Gordon Fitting Surge Functions to Data

FITTING SURGE FUNCTIONS

TO DATA
Sheldon P. Gordon

ADDRESS: Department of Mathematics, Farmingdale State University of

New York, Farmingdale NY 11735 USA. [email protected]

ABSTRACT: The problem of fitting a surge function to a set of data such

as that for a drug response curve is considered. A variety of different
techniques are applied, including using some fundamental ideas from
calculus, the use of a CAS package, and the use of Excel’s regression
features for fitting a multivariate linear function to a set of trans-
formed data. The results of the different approaches are contrasted
and discussed.

KEYWORDS: Surge functions, drug response curves, curve fitting, least

squares, multivariate regression analysis, rational functions.

INTRODUCTION
One of the “new” families of functions that are being introduced in the early
years of the mathematics curriculum is the surge function, which is treated
at the calculus level in [2, 3] and at the precalculus level in [1]. The surge
function, whose graph is shown in Figure 1, has the form

f(x) = Axp e−bx , (1)

where p > 0 and b > 0, which is equivalent to f(x) = Axp cx , 0 < c < 1.
Surge functions are used to model a variety of real-world applications, such
as the response to an initial dose of a drug (the level of the medication in
the bloodstream rises relatively rapidly to a peak and thereafter decays as
the drug is washed out of the body by the kidneys) or the body’s response
to an infection. Surge functions are also used to model the results of an
advertising campaign that initially causes a fast increase in sales, but which
then slowly diminish. From a modeling point of view, the initial surge is

81
PRIMUS March 2006 Volume XVI Number 1

accounted for by the power function term xp and the subsequent slow decay
is accounted for by the decaying exponential term e−bx or cx.

Figure 1. The graph of a surge function.

A surge function such as the one pictured in Figure 1 has a maximum

and two points of inflection for x>0. As shown in [4] using standard calculus
techniques, the maximum occurs at the point where
p
x = , b 6= 0. (2)
b
A second horizontal tangent occurs at the origin where f 0 (0) = 0; de-
pending on the value of p, this could be a turning point or an inflection
point.√ The two inflection points seen in the graph in Figure 1 occur at x
p± p
= b . This can be written in a more insightful way as
√
p p
x= ± , (3)
b b
which indicates that the two inflection points are located symmetrically
about the turning point at x = pb ; however, the inflection points do occur
at different heights since the curve is not symmetric about the vertical line
through the turning point.
Another extremely important theme in the modern mathematics curricu-
lum is that of fitting a function to a set of data. All graphing calculators (as
well as spreadsheet packages such as Excel) have the capability of fitting a
linear, exponential, power, logarithmic and polynomial function (up to 4th
degree on a calculator and up to 6th degree on Excel) to a set of data. Many
calculators also have the capability of fitting a logistic function and a sinu-
soidal function to data. However, unless one uses a specialized computer
package such as Mathematica or Maple, there is no readily available tool
for fitting a surge function to a set of data. This issue was addressed in [4],
where several different techniques were discussed, but none was particularly
satisfactory in the sense of yielding good results in an easily accessible way.

82
Gordon Fitting Surge Functions to Data

In this article, we look at this issue again and consider in detail an approach
that has the advantage of giving reasonably accurate results with a readily
available tool, as well as the approach of applying the least squares criterion
directly with the assistance of a CAS package.

Figure 2. The drug response curve for Viagra.

To illustrate the different approaches, we will use data on the mean

plasma concentration level for sildenafil citrate (Viagra), a drug that our
students would certainly be aware of and which would obviously pique their
interest. We show the drug response curve for Viagra, as posted at the
Pfizer website [5], in Figure 2. In this situation, the independent variable t
is time in hours since a dose of Viagra was taken initially and the dependent
variable, the mean plasma concentration level C for a group of healthy male
volunteers, is measured in nanograms per milliliter (ng/ml). It is evident
that the drug achieves its maximum level slightly more than an hour after
it is taken and thereafter the level decays relatively slowly.
From this graph, we can estimate a set of data points to use as our
target in finding an equation for the surge function that matches the curve
shown in the Pfizer graph. In particular, we estimate the following values
(see Table 1) for t in hours and the corresponding concentration level C in
nanograms per milliliter:

t 0.05 0.4 0.6 1.2 1.8 2.1 3 4 6 8 10 12 18 24

C 1 50 320 440 410 350 250 170 80 50 30 20 12 6
Table 1. Data on level of Viagra.

