United Arab Emirates University

College of Sciences
Department of Mathematical Sciences

HOMEWORK 1 – SOLUTION

Sec 1.1 – Sec 1.3

Mathematical Modeling
MATH 470 SECTION 51 CRN 31749
8:00 – 9:15 on Sunday, Monday, Tuesday, & Wednesday
Due Date: Wednesday, February 17, 2010

ID No: Solution

Name: Solution

Score: Solution
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

Section 1.1 Modeling Change with Difference Equations

1. Find a formula for the nth term of the sequence.
(1.1) {1, 4, 16, 64, 256, . . .}
Answer. Look for the pattern:

a1 = 1 = 20 = (20 )2 a2 = 4 = 22 = (21 )2 a3 = 16 = 42 = (22 )2

a4 = 64 = 82 = (23 )2 a5 = 256 = 162 = (24 )2 an = (2n )2 = 22n . 

(1.2) {1, 3, 7, 15, 31, . . .}
Answer. Look for the pattern:

△a1 = a2 − a1 = 3 − 1 = 2 = 21 △a2 = a3 − a2 = 7 − 3 = 4 = 22

△a3 = a4 − a3 = 15 − 7 = 8 = 23 △a4 = a5 − a4 = 31 − 15 = 16 = 24

△an = an+1 − an = 2n =⇒ an+1 = 2n an . 

• Remark: If we add all the equations up to △an−1 above, by the telescoping cancelation,

∑
n−1 ∑
n−1
△ai = an − a1 = 2 + 2 + 2 + · · · + 2
1 2 3 n−1
= 2i = 2n − 2
i=1 i=1
=⇒ an = 2 − 2 + a1 = 2 − 2 + 1 = 2 − 1
n n n
=⇒ an = 2n − 1.

2. By examining the following sequences, write a difference equation to represent the change during
the nth interval as a function of the previous term in the sequence.
(2.1) {1, 2, 5, 11, 23, . . .}
Answer. Ignore! 
(2.2) {1, 8, 29, 92, . . .}
Answer. Look for the pattern:

△a1 = a2 − a1 = 8 − 1 = 7, △a2 = a3 − a2 = 29 − 8 = 21 = 3 · 7 = 3△a1

△a3 = a4 − a3 = 92 − 29 = 63 = 3 · 21 = 3△a2
∑
n ∑
n
△an = 3△an−1 =⇒ an+1 − a2 = △ai = 3 △ai−1 = 3 (an − a1 )
i=2 i=2
3n 7 − 15
an+1 = 3an − 3a1 + a2 = 3an + 5
formula on linear system
=⇒ an =
6
. 

Page 1 of 6
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

3. Formulate a dynamical system that models change exactly for the described situation.
(3.1) You currently have $5000 in a savings account that pays 0.5% interest each month. You add
another $200 each month.
Answer. Let an be the amount of money in the account after n months. Then we have

△an = 0.005an + 200 =⇒ an+1 = 1.005an + 200 with a0 = 5000.

The formula on the linear dynamical system implies

( )
8
an = 1.005 45000 − 40000 = 45000 1.005 −
n n
9
. 

(3.2) Your parents are considering a 30–year, $100, 000 mortgage that charges 0.5% interest each
month. Formulate a model in terms of a monthly payment p that allows the mortgage (loan) to
be paid off after 360 payments. (Hint: If an represents the amount owed after n months, what
are a0 and a360 ?)
Answer. Let an be the amount owed after n months. Then we have

△an = 0.005an − p =⇒ an+1 = 1.005an − p with a0 = 1000000.

The formula on the linear dynamical system implies

an = 1.005n (1000000 − 200p) + 200p = 200 [1.005n (5000 − p) + p]

[ ]
a0 = 200 1.0050 (5000 − p) + p = 1000000
[ ]
a360 = 200 1.005360 (5000 − p) + p = 200(6.02258(5000 − p) + p)
≈ 6022575 − 1004.52p

Since a360 ≈ 6022575 − 1004.52p = 0 at p ≈ 5995.51, so they need to pay about $6000 each
month for the complete pay off after 360 payments, i.e., 30 years. 
• Remark: If we increase the interest rate from 0.5% to 1%, then how much money should they
pay for the complete pay off after 360 payments, i.e., 30 years?

Page 2 of 6
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

Section 1.2 Approximating Change with Difference Equations

4. The following data were obtained for the growth of a sheep population introduced into a new envi-
ronment on the island of Tasmania.

Year 1814 1824 1834 1844 1854 1864

Population 125 275 830 1200 1750 1650

Plot the data. Is there a trend? Plot the change in population versus years elapsed after 1814. For-
mulate a discrete dynamical system that reasonably approximates the change you have observed.

Answer. Ignore! 
5. Sociologists recognize a phenomenon called social diffusion, which is the spreading of a piece of in-
formation, a technological innovation, or a cultural fad among a population. The members of the
population can be divided into two classes: those who have the information and those who do not.
In a fixed population whose size is known, it is reasonable to assume that the rate of diffusion is pro-
portional to the number who have the information times the number yet to receive it. If an denotes
the number of people who have the information in a population N people after n days, formulate a
dynamical system to approximate the change in the number of people in the population who have
the information.

Answer. It is clear to see

rate of diffusion = kan (N − an ) .
an+1 − an
Here k is the proportionality constant. In the discrete case like this problem, the rate means = △a
n+1−n
(In the continuous case, it means the derivative.) So, the equation is in fact

△an = kan (N − an ) . 

