibit these gross errors

because it was the transformation to A which caused them.
Still, there are errors before transformation because of the
curvature of the T versus C plot. If we inspect the middle
region of this plot (Figure 5), we see that the slope is
intermediate here, and that the steeper this slope the more
accurate the determination should be. What can we say about
the error in concentration introduced by our 0.01 errors in T?
We can see that the error in concentration, C, gets worse as
the concentration increases, because the slope of the line is
decreasing. But analytical chemists are interested in the
relative error, C divided by C. What happens to the relative
error in moderate, high, and in low concentration regions?
Remember that relative error is inversely proportional to C and
also inversely proportional to the slope of the line.

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

84

Chem. Educator, Vol. 4, No. 3, 1999

Bonicamp et. al.

Error at High Concentrations. In the high C region the
relative error would seem to get smaller still because of C
itself getting larger, except that the slope of the line gets
extremely small. This smaller slope increases C for a given
T and increases the error faster than can be compensated for
by the increasing C in the denominator. Thus, the relative error
goes up in this region.
Error at Low Concentrations. Inspecting the low C
region, we see that the slope is favorable to yield a small C
for a given T, but the C itself is small, and because C is in the
denominator of the relative error expression the result is that
the smallness of C overrides the bigness of the slope, and the
relative error becomes increasingly large at these small
concentrations.

The TwymanLothian Plot

F. Twyman and G. F. Lothian recognized all of this in 1933,
and published [6] a simple derivation of the relative error
versus transmittance equation:
1 dC
0.434
=
C dT T log T

B
Figure 4. A) Both plus and minus 0.01 error in T. Perfect line and
nonlinear regression line are the same line. B) Both plus and minus

Data with 0.01 error, Perfect line before

error in T.
error, Linear regression line.

(4)

(Please see Endnote 2 located in the supporting material,

43jb1897.pdf.)
Because dC/dT is always negative, plotting 1/C dC/dT
versus T gives a minimum relative error plot that is familiar to
most instrumental analysis students [7]. (See Figure 6A.) The y
axis of a TwymanLothian plot is the relative fractional error
in the concentration for each fractional error in measurement
of T. Put another way, it is the relative standard deviation of
the concentration answer for each relative standard deviation
of the T values.
Errors are minimized when T is in the range of 0.1 to 0.75,
which corresponds to A values between 1 and 0.12. Outside of
these ranges we expect greater error. Figure 6b is the same plot
of relative error, but plotted versus A. These TwymanLothian
plots show probable relative error in the unknown for a given
experimental error, T. This error will be most pronounced
outside the regions shown because of the exponential nature of
T. It is important to note that amplification of these errors by
conversion to logarithms is not shown on these plots, and that
this extra error is most pronounced in the high A (low T)
region.
Real Data

Figure 5. Three regions of the T versus C plot. !!!!!! Data with

0.01 error, Perfect line before error.

Error at Moderate Concentrations. Because C is

moderate and T/C is moderate, the relative error is a
minimum in this region.

A recent paper in the literature [8] titled The Remarkable

Resilience of Beers Law seems to suggest that Beers law
can be used in these regions of very low and very high
absorbance, despite TwymanLothian errors and transformation errors. This data set of Muyskens and Sevy is not
typical Beers law data as students might collect it. It is data on
very-low-pressure gas samples collected using very
sophisticated laboratory equipment. The light source was a
laser, and the detector was a special time-averaging pulse
detector. Such results may mislead students into thinking that
Beers law should work fine for their solutions at all values of
absorbance (Please see Endnote 3 located in the supporting
material, 43jb1897.pdf).

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

Beers Law Is Not a Straight Line

Chem. Educator, Vol. 4, No. 3, 1999

85

Table 2. Student Percent T Data Taken on a Spectronic 20D for

Nickel(II) Nitrate Solutions

Concentration

Student A

Student B

Student C

Student D

mol L1

%T

%T

%T

%T

0.004
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.20
0.30
0.40
0.50
Unk. soln.

95.2
89.2
79.0
70.0
62.3
55.8
49.6
43.4
39.2
34.5
30.6
26.8
24.2
21.6
19.6
17.9
16.0
13.8
12.4
11.2
10.2
4.8
2.1
1.0
67.0

88.0
78.8
68.5
61.0
54.5
47.5
42.0
38.0
33.0
30.0

88.2
78.6
69.7
61.0
54.8
48.0
42.5
38.0
33.1
29.6

67.0

67.0

88.2
77.8
68.5
60.5
54.2
47.4
41.7
37.1
32.8
29.9
27.4
24.8
21.5
19.4
18.7
16.9
13.2
12.4
11.4
10.6
4.6
2.9
1.1
64.9

What to Do with Beers Law Data

B
Figure 6. A). The Twyman-Lothian plot of relative error versus T. B).
The Twyman-Lothian plot of relative error versus A.