83
PRIMUS March 2006 Volume XVI Number 1

It is interesting to note that the first point shown in Pfizer’s graph in

Figure 2 is not at the origin although we would presume that the initial
Viagra concentration level would be 0 at time 0. We will address this issue
later in the article.
From the data, we conclude that the maximum concentration level of
about 440 ng/ml occurs at about t = 1.2 hours. Furthermore, the two inflec-
tion points, which correspond to the points where the function is changing
most rapidly, occur at about t = 0.4 and t = 2.4 hours after the Viagra is
first taken. Since we know that the inflection points for a surge function
should be symmetrically located about the turning point, we might opt to
average the two deviations about t = 1.2 and so estimate that the inflection
points are 1 hour above and 1 hour below t = 1.2; that is, at t = 0.2 and
t = 2.2.
Alternatively, we might reason that just because the largest value of
C in the table is at t = 1.2 hours does not necessarily mean that this is
the absolute maximum value – the Viagra level might reach a higher level
somewhat beyond t = 1.2, say at t = 1.3 or t = 1.4, and the latter value
might make for a more symmetric format. We leave this possibility for the
interested reader to pursue.
We now substitute the estimates for the location of the turning point
(t = 1.2) and the inflection√points (t = 0.2 and 2.2) into Equations √(2) and
p+ p p
(3) to get pb = 1.2 and b = 2.2 and the latter is equivalent to b = 1.
These two equations can be solved readily to get p = 1.44 and b = 1.2, so
that the form of our surge function is

C(t) = At1.44e−1.2t.

We know that the peak concentration value of approximately 440 ng/ml

occurs at about t = 1.2, so that C(1.2) = A(1.2)1.44e−1.44 = 0.30806A =
440, and so A = 1428.29. The corresponding model for the surge function
is therefore

C(t) = 1428.29t1.44e−1.2t. (4)

We show the graph of this function superimposed over the data points in
Figure 3 and conclude that it is a reasonably good fit for t between 0 and
about 3 hours, although thereafter the surge function dies out much more
rapidly than the concentration of Viagra does.
Probably the most common measure used to assess how well a function
fits a set of data is the sum of the squares of the vertical deviations between
the curve and the data points. The corresponding value for the surge func-

84
Gordon Fitting Surge Functions to Data

tion (4) shown in Figure 3 is 53,459.9. We will use this value for comparison
in our subsequent calculations.

Figure 3. Viagra data and the surge function.

Figure 4. Another surge function with the Viagra data.

We can obtain a better approximation to a best-fit surge function by

using some of the features in Mathematica to minimize the value of the
sum of the squares. (The actual Mathematica session, including the specific
commands used and the corresponding output, is shown in Appendix A.)

85
PRIMUS March 2006 Volume XVI Number 1

The result of this process is the quite different surge function

C(t) = 1100.31t1.6634e−1.08376t. (5)

All three of the parameters have changed considerably and should change
the function significantly compared to the surge function in Equation (4).
The corresponding value for the sum of the squares associated with this
function is 25,815.1, so that it is significantly smaller (about 12 as large)
than the value of 53,459.9 that we got by simply applying the calculus-
based results. We show this function superimposed over the data points in
Figure 4, and observe that it seems to be a better fit to more of the points
than the surge function (4) in Figure 3. Nevertheless, this function is still a
rather poor fit to the data, especially after about t = 5 hours. And, perhaps
more importantly, because this approach requires the use of a specialized
CAS program that might not be available to all students, it may not be the
most effective approach with which to investigate other sets of data that
fall into the pattern of a surge function.