Page 3 of 6
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

Section 1.3 Solutions to Dynamical Systems

6. Find the solution to the difference equations in the following problems.
(6.1) an+1 = 3an /4, a0 = 64
Answer. The formula on the solution of the linear dynamical system implies
( )n ( )n
3 3
an =
4
a0 =
4
64. 

(6.2) an+1 = 0.1an + 3.2, a0 = 1.3

Answer. The formula on the solution of the linear dynamical system implies
203 32
an = −
90
0.1n + .
9


7. For the following problems, find an equilibrium value if one exists. Classify the equilibrium value as
stable or unstable.
(7.1) an+1 = 0.9an
Answer. Suppose the system has the equilibrium value E. Then

E = 0.9E =⇒ E = 0.

In the given system, |r = 0.9| < 1 and so the equilibrium value is stable. 

(7.2) an+1 = −0.8an + 100

Answer. Suppose the system has the equilibrium value E. Then
500
E = −0.8E + 100 =⇒ E= .
9

In the given system, |r = −0.8| < 1 and so the equilibrium value is stable. 

(7.3) an+1 = an − 100

Answer. In the given system, r = 1 and so the equilibrium value does not exist. 

Page 4 of 6
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

8. Build a numerical solution for the initial value problem, an+1 = 0.8an − 100, a0 = 500. Plot your data
to observe pattern in the solution. Is there an equilibrium value? Is it stable or unstable?

Figure 1: Continuous graph of solution an

Answer. Let E be the equilibrium value of the system. Then, we have

100
E = 0.8E − 100 =⇒ E=− = −500.
0.2
By the formula, the dynamical system has the solution
an = 0.8n (a0 + 500) − 500 = 0.8n 1000 − 500.
As n → ∞, 0.8n → 0 and so an → −500 = E. Thus, the equilibrium value is stable. See the
figure 1. 
9. Consider an+1 = an + 2, a0 = −1. Find the solution to the difference equation and the equilibrium
value if one exists. Discuss the long–term behavior of the solution for the various initial values.
Classify the equilibrium values as stable or unstable.

Answer. Suppose E is the equilibrium value of the system. Then,

E = E + 2,
which cannot be true for any E. Thus, there is no equilibrium value.
Now we find the solution of the dynamical system as follows.
∑
n−1 ∑
n−1
△an = an+1 − an = 2 =⇒ △an = 2 =⇒ an − a0 = △ai = 2 = 2n
i=0 i=0

=⇒ an = 2n + a0 = 2n − 1. 

Page 5 of 6
Mathematical Modeling HOMEWORK 1 – SOLUTION Spring, 2010

10. You currently have $5000 in a savings account that pays 0.5% interest each month. You add $200
each month. Build a numerical solution to determine when the account reaches $20, 000.

Figure 2: Continuous graph of solution an

Answer. Let an be the amount owed after n months. Then we have

△an = 0.005an + 200 =⇒ an+1 = 1.005an + 200 with a0 = 5000.

The formula on the linear dynamical system implies

( )
8
an = 1.005 45000 − 40000 = 45000 1.005 −
n n
.
9

Now we solve an = 20000 for n:

( )
8 4 8 ln(4/3)
20000 = an = 45000 1.005 − n
=⇒ = 1.005n − =⇒ n= ≈ 57.6801.
9 9 9 ln 1.005

Thus, after 58 months, the account will reach $20000. 

Page 6 of 6
 United Arab Emirates University
College of Sciences
Department of Mathematical Sciences

HOMEWORK 2 – SOLUTION

Sec 2.2 – Sec 3.4

Mathematical Modeling
MATH 470 SECTION 51 CRN 31749
8:00 – 9:15 on Sunday, Monday, Tuesday, & Wednesday
Due Date: Monday, March 15, 2010
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 2.2 Modeling Using Proportionality

1. If a spring is stretched 0.37 in. by a 14–lb force, what stretch will be produced by a 9–lb force? By a 22–lb force?
Assume Hooke’s law, which asserts the distance stretched is proportional to the force applied.

Answer. Hooke’s Law says F = kS, where F is the restoring force in a spring stretched or compressed a distance
S. The given condition implies

4 lb
4 lb = k(0.37 in) =⇒ k= = 10.8108 lb/in.
0.37 in
Using this proportional constant, we get
9
9 lb = (10.8108 lb/in)S, i.e., S= in = 0.8325 in.
10.8108
22
22 lb = (10.8108 lb/in)S, i.e., S=
10.8108
in = 2.035 in. 

2. If an architectural drawing is scaled so that 0.75 in. represents 4 ft, what length represents 27 ft?
Answer. Let x be the length representing 27 ft. Then we observe a simple ratio:

4 ft 27 ft 0.75 in
0.75 in
=
x in
, i.e., x in = 27 ft
4 ft
= 5.0625 in. 

3. Determine whether the following data support a proportionality argument for y ∝ z 1/2 . If so, estimate the
slope.
y 3.5 5 6 7 8
z 3 6 9 12 15
Answer 1. Transformation via Square. First we have to determine whether or not y and z 1/2 are proportional,
i.e., whether or not there is a positive constant k satisfying y = kz 1/2 . If they are not, we don’t have to proceed.
For this purpose, we compute the ratio y 2 /z, because y 2 and z are proportional if and only if y and z 1/2 are
proportional. (One can compute the ratio y/z 1/2 or z 1/2 /y or z/y 2 or (y/z 1/2 )p , where p is any real number
except 0.)
From the following table, we are allowed to say that the given data can be approximated by

(y 2 , z) (12.25, 3) (25, 6) (36, 9) (49, 12) (64, 15)

y 2 /z 4.0833 4.1667 4 4.0833 4.2667
y2 √
= 4.12, i.e., y 2 = 4.12z, i.e., y= 4.12z 1/2 = 2.0298z 1/2 ,
z

where 4.12 is the average of the ratio yi2 /zi . The figure 1 below supports our result. 