Student Data
We hope to answer the question, Should students plot
Beers law data nonlinearly (T versus C) and do nonlinear
regression, or is it satisfactory to plot A versis C and do linear
regression, as probably most do? Real student solution data
are expected to show errors, particularly outside of the
recommended concentration range. In order to answer the
question posed above, Does it make any difference?, we
analyzed student data taken on an inexpensive
spectrophotometer. The instructor is presumed to have the
right answer to the unknowns. Students A, B, C, and D took
data (Table 2) on a Milton Roy Company Spectronic 20D
spectrophotometer at 395 nm using standard solutions, which
they prepared by serial dilution from a 1.00 M nickel(II) nitrate
stock solution [9]. The path length of the cells was 1.165 cm.
Two of the data sets were kept in the recommended range, but
the other two sets were extended intentionally beyond this
region to answer the question about which plotting method is
superior. The students were also given an unknown solution
which fell in the range 0.01 to 0.10 M nickel(II) nitrate.

We recommend that when a student has finished taking

data, the next step should be to make a hand-drawn graph of
these data on graph paper. For the graph, absorbances should
be calculated from transmittance readings and a plot of A
versus C prepared. The graph should have points only, no
lines, and certainly not lines connecting each point. Inspection
of the graph should lead the student to one of five courses of
action as follows:
1) The points seem to fall on a line, and the unknown is
within these calibration points. Conclusion: Use a
computer to do standard linear regression on the
calibration points and deduce the concentration of the
unknown from the parameters i and k for this leastsquares line.
2) The points do not fall on a straight line. The deviations
are in the high A region. The unknown absorbance falls
within the straight region. Conclusion: Do standard linear
regression, ignoring (that is not entering) the high A
outliers.
3) The points do not fall on a straight line. The deviations
are in the high A region. The unknown absorbance falls in
or near the curved region. Conclusion: Do one of the
following procedures.
a) Perform a nonlinear regression analysis on the T vs.
C data. Use the T0 parameter to calculate i from the
equation logT0 = i. Use i and k as you would with
linear regression to calculate the unknown
concentration.

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

86

Chem. Educator, Vol. 4, No. 3, 1999

Bonicamp et. al.

Table 3. Student Results (from Table 2) Treated Two Ways

Student Results: Nonlinear Data Analysis
Student A
Parameters returned
i = 0.0882
k = 4.095
Parameter confidence
i = 0.3
limits
k = 0.018
Equation of best line
A = 0.088+
4.1 C
Unknown answer
0.022
reported
0.01
Percent error
-33%

Parameters returned
Parameter confidence
Limits (95%)
Equation of line
Unknown answer
reported
Percent error

Student A
T0 = 0.997
k = 5.069
T0 0.008
k 0.07
T = 0.997
10-5.069 C
0.034 M
0.001
3%

Student B
i = 0.0034
k = 5.26
i = 0.003
k = 0.051
A = 0.003+
5.26 C
0.032
0.0009
-3%
Student Results: Nonlinear Data Analysis
Student B
T0 = 0.996
k = 5.29
T0 0.01
k 0.09
T = 0.996
10-5.29 C
0.033 M
0.001
0%

Student C
i = -0.00029
k = 5.298
i = 0.35
k = 0.021
A = 0+
5.298 C
0.033
0.0007
0%

Student D
i = 0.1207
k = 3.868
i = 0.002
k = 0.029
A = 0.12+
3.8 C
0.017
0.02
-48%

Student C
T0 = 0.999
k = 5.288
T0 0.0007
k 0.07
T = 0.999
10-5.288 C
0.033 M
0.0002
0%

Student D
T0 = 0.973
k = 5.045
T0 0.017
k 0.138
T = 0.97
10-5.04 C
0.035 M
0.003
6%

Figure 7. Student preliminary plots from data in Table 2.

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

Beers Law Is Not a Straight Line

Chem. Educator, Vol. 4, No. 3, 1999

Table 4. Five Methods Applied to Student Data. Numbers are the

Percent Error Using Each Method on the Unknown Solution
(Concentration 0.033 M)
Student
Nonlinear (exponential)
Nonlinear (quadratic)
Linear (weighted)
Linear (cropped)
Linear (regular)

A
3%
-5%
6.7%
3%
-33%

B
0%
-2%
-1.5%
4.8%
-3%

C
0%
0.6%
-0.6%
0%

D
6%
4.2%
11.2%
-48%

87

analysis of the confidence limits for the answer is somewhat

easier for linear plots [12], it may not be worth the trouble to
go to nonlinear regression when the data do not exhibit
curvature.
In the two cases of the extended data ranges, the
transformed data is definitely not linear, and the simple linearleast-squares data analysis gives awful answers (Figure 8). The
students were able to see that the exaggerated curvature
(Figure 7) caused the poor results by linear analysis. These are
the data sets that needed cropping, since the unknown falls in
the linear region.
Conclusion

Figure 8. Error Bars for Linear and Nonlinear Regression.