USING MULTIVARIATE REGRESSION

We now consider a different way to find the equation of a surge function
that fits a set of data. To do so involves using some ideas on fitting a
linear function of two or more variables to a set of multivariate data. In
particular, suppose we have a set of (x1 , x2, . . . , xn , y) data, where y
depends on the n independent variables x1 , x2 , . . . , xn . Multivariate linear
regression is a standard tool that is used to find the linear function Y =
c0 + c1 X1 + c2X2 + . . . + cn Xn that is the best fit to the data. Think of
this as finding the hyperplane in n + 1 dimensional space that comes closest
to all of the points in the data set. This capability is available in many
software packages, including Excel, so it is readily at hand for use. (Note
that this feature is not automatically loaded when Excel is first installed;
rather, it must be loaded one time as an Add-In under Excel’s Tools menu
– just select Analysis ToolPak, and it is thereafter available under Tools.
One of the options you then get is Regression. We describe the use of this
feature in Appendix B.)
One somewhat unexpected application of multivariate linear regression
is in fitting a polynomial in one variable x to a set of (x, y) data. Suppose
we wish to find a polynomial of degree 3, y = a0 + a1x + a2x2 + a3 x3,
that fits such a table of values. We can think of the polynomial expression
as a linear function of x, x2 , and x3 and then apply multivariate linear
regression, where X1 = x, X2 = x2, and X3 = x3. The resulting coefficients

86
Gordon Fitting Surge Functions to Data

for the constant term and for the three variables X1 , X2 , and X3 are then
the coefficients of the desired polynomial.
We now apply a similar approach to fitting a surge function to a set of
(x, y) data, and will then apply the procedure to the data on the concen-
tration levels of Viagra. Since the surge function we seek has the form

y = Axp e−bx ,

we can take logarithms of both sides to get

log(y) = log(A) + log(xp ) + log(e−bx ) = log(A) + p log(x) − bx

and so log y is a linear function of x and log x. Thus, we can set Y = log y,
X1 = x and X2 = log x and apply multivariate linear regression to an
extended table of values that also includes a column of log x values and a
column of log y values. In order to take logs of the t and the C values,
we need to avoid the obvious starting point where t = 0 and C = 0; we
do this by making a very minor change in the values of the two variables
at that point and use t = 0.05 instead of 0 and C = 10 instead of 0. We
presume that the researchers at Pfizer did the same in producing the graph
in Figure 2 on their website; otherwise, it is far more natural to use (0, 0) as
the starting point. For the Viagra data, we then have the extended Table
2.

C t log t log C
10 0.05 -1.30103 1
50 0.4 -0.39794 1.69897
320 0.6 -0.22185 2.50515
440 1.2 0.079181 2.643453
410 1.8 0.255273 2.612784
350 2.1 0.322219 2.544068
250 3 0.477121 2.39794
170 4 0.60206 2.230449
80 6 0.778151 1.90309
50 8 0.90309 1.69897
30 10 1 1.477121
20 12 1.079181 1.30103
12 18 1.255273 1.079181
6 24 1.380211 0.778151

Table 2. Original and transformed data.

87
PRIMUS March 2006 Volume XVI Number 1

When we “hit” this set of transformed data with the multivariate regression
features of Excel, we get the linear regression equation

Y = 2.3190 − 0.1242X1 + 0.7613X2,

which is equivalent to

log C = 2.3190 − 0.1242t + 0.7613 logt.

We can eliminate the logs algebraically by undoing the original transforma-
tion using powers of 10 and so obtain

10log C = C = 102.3190−0.1242t+0.7613log t
= 102.319010−0.1242t100.7613 log t
0.7613
= 208.45(10−0.1242)t10log t
= 208.45(0.7513)tt0.7613.

Figure 5. The surge function based on multivariate regression.