Figure 1: Data and Fitting Model y = 2.0298z 1/2

Page 1 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

(y, z) (3.5, 3) (5, 6) (6, 9) (7, 12) (8, 15)

ln y − (ln z)/2 0.7035 0.7136 0.6931 0.7035 0.7254

Answer 2. Transformation via Natural Logarithmic Function. We introduce another way to determine whether
or not y and z 1/2 are proportional and to find the constant of proportionality. A simple observation shows that
y and z 1/2 are proportional if and only if the difference between ln y and ln z 1/2 = (ln z)/2 is constant.
From the following table,

the third row shows that ln y − (ln z)/2 is close to a constant 0.7078, which is the average of them. Hence, we
can deduce
1
ln y = ln z + 0.7078 =⇒ y = e0.7078 z 1/2 = 2.0295z 1/2 .
2
The figure 2 below supports our result. 

Figure 2: Fitting Model y = 2.02954z 1/2

4. A new planet is discovered beyond Pluto at a mean distance to the sun of 4004 million miles. Using Kepler’s
third law, determine an estimate for the time T to travel around the sun in an orbit.

Answer. Kepler’s Third Law says T = cR3/2 , where T is the period (days) and R is the mean distance to the
sun. From the result of the Example 1 on page 67 of the textbook, we take c = 0.410. Then we have T =
0.410(40043/2 ) = 103878. 

Page 2 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 2.3 Modeling Using Geometric Similarity

5. The stilt, a little long–legged bird, was described in Gulliver’s Travels as weighing 4.5 ounces and having legs that
are 8 in long. A flamingo has a similar shape and weighs 4 lb. Apply scaling arguments to show that flamingo’s
legs should be about 20 in long (as they actually are!).

Answer. Since a stilt and a flamingo have the similar shape, we can assume the geometric similarity between
two species and moreover, we assume the weight is proportional to the cube of the leg length,

W ∝ l3 W = kl3 ,

where l is the length of the leg and W is the weight and k is the constant of the proportionality. Then, the given
information implies
W 4.5 oz 3
k = 3 = 3 3 = 0.00879 oz/in .
l 8 in
3
Using the equation W = (0.00879 oz/in ) l3 , we deduce
3
4 lb = (0.00879 oz/in ) l3
( )1/3 ( )1/3
4 lb 64 oz
=⇒ l= 3 = 3 = 19.3826 in,
0.00879 oz/in 0.00879 oz/in

which is almost 20 in and here, 1 lb = 16 oz is used. 

Page 3 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 3.1 Fitting Models to Data Graphically

6. In the following data, V represents a mean walking velocity and P represents the population size. We wish to
know if we can predict the population size P by observing how fast people walk.
(1) Plot the data.
(2) What kind of a relationship is suggested?
(3) Test the given models by plotting the appropriate transformed data.

V 2.27 2.76 3.27 3.31 3.70 3.85 4.31 4.39 4.42 4.81 4.90
P 2500 365 23700 5491 14000 78200 70700 138000 304500 341948 49375
V 5.05 5.21 5.62 5.88
P 260200 867023 1340000 1092759

(6.1) P = aV b .
Answer. (1) See the figure 3.
(2) Based on the plot of data, we may expect the exponential relationship such as P = aV b .
(3) It is not easy to determine the parameters a and b such that the graph of P = aV b is close to the data. So
we take the natural logarithmic function on the suggested model P = aV b , ln P = ln a + b ln V , which is
the equation of a line on the (ln V )(ln P )–plane with the slope b and (ln P )–intercept (0, ln a). To use this
model, we need to transform the data (ln Vi , ln Pi ). We plot the transformed data and find a straight line
which is close to data as many as possible. It seems that the slope b = 9.55 and the intercept ln a = −2.88.
That is, we deduce

ln P = −2.88 + 9.55 ln V =⇒ P = e−2.88 V 9.55 = 0.05613V 9.55 . (6.1)

One may have other a and b as long as those values of the parameters are reasonable. 

Figure 3: Fitting Model P = 0.05613V 9.55

Page 4 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

(6.2) P = a ln V .
Answer. The plot of transformed data (ln V, P ) is almost similar to the plot of (V, P ), which looks like
the exponential graph. For this reason, the suggested model P = a ln V cannot be used to predict the
population size P by the walking speed V . See the figure. See the figure 4. 
(6.3) P = a(bV ).
Answer. (1) See the figure 5.
(2) Based on the plot of data, we may expect the exponential relationship such as P = abV .
(3) Since the previous suggested model is not good, let us consider another exponential type of model.
Transformation:
ln P = V ln b + ln a,
which is a line on the V versus ln P plane. It has the slope ln b and the vertical intercept (0, ln a). Using
two points of (V, ln P ), we find the slope and the intercept. If we choose (V, ln P ) = (4.31, ln 70700) and
(5.88, ln 1092759), we have
ln 1092759 − ln 70700
ln P = (V − 4.31) + ln 70700 = 1.7440V + 3.6497
5.88 − 4.31
=⇒ P = 38.4646(5.7199V ). (6.2)

Figure 4: Fitting Model P = a ln V

Figure 5: Fitting Model P = 38.4646(5.7199V )

Page 5 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

7. The following data represent the growth of a population (P ) of fruit flies over a 6–week period.
(1) Test the following models by plotting an appropriate set of data.
(2) Estimate the parameters of the given model.

t (days) 7 14 21 28 35 42
P (population) 8 41 133 250 280 297
(7.1) P = ct
Answer. (1) It does not seem that the plot of the data have the linearity. So the model P = ct is not
appropriate for the given data.
(2) If we insist the linear model P = ct, then by choosing one of the data points, we can estimate c. When
we choose the third (t, P ) = (21, 133), it seems that the line P = 6.33t passing through the origin and
looks close to all data. See the figure 6. 

