You are on page 1of 24

Models on Conservation Laws

Joaquim MC. Correia and Jayrold P. Arcede


September 5, 2015

Abstract
This talk is all about Conservation Laws models. At the beginning, we conceptualized a toy model
where we build a mathematical equation and find its solution.

Introduction

We will begin rst to talk about matter. What is matter? In highschool physics, matter is dened to
be anything that occupies space. All objects take up space. Your computer is taking up space on the
desk. You are taking up space on the chair. Objects have mass. Mass is how much there is of an object.
Mass is related to how much something weighs. But note that mass and weight are two dierent things.
The unit for mass is a gram. For example, a nickel has the mass of about one gram. Objects that take
up space and have mass are called matter1 . We also note that the matter of which we, human being,
is made of remains the same, whether on Earth or on the moon. Weight is not a good measure of how
much matter one has. Hence mass, independent of weight, is used to describe the amount of matter.
Now most variable like matter in our case were looked as continuous variable. Here, we will treat
matter to be composed of points (which can thought of atoms, molecules, etc) that dont have mass,
otherwise, the aggregation (or sum) of masses would be innite. Imagine a chair, which has an innite
mass!(??). This is absurd, and that not make sense. But its better to treat matter as continuous variable,
seeing in macroscopic view, so that later, we can talk about density.
Now, the idea of looking matter as a continuous variable originate from Hemholtz. In [McCormmach],
we quote the following in verbatim:
In the introductory lectures, Helmholtz discussed in general terms the two ways physicist
have of treating bodies. Depending on the problem at hand, they regard bodies either as
aggregates of material points or as volume elements lled with matter....In the rst set of
lectures..., Helmholtz treated the dynamics of discrete material points, or mass points. The
mass point is a useful concept for solving problems in which the real form and extension of a
body can be neglected.
For the description of the motion of a mass point, which is one of the rst and most
important tasks of theoretical physics, the position of the point must be dened by continuous
and dierentiable functions of time; otherwise, the point could be in two places at the same
time, disappearing here and instantly reappearing there. Such a discontinuity would violate
the identity of the mass point, which derives from a fundamental law of experience of all
natural phenomena:matter cannot be created or destroyed....
In the second set of dynamical lectures, Helmholtz elucidated the central concept of continuously distributed masses, comparing it with the concept of discrete mass points. The
dierence between the two is evident from the calculation of density, or the ratio of mass to
volume: within the picture of continuous masses, we can imagine a close volume of diminishing
From

his Evora Lecture Series


is discontinuous and broken. It is composed of tiny discrete particles called atoms (looking at microscopic level).
These atoms are held together by strong attractive forces called bondsthis is what gives matter its appearance of continuity
(from Gideon Ifianayi, Professor of Chemistry).
1 Matter

smallness in which the density approaches a limiting value at a given position; by contrast,
within the picture of discrete masses, we can imagine a suciently small closed volume containing only a few mass points, so that in this case it makes no sense to speak of a limiting
value of the density.
Although the concept of continuously distributed masses corresponds to our direct sense
impressions of sight and touch, Helmholtz cautioned that we cannot conclude that matter is
actually continuous. Of the nal division of matter we know nothing; all we can do is form
hypotheses about it to explain the properties of bodies, mechanical, thermal, and chemical.
We generally work with the hypothesis that the nal division is into atoms and molecular
groupings of atoms, the main properties of which correspond to the picture of discrete mass
points interacting through characteristic central forces. But when we treat phenomena by
mathematically dividing bodies into volume elements that are large compared with molecular
separations, as we do, for example in hydrodynamics and elasticity theory, we then work within
the picture of continuously distributed masses rather than atomistically structured ones 2 . Even
here, Helmholtz noted, there are problems such as dispersion of light in which the simple
picture of continuously distributed masses is insucient and the molecular hypothesis must
be invoked.
After introducing the fundamental concepts for treating the dynamics of continuously
distributed masses, Helmholtz described how we apply mathematics to them. We introduce
partial dierential equations, the mathematical language of dynamics of systems with more
than one independent variable. We imagine cell walls distributed throughout a continuous
body, dividing it into masses that retain their identity as the body moves as a whole. Without
misunderstanding, we may speak of mass point in the present case as we can in that of the
dynamics of discrete masses. Only the meaning is dierent here: the mass point now has no
denite mass; it is a shorthand expression for the corner point of a volume element lled with
mass. Exactly as in molecular picture, in the picture of continuous masses the positions of the
mass points can be dierentiated with respect to time; since however, these mass points form
a continuum, as they do not in the molecular picture, their velocities can be dierentiated
not only with respect to time, to construct their acceleration, but also with respect to the
three spatial coordinates. It is this mathematical distinctionthe presence here of spatial
coordinates as variables with which the velocities can be dierentiatedthat is the essential
characteristic of continuously distributed masses in contrast to systems of discrete mass points
in which time is the only primitive variable.