b) Use nonlinear regression to fit a quadratic equation to

the A vs. C data. Solve the quadratic equation to find
the concentration of the unknown. This technique is
inferior to 3(a) above and 3(c) below, but it is better
than simple linear regression.
c) Perform a weighted linear regression on the A vs. C
2
data. Use a weight factor of T [16]. Use the parameters
reported to calculate the unknown concentration.
(Please see Endnote 4 located in the supporting material,
43jb1897.pdf)
Taking Our Advice
Table 2 contains the student data. Figure 7 shows the
student data plotted linearly. Students B and C could see that
the short data sets did not need elaborate techniques. Their use
of a simple A versus C plot gave good answers (Table 3).
Students A and D had data sets that were definitely curved.
They chose correctly to chop off the high A data (above 0.10
M) and thus got reasonable results (Table 4, cropped). We then
took the four data sets without truncation to test the techniques
of nonlinear, weighted linear, and quadratic fit. We wished to
see how effectively each of these could compensate for the
curvature in two of the data sets. Table 4 shows the percent by
which the student would miss the unknown concentration
(0.033 M) using all the methods.
Several things are clear from the results of the students data
sets. If students stayed in a small range within the Twyman
Lothian limits they had no difficulty getting good results. The
inexpensive spectrophotometer was quite sufficient for the
experiment. Using nonlinear data analysis was only slightly
better than linear regression for the short data sets. Because

Beers law is not a straight line. Better fits to data with

serious errors are found by nonlinear regression on T versus C,
rather than by transformation to A followed by simple linear
regression on the A versus C plot (Table 3). If the data are kept
within the TwymanLothian limits of T = 0.1 to 0.75 (A = 1 to
0.12) then the differences between the two methods are small,
even with student data from an inexpensive instrument. If
there is obvious curvature in the high absorbance region of a
linear plot, and if these points cannot be discarded because the
unknown is also in this region, the student would be well
advised to use nonlinear regression of T on C. What seem
gross errors on a linear plot are not always bad points, unless
they are well beyond the TwymanLothian limits. State-of-theart instrumentation will allow experimentation in the higher
absorbance region [7], but with ordinary spectrophotometers
the student is advised to avoid the region beyond A = 1, either
by dilution of the samples or by changing the cell path length
if that is an option. If these options are not available, then
nonlinear regression will minimize the errors in the low T
region.
Acknowledgment. We are grateful to Gary D. White for
helpful suggestions that improved this manuscript.
References and Notes
1.

Zielinski, T. J.; Allendorfer, R. D. J. Chem. Educ. 1997, 74, 1001.

2.

Smith, E. D.; Matthews, D. M. J. Chem. Educ. 1967, 44, 757.

3.

Chong, D. P. J. Chem. Educ. 1994, 71, 489.

4.

Martin, B. R. J. Chem. Educ. 1997, 74, 1238.

5.

Taylor, J. R. An Introduction to Error Analysis; University Science

Books: Mill Valley, CA, 1982; pp 164167.

6.

Twyman, F.; Lothian, G. F. Proc. Phys. Soc. (London) 1933, 45,

643.

7.

Skoog, D. A.; Holler, F. J.; Nieman, T. A. Principles of Instrumental

Analysis, 5th ed.; Saunders: Philadelphia, 1998: pp 307309.

8.

Muyskens, M. A.; Sevy, E. T. J. Chem. Educ. 1997, 74, 1138.

9.

This data was collected at Middle Tennessee State University as part

of a quantitative analysis research project directed by Judith M.
Bonicamp.

10. Harris, D. C. J. Chem. Educ. 1998, 75, 119.

11. NLREG (Phillip Sherrod, 4410 Gerald Place,) is an excellent
nonlinear regression software package by Phillip Sherrod, Nashville,
TN.
12. Miller, J. C.; Miller, J. N. Statistics for Analytical Chemistry, 3rd
ed.; Ellis Horwood Ltd.: West Sussex, England, 1993; pp 109115.
13. Harris, D C. Quantitative Chemical Analysis, 5th ed; W. H. Freeman
& Co., 1995; p 517.

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf

88

Chem. Educator, Vol. 4, No. 3, 1999

14. Skoog, D. A.; Holler, F. J.; Nieman, T. A., Principles of

Instrumental Analysis, 5th ed.; Saunders: Philadelphia, 1998; pp
303312.

Bonicamp et. al.

15. Werner, T. C.; Carroll, M. K. Chem. Educator 1997, 2 (4): S14304171 (97) 04139-9. Avail. URL: http://journals.springerny.com/chedr
16. Miller, J. C.; Miller, J. N., Statistics for Analytical Chemistry, 3rd
ed.; Ellis Horwood Ltd.: West Sussex, England, 1993; pp 132133.

1999 Springer-Verlag New York, Inc., S1430-4171(99)03295-8, 10.1007/s00897990295a, 430081jb.pdf