Figure 6: Model P = 6.33t

Figure 7: Model P = 11.3903e0.0915t

(7.2) P = aebt
Answer. (1) The plot of the data looks like the exponential graph and so the suggested model is more
appropriate. See the figure 7.
(2) We transform the data and also the model by taking the natural logarithmic function:
ln P = ln a + bt,
which is a line on the t versus ln P plane. The line has the slope b and the (ln P )–intercept (0, ln a). We plot
the data (t, ln P ) and look for the line close to all the data. Using the second and the four data, (t, ln P ) =
(14, ln 41) and (t, ln P ) = (45, ln 280),
ln P = 0.0915t + 2.4328 =⇒ P = 11.3903e0.0915t . (7.1)

Page 6 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010


• Aside: As we studied in Section 1.2 Approximating Change with Difference Equations, specifically, Example 2
Growth of a Yeast Culture Revisited on page 10 of the textbook, the population p(t) follows the differential equation,

dp(t)
= (µ − bp) p
dt
where µ and b are constants. Solving the differential equation, we have
µ
p(t) = ,
b + ce−µt
where c is another constant. Thus, this model should be better than both of them tested above. However, it is
not easy to determine the constants/parameters µ, b and c. See the figure 8. 

298.671
Figure 8: Model P =
1 + 210.059e−0.251t

Page 7 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

8. The following data represent (hypothetical) energy consumption normalized to the year 1900. Plot the data.
Test the model Q = aebx by plotting the transformed data. Estimate the parameters of the model graphically.

x Year Consumption Q
0 1900 1.00
10 1910 2.01
20 1920 4.06
30 1930 8.17
40 1940 16.44
50 1950 33.12
60 1960 66.89
70 1970 134.29
80 1980 270.43
90 1990 544.57
100 2000 1096.63

Answer. When we transform the data and the model, ln Q = bx + ln a, we observe a straight line passing
through the origin, because (x, Q) = (0, 1) is transformed to (x, ln Q) = (0, 0). Using the origin and the last
data (x, ln Q) = (100, 7), we have the equation of the line,

ln Q = 0.07x =⇒ Q = e0.07x .

See the figure 9. 

Figure 9: Model Q = e0.07x

Page 8 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 3.2 Analytic Methods of Model Fitting

9. 1 Using elementary calculus, show that the minimum and maximum points for y = f (x) occur among the
minimum and maximum points for y = f 2 (x). Assuming f (x) ≥ 0, why can we minimize f (x) by minimizing
f 2 (x)?

Answer. Logic Behind the Strategy: Let A and B be the sets of all extremum points of y = f (x) and y = f 2 (x),
respectively. Then the problem says A is a subset of B. So we choose any element of A and show that it belongs
to B.
Let (x0 , y0 ) be an extremum of y = f (x).
Step 1: From Calculus, x = x0 must be a critical number and so the derivative of y = f (x) at x = x0 should
vanish, i.e.,
f ′ (x0 ) = 0 =⇒ 2f (x0 )f ′ (x0 ) = 0, i.e., (f 2 )′ (x0 ) = 0,
implying that x = x0 is a critical number of y = f 2 (x).
Step 2: We introduce the sign function sgn(x) defined by

 −1 for x < 0
sgn(x) = 0 for x = 0

1 for x > 0.

It is straightforward to understand sgn(ab) = sgn(a) sgn(b) for any real numbers a and b.
We recall the First Derivative Test from Calculus saying that for a critical number x = x0 ,

sgn(f ′ (a)) sgn(f ′ (b)) < 0 at any a and b near x0 with a ≤ x0 ≤ b (9.1)

if and only if the critical number yields the extremum.

Since the critical number x = x0 gives the extremum of y = f (x), there exists a small interval x1 < x0 < x2
such that for any a ∈ (x1 , x0 ) and b ∈ (x0 , x2 ) in the interval (x1 , x2 ),

sgn(f (a)) sgn(f (b)) > 0. (9.2)

If there does not exist such an interval (x1 , x2 ), then x = x0 cannot give the extremum of y = f (x). Why?

Proof. Suppose there exists an interval x1 < x0 < x2 such that sgn(f (a)) sgn(f (b)) < 0, specifically, sgn(f (a)) >
0 and sgn(f (b)) < 0, for any a ∈ (x1 , x0 ) and any b ∈ (x0 , x2 ).
Then it implies f (x0 ) = 0. (Why?) So we have f (b) < f (x0 ) = 0 < f (a) for any a ∈ (x1 , x0 ) and any
b ∈ (x0 , x2 ). This is a contradiction on the big assumption that f (x0 ) is a extremum. 
In a nutshell, in general, if a function g has a critical number x = c, then

g has an extremum at x = c
⇐⇒ sgn(g ′ (a)) sgn(g ′ (b)) < 0 and sgn(g(a)) sgn(g(b)) > 0

for any a ∈ (x1 , c) and b ∈ (c, x2 ) near x = c.