Setting

Now, consider [x1 , x2 ] R, which can be thought of a (homogeneous) medium, maybe some metal rod,
etc. with boundary points at x1 and x2 . Then we characterized continuous mass in [x1 , x2 ] by some
density (x, t), where t can be thought of as time and x as spatial variable. Note that x can be a vector.
The density (x, t) is a quantity of interest and usually is unknown. In some applications, (x, t) may be
the temperature of a rod, the pressure of a uid, the concentration of a chemical or a group of cells, or
the density of cars in a trac setting.
Suppose we interpret (x, t) as mass points in [x1 , x2 ], then the total mass inside [x1 , x2 ] for some
time t can be expressed mathematically as
x2
(x, t)dx.
(1)
x1

Now, since (x, t) is a function of time, then two things can happen to our medium: there will be some
instance which temperature rises which make the medium expands, or other times, make the medium
contracts, that is there will be some amount of density that goes in or go out in the domain [x1 , x2 ]. This
ow of density created in expanding or contracting of medium is called ux, denoted by q(x, t). This ux
is observable quantity, which is a function of space and time. Flux represents the amount of density that
2 emphasis

mine.

either goes in, q(x1 , t), or comes out, q(x2 , t), of the domain [x1 , x2 ]. To set the orientation of the ow
of ux we will agree that to the right of x-axis is positive, otherwise, ux would be negative.
Thus, we can express mathematically, the variation in time of the total density in [x1 , x2 ] or the
change with respect to t of (x, t) or the change of mass in time of Equation (1) as the following:

d x2
(x, t)dx = q(x1 , t) q(x2 , t).
(2)
dt x1
Now, we can interpret, Equation (2) as when temperature rises, then the medium will expand. In this
case, some quantity (x, t) goes out in the boundary. In x1 , q(x1 , t) will be negative and in x2 , q(x2 , t) is
positive. Therefore, equation (2) will be negative. This means that there are more (x, t) quantity that
goes out in the system, otherwise, (2) is positive.
According to (2), the quantity is neither created nor destroyed: the total amount of contained
inside any given interval [x1 , x2 ] can change only due to the ow of (x, t) (ux) across the two boundary
points x1 and x2 . That is why, we call (x, t) a conserved quantity. See the diagram below.

Figure 1: Some insulated medium


Thus, we could easily observe that ux, q(x, t) is a function of density , space x and time t, and can
be written in the following useful notation:
q(x, t) := (, x, t).

2.1

(3)

Source or sink

Another factor for the change in (x, t) is due to the source or sink, denoted by (x, t). This is the
quantity of (x, t) introduced in a medium. This quantity is observable. Suppose in our medium, some
creation of mass happens inside [x1 , x2 ]. Then we could sum up the mass creation inside our domain to
be
x
2

(x, t)dx,

(4)

x1

incorporating (4) to (2), we have


x2

d x2
(x, t)dx.
(x, t)dx = q(x1 , t) q(x2 , t) +
dt x1
x1

(5)

To give an emphasis, imagine a closed room of two doors where some crevices in the door or window.
We can treat the doors as would be our endpoint x1 and/or x2 and some crevices as source or sink.
In this case, our density is the amount of heat enters and leaves the room. To monitor the amount of
heat, we suppose that there is some air-conditioning device installed of this room. Then amount of heat
entered or escaped from the room will be measured by the air-con device thru its sensor which regulates
the temperature inside the room.

On some technicalities

To proceed mathematically, as we want to get a partial dierential equation out of equation (5), we seek
to answer the following question.
Question 1. What assumptions in the Henstock sense does the equality holds:
d
dt

x2

??

z}|{
(x, t)dx =

x1

x2
x1

(x, t)dx.
t

(6)

This is called Leibniz rule in Calculus.