The base 0.7513 in the exponential term can be converted to an ap-

propriate power of e by solving the exponential equation e−b = −0.7513,
which leads to b = 0.2859. Thus, we have the surge function C(t) =
208.45t0.7613e−0.2859t, which is shown superimposed over the data points
in Figure 5. It is obviously a very poor fit to the data, other than at the
very beginning and the very end. Furthermore, the associated value for the
sum of the squares is 192,812.4, which is considerably larger than anything
we had before and which therefore corroborates our visual conclusion that

88
Gordon Fitting Surge Functions to Data

the fit is extremely poor. Yet, the logic leading up to this result seems rea-
sonable in the sense that the multivariate regression process produces the
best-fit plane to the transformed data and therefore leads us to expect a
much better fit. Let’s see what has gone wrong.
In Figure 6, we show the plot of the points (log t, log C). Notice that,
other than the left-most point (−1.5, 1), the remaining points are clustered
relatively tightly and mostly display a clear pattern. This suggests that the
results we get for the regression equation might be very sensitive to small
changes in the values of the coordinates at the left-most point in the sense
that this point may have a disproportionate effect on the coefficients in the
regression equation.
Moreover, the left-most point is what we estimated to avoid the problem
with taking logs of 0. And, because it involves a negative value for t, a
relatively minor change in the value of t near 0 would likely result in a
major change in the value of log t. In addition, when you look back at
Figure 2 (the Pfizer website graph), it is evident that this initial point is
the one for which it is hardest to estimate an accurate value.

Figure 6. Plot of log C vs. log t.

Let’s see just how much of an effect we get by trying a slightly different
estimate for the value of t for this point. Instead of using t = 0.05, suppose
we try t = 0.10 and maintain the value C = 10. The resulting surge
function is C(t) = 619.44t0.8236e−0.2924t. The value for the coefficient has
changed dramatically from A = 208.45; the power in the power function
term has changed a bit from p = 0.7613 to p = 0.8236; and the multiple
in the exponential term has changed fairly minimally from b = −0.2859 to
b = −0.2924. However, the corresponding value of the sum of the squares
is now 700,741.1, which is almost four times as large as the previous value
of 192,812.4 and the resulting surge function is a far poorer fit to this data.
More significantly, a relatively small change in the estimate of the point

89
PRIMUS March 2006 Volume XVI Number 1

near the origin clearly results in a huge change in the results.

One way to circumvent this issue is to realize that, by its very nature, ev-
ery surge function must pass through the origin provided that the power p in
the power function term is positive. As a consequence, it might make sense
to ignore the point near the origin altogether and see what happens if we use
only the remaining points for the multivariate regression analysis. When we
do this, the corresponding linear function is Y = 2.4276−0.092X1+0.225X2,
which is equivalent to log C = 2.4276 − 0.092t + 0.225 log t.
When we undo the logarithmic transformation by taking powers of 10,
we eventually get the surge function

C(t) = 267.67(0.8091)t t0.225 = 267.67t0.225e−0.2118t.

The associated value for the sum of the squares is 153,650.7. This is a
substantial improvement over the two preceding surge functions using mul-
tivariate regression with estimates of the point near the origin. However, it
is still considerably larger than the value of 53,459.9 we initially obtained
using the calculus argument, let alone the value of 25,815.1 that resulted
from the Mathematica routine for minimizing the sum of the squares.
Incidentally, there is another statistical measure used to assess how well
a multivariate linear function fits a set of data; it is the coefficient of multiple
determination and is denoted by R. It is the extension of the correlation
coefficient r to multivariate data. For the three functions we have created
using multivariate regression, the corresponding values are R = 0.8768,
R = 0.8581, and R = 0.8988, respectively. While all three values are
statistically significant, the fact that they are all quite close to one another
means that we should not make a definitive call on which of the three surge
functions is the best fit based solely on the value of R.
This still leaves one rather perplexing question. How can all three of
these surge functions based on multivariate linear regression be so much
poorer fits than the one based on calculus, let alone the one obtained us-
ing the computer search method? After all, multivariate linear regression
is supposed to produce the best fit! The key is that it does produce the
best linear fit to the set of transformed (t, log t, log C) data. If all we did
was to stop there, we would indeed have the best possible fit. However,
we started with (t, C) data and, in the process of transforming it via log-
arithms, we stretched the data values in a non-linear way. After we got
the corresponding multivariable linear regression equation, we undid the
original transformation, which entails another non-linear stretch, but this
time the inverse transformation is applied to the function, not to the data.
So, although the three regression planes we obtained were the best linear