Step 3: Now we prove that x = x0 yields the extremum of y = f 2 (x) by showing (9.1) and (9.2) on y = f 2 (x):

(f 2 )′ (x) = 2f (x)f ′ (x), sgn((f 2 )′ (x)) = sgn(2) sgn(f (x)) sgn(f ′ (x)) = sgn(f (x)) sgn(f ′ (x)).

(1) For any a ∈ (x1 , x0 ) and b ∈ (x0 , x2 ) in the small interval (x1 , x2 ) of x0 ,

sgn((f 2 )′ (a)) sgn((f 2 )′ (b)) = sgn(f (a)) sgn(f ′ (a)) sgn(f (b)) sgn(f ′ (b))
1
Through the books or professors or by yourself, one may find a nicer and simpler proof of the given problem.

Page 9 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

= sgn(f (a)) sgn(f (b)) sgn(f ′ (a)) sgn(f ′ (b)) < 0

by the inequalities (9.1) and (9.2) on the extremum of x = x0 for y = f (x).

(2) For any a ∈ (x1 , x0 ) and b ∈ (x0 , x2 ) in the small interval (x1 , x2 ) of x0 ,

sgn(f 2 (a)) sgn(f 2 (b)) = sgn(f (a)) sgn(f (a)) sgn(f (b)) sgn(f (b)) = [sgn(f (a)) sgn(f (b))]2 > 0,

by the inequality (9.2) on the extremum of x = x0 for y = f (x).

Step 4 Conclusion: Thus, we can conclude x = x0 belongs to B, i.e., A is a subset of B, i.e., any extremum of
y = f (x) is an extremum of y = f 2 (x). Specifically, if y = f (x) has a (local) maximum at x = c, then y = f 2 (x)
has also a (local) maximum at x = c. If y = f (x) has a (local) minimum at x = c, then y = f 2 (x) has also a
(local) minimum at x = c.
False Converse: Is the converse true? That is, if y = f 2 (x) has an extremum at x = c, then can y = f (x) also
have an extremum at x = c? No. We can explain it by giving a counterexample.
Counterexample) Let f (x) = x, i.e., y = f 2 (x) = x2 . Then it has the minimum at x = 0. But y = f (x) = x
does not have any extremum.
True Converse: Suppose f ≥ 0 for all x in the domain. Then those two sets A and B are same, i.e., the converse
becomes true, i.e., if y = f 2 (x) has an extremum at x = x0 , then y = f (x) also has an extremum at x = x0 .

Proof. Case 1. Let x = x0 give an extremum of y = f 2 (x) and f 2 (x0 ) = 0. Then clearly it gives the absolute
minimum value of y = f 2 (x), because f 2 (x0 ) = 0 ≤ f 2 (x) for all x. Moreover, due to the assumption f (x) ≥ 0
for all x, x = x0 gives the absolute minimum value of y = f (x), too, because of f 2 (x0 ) = 0 = f (x0 ) ≤ f (x) for
all x.
Case 2. Let x = x0 give an extremum of y = f 2 (x) but f 2 (x0 ) ̸= 0. Then since x = x0 is the critical number of
y = f 2 (x), so it is easy to see f ′ (x0 ) = 0, i.e., x = x0 is a critical number of y = f (x):

0 = (f 2 )′ (x0 ) = 2f (x0 )f ′ (x0 ) and f (x0 ) ̸= 0 =⇒ f ′ (x0 ) = 0.

Since y = f 2 (x) has an extremum at x = x0 , it satisfies the inequalities (9.1) and (9.2), i.e., for any a ∈ (x1 , x0 )
and b ∈ (x0 , x2 ) near x = x0 (where we choose the small interval (x1 , x2 ) containing x0 not to have any root of
f 2 (x) = 0):
sgn((f 2 )′ (a)) sgn((f 2 )′ (b)) < 0 and sgn(f 2 (a)) sgn(f 2 (b)) > 0.
When we expand them, we have

0 > sgn((f 2 )′ (a)) sgn((f 2 )′ (b))

= sgn(2f (a)f ′ (a)) sgn(2f (b)f ′ (b))
= sgn(2) sgn(f (a)) sgn(f ′ (a)) sgn(2) sgn(f (b)) sgn(f ′ (b))
= sgn(f (a)) sgn(f (b)) sgn(f ′ (a)) sgn(f ′ (b)).

Since f ≥ 0 for all x and the interval (x1 , x2 ) does not have any root of f 2 (x) = 0, i.e., f 0 in the interval
(x1 , x2 ), so the result above becomes

0 > sgn((f 2 )′ (a)) sgn((f 2 )′ (b))

= sgn(f (a)) sgn(f (b)) sgn(f ′ (a)) sgn(f ′ (b))
= sgn(f ′ (a)) sgn(f ′ (b)), and sgn(f (a)) sgn(f (b)) > 0.

Therefore, again by the inequalities (9.1) and (9.2) on y = f (x), y = f (x) has an extremum at x = x0 . 


Page 10 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

10. For each of the following data sets,

1. formulate the mathematical model that minimizes the largest deviation between the data and the line y =
ax + b.
2. If a computer is available, solve for the estimates of a and b.

x 1.0 2.3 3.7 4.7 6.1 7.0

(a)
y 3.6 3.0 3.2 5.1 5.3 6.8
Answer. Part 1. Observing Plot: From the plot of data points, we find a straight line y = ax + b which is close to
all the points. One can choose (x, y) = (2.3, 3.0) and (x, y) = (6.1, 5.3) to find the slope a and the y–intercept
b:
y = 0.6053x + 1.6079. (10.1)
It is fine that you use any other choice of pairs of points to find the slope a and the intercept b.
Part 2. Chebyshev (Optional): When we use the Chebyshev criterion via the computer,

y = 0.5333x + 2.1467 (10.2)

minimizes the largest absolute deviations of the given data with the minimum value 0.92. See the two plots in
the left–hand side of the figure 10. 