Supposing Leibniz rule is true in this some setting, we can express the left hand side of (5) in the
following manner:
x2
x2

(x, t)dx = q(x1 , t) q(x2 , t) +


(x, t)dx.
(7)
t
x1
x1
By some Fundamental Theorem of Calculus and using the notation in Equation (3), we can write the
inow-outow of q(x, t) in the domain as
x1
x2

((, x, t)) dx =
((, x, t)) dx.
(8)
q(x1 , t) q(x2 , t) =
x
x
x2
x1
So equation (7) becomes
x2

(x, t)dx =
t

x2

((, x, t)) dx +
x

x2

(x, t)dx.

(9)

In the subsequent, it would be useful to have some shorthand notation:


Lets denote

(x, t) as t (x, t);


t
and

((, x, t)) as x (, x, t).


x
By these notations and some algebra, (9) can be express as
x2
(t (x, t) + x (, x, t) (x, t)) dx = 0 for all x1 , x2 R

(10)

x1

x1

x1

x1

By highlighting that the integral in (10) is zero for all x1 , x2 R, by hiding the parameters , x, and t,
we can have the following expression
x2
(t + x ) dx = 0 for all x1 , x2 R.
(11)
x1

Furthermore, we can write (11) as

(t + x ) [x1 ,x2 ] (x)dx = 0.


{

where
[x1 ,x2 ] (x) =

(12)

if x [x1 , x2 ]
.
if x
/ [x1 , x2 ]

1
0

Now, Equation (12) can only mean one thing, that is, the integrand should be zero. Therefore,
t + x = 0,
or in terms of the ux function q, we have the following expression
4

(13)

t + qx = 0.

(14)

Equations (14) or (13) is the partial dierential equation (pde) that we are interested about. We want
to nd the solution that satises this pde.
Consider the following example.
Example 3.1. Suppose in (14), q(x, t) =

2 (x,t)
2

with no source or sink, i.e, = 0. Then

2 (x, t)
q(x, t) =
= x .
x
x
2
So (14) becomes
t + x = 0.

(15)

What should be the density , which is the solution of equation (15)? The conservation law in (15) is
called the inviscid Burgers equation.
We want to remark that sometimes we get the derivative of with respect to x, x , and not the usual
derivative of with respect to t, i.e., t . In this case, we could interpret x as a transfer of density with
respect to x. Furthermore, if we take another derivative of x , which is xx , then we have diffusion, since
diusion equations involve a second order derivative with respect to space.
To make it clear, consider the ux (x , , x, t) which now a function of x in addition to , x, and
t. By dierentiating q with respect to x, we have

(x , , x, t) = x (x , , x, t) = xx D
x

for some coecient D

(16)

Then equation (14) can be written as


t + x = 0 t + Dxx = .

(17)

If = 0, then (17) can be written in the following manner:

2
= D 2
t
x

(18)

Equation (18) is called diusion equation for some particle density and D > 0 is called a diusion
coecient.

3.1

On some general form of flux and source/sink

Now, recall the form of ux q(x, t) = ((x, t), x, t) and suppose the source or sink (x, t) = (, x, t),
that is, both are function of density (x, t), which was not considered previously. Then equation (14) can
be written as

t + x ((x, t), x, t) = +
t

z(

chain rule

}|
){

+
1 =

x
x

(19)

t + ( x + x ) =
t + x = x
which can also be written as
t + c(, x, t)x (x, t) = g(, x, t)

(20)

where
c(, x, t) = (, x, t)
and
g(, x, t) = (, x, t) x (, x, t).
5

(21)

Now, we have some caveat with regards to the above equations (19) and (14). And we feel strongly
that this should be taken with extra care. We have to be vigilant in the following manner: if = 0,
equation (14) becomes
t + qx = 0 a homogeneous pde.
(22)
However, equation (20) is not a homogeneous pde for when = 0, g = x in (21).
Hence,
t + cx = g = 0.

(23)

Thus, an extra care should be taken when considering such forms of pde.
To put more emphasis, common mistake would be like in the following. Consider the following equation
t + qx = .

(24)

Then equation (24) is a generalization of


t + x = 0,
which is false! For one to not be able to commit this blunder, one should go back to the modelization of
these equations to get things clear.