90
Gordon Fitting Surge Functions to Data

fits to the three different sets of transformed data, the corresponding surge
functions are not necessarily the best, or even extremely good, fits to the
original data. They may be good fits if the original data fall very closely
into a surge function pattern; however, if the data is not extremely close to
such a pattern, the resulting function based on multivariate regression may
be a surprisingly poor fit.
A comparable situation arises with the curve fitting routines in calcula-
tors and in Excel; rather than directly fitting an exponential, logarithmic,
or power functions to a set of data, these routines transform the data (either
a semi-log plot or a log-log plot), find the regression line for the transformed
data, and then undo the transformation algebraically. In the process, one
obtains the best possible line for the transformed data, but in the process of
undoing the transformation, a nonlinear stretch takes place and the result-
ing function is not necessarily the best fit within that family of functions.

FITTING A RATIONAL FUNCTION TO THE DATA

Gordon and Gordon [4] also discuss the possibility of fitting a rational func-
tion of the form
at2
C(t) = 4 , (6)
t + b2
to a set of data on drug concentration levels over time, where a and b are
two constants. The quadratic term in the numerator is needed to reflect
the curvature of the drug data at and near the origin; the quartic term in
the denominator reflects the fact that the data eventually die out as time
progresses.
In [4], it was found that such a rational function was actually a con-
siderably better fit than a surge function is to a set of data on the drug
concentration level for a form of L-Dopa used to treat patients with Parkin-
son’s disease. Let’s see how well such a function fits the data on Viagra. We
again use Mathematica to perform a direct least squares fit with a rational
function of the form in Equation (6). The resulting function is
1822.6t2
C(t) = .
t4 + 3.191292
where the corresponding value for the sum of the squares is 28,468.6. We
note that this value is slightly larger than the value of 25,815.1 we obtained
before for the best fitting surge function. So, in this case, the rational
function gives slightly poorer accuracy. In Figure 7, we show both the best
surge function and this best rational function of the form in Equation (5)
to compare the relative fits. From this, we see that the rational function

91
PRIMUS March 2006 Volume XVI Number 1

(the darker curve) spikes to a considerably higher level than either the data
or the surge function do; however, it dies out more slowly than the surge
function and so is a better fit to the data after about t = 12 hours.

Figure 7. The rational function (darker curve) vs. the surge function.

CONCLUSIONS

In general, the focus of this kind of investigation should not be on simply

finding a function that is the best possible fit to a set of data, but rather
on finding a function that is a reasonable fit and whose properties provide
insight into the situation being modeled. A surge function provides that
kind of insight in the sense that the power function term models the initial
impetus and the exponential term models the eventual exponential-type
decay of the drug concentration levels. The rational function certainly fits
the data well, but there are considerably less compelling interpretations for
why it follows the desired behavior pattern – the term at2 does model the
initial impetus and the t41+b term does die out relatively quickly, but the
latter is nowhere as convincing a description as exponential decay.
So the key is the realization that all we are producing is a mathematical
model. There are different routes to developing such a model, not only the
ones we discussed here. For instance, there may be a more sophisticated
pharmacokinetics model based on differential equations, but that is beyond
the scope of what we are considering here and likely beyond the scope
of the students we serve in introductory courses. In the final analysis,
what is most important is not the keystrokes used to produce a model,
but an understanding of the mathematics underlying that model and the

92
Gordon Fitting Surge Functions to Data

development of the judgment necessary to decide how well the function

actually fits the data and how the model gives us an understanding of the
process.
ACKNOWLEGEMENT
The work described in this article was supported by the Division of Un-
dergraduate Education of the National Science Foundation under grants
DUE-0089400, DUE-0310123, and DUE-0442160. The author also appre-
ciates the assistance provided by Brian Winkel, editor of PRIMUS , with
Mathematica.
REFERENCES