Figure 10: Model y = ax + b

x 2.5 3.0 3.5 4.0 4.5 5.0 5.5

(b)
y 4.32 4.83 5.27 5.74 6.26 6.79 7.23
Answer. Part 1. Observing Plot: From the plot of data points, we find a straight line y = ax + b which is close to all
the points. One can choose (x, y) = (2.5, 4.32) and (x, y) = (5.5, 7.23) to find the slope a and the y–intercept
b:
y = 0.97x + 1.895. (10.3)
It is fine that you use any other choice of pairs of points to find the slope a and the intercept b.
Part 2. Chebyshev (Optional): When we use the Chebyshev criterion via the computer,

y = 0.98x + 1.855. (10.4)

minimizes the largest absolute deviations of the given data with the minimum value 0.035. See the two plots in
the right–hand side of the figure 10. 

Page 11 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

11. For the following data,

1. formulate the mathematical model that minimizes the largest deviation between the data and the model
y = c1 x2 + c2 x + c3 .
2. If a computer code is available, solve for the estimates of c1 , c2 and c3 .
x 0.1 0.2 0.3 0.4 0.5
y 0.06 0.12 0.36 0.65 0.95
Answer. Part 1. Observing Plot: From the plot of data points, we find a quadratic curves y = c1 x2 + c2 x + c3
which is close to all the points. One can choose (x, y) = (0.1, 0.06) and (0.3, 0.36) and (0.5, 0.95) to find the
coefficients c1 , c2 and c3 :
y = 3.625x2 + 0.05x + 0.0188. (11.1)
It is fine that you use any other choice of points to find the coefficients.
Part 2. Chebyshev (Optional): When we use the Chebyshev criterion via the computer,

y = 4x2 − 0.0333x − 0.0050 (11.2)

minimizes the largest absolute deviations of the given data with the minimum value 0.0283. See the figure 11.


Figure 11: Model y = c1 x2 + c2 x + c3

Page 12 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

12. For the following data,

1. formulate the mathematical model that minimizes the largest deviation between the data and the model
P = aebt .
2. If a computer code is available, solve for the estimates of a and b.
t 7 14 21 28 35 42
P 8 41 133 250 280 297
Answer. Taking the natural logarithmic function on the suggested model P = aebt , it becomes

ln P = ln(aebt ) = ln a + ln ebt = ln a + bt ln e = ln a + bt, i.e., ln P = bt + ln a.

Now we transform the data, only P values:

t 7 14 21 28 35 42
P 8 41 133 250 280 297
ln P 2.0794 3.7136 4.8903 5.5215 5.6348 5.6937

Part 1. Observing Plot: From the plot of transformed data points, we find a straight line ln p = bt + ln a which is
close to all the points. One can choose (t, ln P ) = (14, 3.71357) and (t, ln P ) = (35, 5.63479) to find the slope b
and the (ln P )–intercept ln a:

ln P = 0.0915t + 2.4328, i.e., P = e2.4328 e0.0915t = 11.3903e0.0915t . (12.1)

It is fine that you use any other choice of pairs of points to find the slope b and the intercept ln a.
Part 2. Chebyshev (Optional): When we use the Chebyshev criterion via the computer,

ln P = 0.1033t + 2.0387, i.e., P = e2.0387 e0.1033t = 7.6810e0.1033t (12.2)

minimizes the largest absolute deviations of the given data with the minimum value 0.6822. See the figure 12.


Figure 12: Model P = aebt

• Aside: Again, the data on the population can be approximated best by the model
µ
P = .
b + ce−µt
However, it is not easy to estimate the parameters µ, b and c. See the Aside on page 7.

Page 13 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 3.3 Applying the Least–Squares Criterion

13. 1. Use the solutions of the normal equations coming from the least–squares criterion to estimate the coeffi-
cients of the line y = ax + b such that the sum of the squared deviations between the line and the following
data points is minimized.
For each problem,
√∑
2. compute D = ( m i=1 di )/m and dmax to bound cmax , where di = |yi − axi − b| and dmax = maxi di .
2

3. Compare the results to your solutions to Problem 10 in Section 3.2 on page 11.

x 1.0 2.3 3.7 4.7 6.1 7.0

(a)
y 3.6 3.0 3.2 5.1 5.3 6.8
Answer. Part 1. Least–Squares: For the suggested model y = ax + b, the Least–Squares criterion gives the normal
equations,
a∥x∥2 + bx · i = x · y and ax · i + b∥i∥2 = y · i,
where x = ⟨x1 , x2 , . . . , xm ⟩ and y = ⟨y1 , y2 , . . . , ym ⟩ and i = ⟨1, 1, . . . , 1⟩ and m is the number of the given
data. Solving the equations for a and b, we have
1 [ ] 1 [ ]
a= (x · y)∥i∥2 − (x · i)(y · i) , b= (y · i)∥x∥2 − (x · y)(x · i) , (13.1)
z z
where z = ∥x∥2 ∥i∥2 − (x · i)2 .
Putting x = ⟨1.0, 2.3, 3.7, 4.7, 6.1, 7.0⟩ and y = ⟨3.6, 3.0, 3.2, 5.1, 5.3, 6.8⟩ and i = ⟨1, 1, 1, 1, 1, 1⟩ into the
formula (13.1) on a and b above, we have

a = 0.5680, b = 2.1522

and so the suggested model is in fact y = 0.5680x + 2.1522.

Part 2. Bounds: A computation gives the table:

x 1.0 2.3 3.7 4.7 6.1 7.0

y 3.6 3.0 3.2 5.1 5.3 6.8
di 0.8798 0.4586 1.0539 0.2781 0.3171 0.6717

where di = |yi − 0.5680xi − 2.1522|. From the table, the maximum value is dmax = 1.0539 and D = 0.6738.

D = 0.6738 ≤ cmax ≤ 1.0539 = dmax .

Part 3. Comparison with Problem 10: Using the result (10.2), y = 0.5333x − 2.1467 (One can use the line (10.1) by
observing the plot), we have the table:

x 1.0 2.3 3.7 4.7 6.1 7.0

y 3.6 3.0 3.2 5.1 5.3 6.8
ci 0.92 0.3733 0.92 0.4467 0.1 0.92

where ci = |yi − 0.5333xi − 2.1467|. From the table, the maximum value is cmax = 0.92. Hence,

D = 0.6738 ≤ cmax = 0.92 ≤ 1.0539 = dmax .

See the left one of the figure 13. 

Page 14 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

x 2.5 3.0 3.5 4.0 4.5 5.0 5.5

(b)
y 4.32 4.83 5.27 5.74 6.26 6.79 7.23

Answer. Part 1. Least–Squares: Since we have the same type of model as above, we use the formula (13.1) in the
solution above. So putting the vectors coming from the given data,

x = ⟨2.5, 3.0, 3.5, 4.0, 4.5, 5.0, 5.5⟩ , y = ⟨4.32, 4.83, 5.27, 5.74, 6.26, 6.79, 7.23⟩ ,

and i = ⟨1, 1, 1, 1, 1, 1, 1⟩, into the formula (13.1), we have

a = 0.9743, b = 1.88

and so the suggested model is in fact y = 0.9743x + 1.88.

Part 2. Bounds: A computation gives the table:

x 2.5 3.0 3.5 4.0 4.5 5.0 5.5

y 4.32 4.83 5.27 5.74 6.26 6.79 7.23
di 0.0043 0.0271 0.02 0.0371 0.0043 0.0386 0.0086

where di = |yi − 0.9743xi − 1.88|. From the table, the maximum value is dmax = 0.0386 and D = 0.0242.

D = 0.0242 ≤ cmax ≤ 0.0386 = dmax .

Part 3. Comparison with Problem 10: Using the result (10.4), y = 0.98x + 1.855 (One can use the line (10.3) by
observing the plot), we have the table:

x 2.5 3.0 3.5 4.0 4.5 5.0 5.5

y 4.32 4.83 5.27 5.74 6.26 6.79 7.23
ci 0.015 0.035 0.015 0.035 0.005 0.035 0.015

where ci = |yi − 0.98xi − 1.855|. From the table, the maximum value is cmax = 0.035. Hence,

D = 0.0242 ≤ cmax = 0.035 ≤ 0.0386 = dmax .

See the right one of the figure 13. 

Figure 13: Model y = ax + b

Page 15 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Figure 14: Model P = aebt

14. 1. Make an appropriate transformation to fit the model P = aebt using the normal equations coming from the
least–squares criterion.
2. Estimate a and b.
√∑
3. Compute D = ( m i=1 di )/m and dmax to bound cmax , where di = |yi − axi − b| and dmax = maxi di .
2

4. Compare the results to your solutions to Problem 12 in Section 3.2 on page 13.
t 7 14 21 28 35 42
P 8 41 133 250 280 297
Answer. Part 1. Least–Squares: For the suggested model P = aebt , we transform the data and the model (resulting
in ln P = bt + ln a) as we already did in the Problem 12, Section 3.2 on page 13. To the transformed data, we
apply the Least–Squares criterion, which gives the normal equations,
b∥X∥2 + (ln a)X · i = X · Y and bX · i + (ln a)∥i∥2 = Y · i,
where X = ⟨t1 , t2 , . . . , tm ⟩ and y = ⟨ln P1 , ln P2 , . . . , ln Pm ⟩ and i = ⟨1, 1, . . . , 1⟩ and m is the number of the
given data. Solving the equations for b and ln a, we have
1 [ ] 1 [ ]
b= (X · Y )∥i∥2 − (X · i)(Y · i) , ln a = (Y · i)∥X∥2 − (X · Y )(X · i) , (14.1)
Z Z
where Z = ∥X∥2 ∥i∥2 − (X · i)2 .
Putting x = ⟨7, 14, 21, 28, 35, 42⟩ and y = ⟨ln 8, ln 41, ln 133, ln 250, ln 280, ln 297⟩ and i =< 1, 1, 1, 1, 1, 1 >
into the formula (14.1) on b and ln a above, we have
b = 0.0999, ln a = 2.1423.
and so the suggested model is in fact ln P = 0.0999t + 2.1423, i.e., P = e2.1423 e0.0999t = 8.5188e0.0999t .
Part 2. Bounds: A computation gives the table:

(t, P ) (7, 8) (14, 41) (21, 133) (28, 250) (35, 280) (42, 297)
di 9.1381 6.5214 63.6356 110.452 0.7440 267.803

where di = |Pi − 8.5188e0.0999ti |. From the table, the maximum value is dmax = 267.803 and D = 121.171.
D = 121.171 ≤ cmax ≤ 267.803 = dmax .
Part 3. Comparison with Problem 12: Using the result (12.2), P = 7.6810e0.1033t (One can use the line (12.1) by
observing the plot), we have the table:

(t, P ) (7, 8) (14, 41) (21, 133) (28, 250) (35, 280) (42, 297)
ci 7.8252 8.3951 65.8236 111.5956 5.1564 290.5115

where ci = |Pi − 7.6810e0.1033ti |. From the table, the maximum value is cmax = 290.5115.
D = 121.171 ≤ cmax = 290.5115 ≤ 267.803 = dmax .
See the left one of the figure 14. 
Page 16 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

Section 3.4 Choosing a Best Model

15. For each of the following problems, answer to the following questions.

1. Find a model using the least–squares criterion either on the data or on the transformed data (as appropri-
ate).
2. Compare your results with the graphical fits obtained in the problem set 3.1 by computing the deviations,
maximum absolute deviations and the sum of the squared deviations for each model.
3. Find a bound on cmax if the model was fit using the least–squares criterion.