1. Gordon, Sheldon P., Florence S. Gordon, et al. 2004. Functioning

in the Real World: A Precalculus Experience, 2nd Ed. Boston: Addison-
Wesley.
2. Hughes-Hallett, Deborah, Andrew Gleason, et al. 2002. Calculus, 3rd
Ed. New York: John Wiley & Sons, New York.
3. Hughes-Hallett, Deborah, Andrew Gleason, et al. 1999. Applied Cal-
culus, New York: John Wiley & Sons.
4. Gordon, Sheldon P. and Florence S. Gordon. 2003. A Spoonful of
Medicine Makes the Mathematics Go Down. The AMATYC Review. 24:
9-24.
5. http://www.pfizer.com/download/uspi viagra.pdf

APPENDIX A:
PERFORMING DIRECT CURVE FITTING
IN MATHEMATICA

Direct least squares fitting of surge function to data.

data = {{.05, 1}, {.4, 50}, {.6, 320}, {1.2, 440}, {1.8, 410},
{2.1, 350}, {3, 250}, {4, 170}, {6, 80}, {10, 30}, {12, 20},
{18, 12}, {24, 6}}
m[x , a , p , b ] = a x^p Exp [-b x]
ss[a , p , b ] = Sum [(data[[i, 2]] - m[data[[i, 1]], a, p, b])^2,
{i,1,Length[data]} ]
sm = FindMinimum[ss[a, p, b], {a, 1}, {b, 1}, {p, 2}]
{23851.6, {a → 1133.57, b → 1.10895. p → 1.69499}}
mf[x ] = m[x,a,p,b]/.FindMinimum[ss[a,p,b],{a,1},{b, 1},{p,1}][[2]]
1133.57 e−1.10895x x1.69499

93
PRIMUS March 2006 Volume XVI Number 1

Figure 8. Plot, in Mathematica, of data with fitted curve.

APPENDIX B:
PERFORMING MULTIVARIATE
REGRESSION IN EXCEL
With Excel’s Analysis ToolPak installed, enter the data values for the de-
pendent variable C in Column A, say, and those for the dependent variable
t in Column B and then create lists of values for log t in Column C and
log C in Column D, as is done in Table 2. Then click on Tools, followed
by Data Analysis, and finally Regression and OK. This will bring up the
Excel dialog box shown in Figure 9. In this dialog, the first box asks you
to Input Y Range; the Y-values are the values for the desired dependent
variable, here log C, which are in Column D. The second box asks you to
Input X Range; the X-values for t and log t are in Columns B and C. Next,
select the first option, Output Range, under Output options; this will give
the cells in which all the regression analysis output will appear. Select a
collection of cells that are empty, say from A20 to I47.
When you click on OK, Excel will perform the complete regression anal-
ysis and print the results in the cells that were indicated. Sample results
are shown in Figure 10. Of all the output results, the only ones that are of
significance to this discussion are the values for the regression coefficients
in Rows 36-38 and possibly the value for the coefficient of multiple deter-
mination R in Row 23. In particular, the constant coefficient is 2.428, the
coefficient of the first independent variable t is -0.0920 and the coefficient
of the second independent variable log t is 0.2251, leading to the regres-
sion equation Y = 2.428 − −0.0920X1 + 0.2251X2, which is equivalent to
log C = 2.428 − 0.0920t + 0.2251 log t.

94
Gordon Fitting Surge Functions to Data

Figure 9. Excel dialog box for regression analysis.

Figure 10. Excel’s regression output display.

95
PRIMUS March 2006 Volume XVI Number 1

BIOGRAPHICAL SKETCH
Sheldon Gordon is Professor of Mathematics at Farmingdale State Univer-
sity of New York. He is a member of a number of national committees
involved in undergraduate mathematics education and is leading a national
initiative to refocus the courses below calculus. He is the principal author
of Functioning in the Real World and a co-author of the texts developed
under the Harvard Calculus Consortium.

96