(15.1) Problem (6.1) and (6.3) in Section 3.1 on page 4.

Answer. • Problem (6.1) P = aV b : 1. For the transformed data (ln V, ln P ), the least–squares criterion
gives
ln P = 8.0063 ln V − 0.2266 =⇒ P = 0.7973V 8.0063 . (15.1)
2. Let fG and fL be the models obtained by the graphical fits in Section 3.1 and the one obtained by the
least–squares criterion, respectively. Then, by the results (6.1) and (15.1),
P = fG (V ) = 0.0561V 9.55 , P = fL (V ) = 0.7973V 8.0063 .
∑
max |Pi − f (Vi )| i (Pi − f (Vi ))
2

Model from (6.1) 528639 6.4005 × 1011

Least–Squares (15.1) 537950 5.8242 × 1011
That is, the least–squares result is better than the graphical fit in the sense of minimizing the sum of squares
of deviations.
3. Bounds.
D = 197049 ≤ cmax ≤ 537950 = dmax .
• Problem (6.3) P = a(bV ): 1. For the transformed data (V, ln P ), the least–squares criterion gives
( )
ln P = 2.0691V + 2.3097 =⇒ P = 10.0717 7.9180V . (15.2)
2. Let fG and fL be the models obtained by the graphical fits in Section 3.1 and the one obtained by the
least–squares criterion, respectively. Then, by the results (6.2) and (15.2),
( ) ( )
P = fG (V ) = 38.4646 5.7199V , P = fL (V ) = 10.0717 7.9180V .
∑
max |Pi − f (Vi )| i (Pi − f (Vi ))
2

Model from (6.2) 645613 8.0043 × 1011

Least–Squares (15.2) 843499 1.0182 × 1012
That is, the least–squares result is worse than the graphical fit in the sense of minimizing the sum of squares
of deviations.
3. Bounds.
D = 260534 ≤ cmax ≤ 843499 = dmax . 
(15.2) Problem 7.2 in Section 3.1 on page 6.

Answer. 1. For the transformed data (t, ln P ), the least–squares criterion gives
ln P = 0.0999t + 2.1423 =⇒ P = 8.5187e0.0999t (15.3)
2. Let fG and fL be the models obtained by the graphical fits in Section 3.1 and the one obtained by the
least–squares criterion, respectively. Then, by the results (7.1) and (15.3),
P = fG (t) = 11.3903e0.0915t , P = fL (t) = 8.5187e0.0999t .

Page 17 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

∑
max |Pi − f (ti )| − f (ti ))2
i (Pi
Model from (7.1) 234.531 68714.3
Least–Squares (15.3) 268.7 88536.4

That is, the least–squares result is worse than the graphical fit in any sense. How can this happen? See the Aside
following the answer.
3. Bounds.
D = 121.475 ≤ cmax ≤ 268.7 = dmax .
See the figure 15. 

Figure 15: Blue: fG (Graphical Fit), Red: fL (Least-Squares Fit)

• Aside In the problem above, we have the case that the least–square fit is worse than the graphical fit. How
can this happen? Let us redescribe what we have done.
1. Graphical Fit: We chose two points (t, P ) = (14, 41) and (35, 280). Using these two points, we found a and
b in the model P = aebt . Even if we find ln a and b in the transformed data and transformed model, the values
of a and b are same. That is, before and after the transformation, the values of a and b do not change and the
graphical fit model passes through two points, the second and the fifth.
2. Least–Squares Fit: We used the least–squares fit on the original data (t, P ) and got a = 8.51901 and b =
0.0999 via a computer. Even if we use the least–squares fit on the transformed data (t, ln P ), we have the
same parameters. But the resulting model is worse than the graphical fit. Why? When we look for a line,
ln P = bt + ln a, the least–squares fit gives the best a and b. However those best constants become worse via the
backward transformation, i.e., P = aebt , because those constants were chosen to be THE BEST FOR THE STRAIGHT
LINE rather than a curve. As a proof, look at the table:

t 7 14 21 28 35 42
ln P ln 8 ln 41 ln 133 ln 250 ln 280 ln 297
| ln Pi − 0.0915ti − 2.43276| (Graphical) 0.9937 0 0.5364 0.5271 0 0.5815
| ln Pi − 0.0999ti − 2.14227| (Least–Square) 0.7619 0.1732 0.6510 0.5831 0.0027 0.6427

∑
max | ln Pi − bti − ln a| i (ln Pi − bti − ln a)2
Graphical Fit on Transformed Data 0.9937 1.8911
Least–Squares Fit on Transformed Data 0.7619 1.7873

Page 18 of 19
Mathematical Modeling HOMEWORK 2 – SOLUTION Spring, 2010

That is, on the transformed data, the Least–Squares Fit is better in any sense. See the one in the right–hand side
of the figure 15.

Page 19 of 19