Differentiation under Integral sign

Before we go further, lets address the Question 1 that was posed previously.
The question is: On what assumptions does the the following equation holds, in the sense of Henstock:

x2
d x2

(x, t)dx =
(x, t)dx.
(25)
dt x1
t
x1
The rule: the t-derivative of the integral of (x, t) is the integral of the t-derivative of (x, t), is called
dierentiation under the integral sign. This is what equation (25) is all about.
To answer the question we will see how it works in Riemann sense, in Lebesgue and nally on Henstock
integration. In this manner, we will have a good grasp and foundation of the theory. To accomplish this,
we shall gives examples and counterexamples along the way.
To give a brief background, the method of dierentiation under the integral sign is due to Leibniz. It
concerns integrals depending on a parameter, such as
1
x2 etx dx.
(26)
0

Here t is the extra parameter. (Since x is the variable of integration, x is not a parameter.) In light of
(25), we might write such an integral as
b
(x, t)dx,
(27)
a

where (x, t) is a function of two variables like (x, t) = x2 etx . The integral in (26), is called a gamma
function, which is a generalization of a factorial, it can be solved using integration by parts but using
dierentiation under integral sign is also quite elegant, we refer the reader to the excellent exposition of
Keith Conrad [Conrad], where our rst example is due to him. So we begin.
Example 4.1. Let (x, t) = (2x + t3 )2 . Then

(2x + t3 )2 dx.

(x, t)dx =
0

An anti-derivative of (2x + t3 )2 with respect to x is 16 (2x + t3 )3 , so

x=1

1
3 3
(2x + t ) dx = (2x + t )
6
x=0

3 2

(2 + t3 )3 t9
6
4
= + 2t3 + t6 .
3

(28)

This answer is a function of t, which makes sense since the integrand depends on t. We integrate over x
and are left with something that depends only on t, not x.
b
An integral like a (x, t)dx is a function of t, so we can ask about its t-derivative, assuming that (x, t)
is nicely behaved. In this example, the integral is nicely behaved in the since that (28) is a continuous
function, so no question that we can get the t-derivative. Now, we will compute the t-derivative of (28):
d
dt

d 4
+ 2t3 + t6
dt 3
=6t2 + 6t5 .

(2x + t3 )2 dx =
0

(29)

Now, to check if (25) is true, we also compute the t-derivative of the integrand and then get the integral
wrt x and we have

1
0

(2x + t3 )2 dx =
2(2x + t3 )(3t2 )dx
t
0
=6t2 + 6t5 ,

(30)

which agrees to equation (29).


If we are used to thinking mostly about functions with one variable, not two, keep in mind that (25)
involves integrals and derivatives with respect to separate variables: integration with respect to x and
differentiation with respect to t.
We have seen in the example above where dierentiation under the integral sign can be carried out
awlessly, but we have not actually stated conditions under which (25) is valid. Something does need
to be checked. In [Talvila2], an incorrect use of dierentiation under the integral sign due to Cauchy is
discussed, where a divergent integral is evaluated as a nite expression. Here are two other examples
where dierentiation under the integral sign does not work. We begin with the following counterexample.
Example 4.2. It is pointed out in [Goel, Example 6] that the formula

sin x

dx = ,
x
2
0
leads to an erroneous instance of differentiation under the integral sign. To see this, rewrite the formula
as

sin(ty)
dy =
(31)
y
2
0
for any t > 0, by the change of variables x = ty. Then differentiation under the integral sign implies

cos(ty)dy = 0,
0

which doesnt make sense.


The next example shows that even if both sides of (25) make sense, they need not be equal.
7

Example 4.3. For any real numbers x and t, let


{
xt3 /(x2 + t2 )2 ,
(x, t) =
0,
Let

For instance, F (0) =

(x, t)dx.
0

(32)

F (t) =

if x = 0 or t = 0,
if x = 0 and t = 0.

0dx = 0. When t = 0,

(x, 0)dx =
0

xt3 /(x2 + t2 )2 dx

F (t) =
0

1+t2

=
x2

t3
dx
2u2

(where u = x2 + t2 )

(33)

t
.
2(1 + t2 )

This formula also works at t = 0, so F (t) = t/(2(1 + t2 )) for all t. Therefore, F (t) is differentiable
and
1 t2
F (t) =
2(1 + t2 )2
for all t. In particular, F (0) = 12 .

(x, t)dx. Since (0, t) = 0 for all t, (0, t) is differentiable


0 t

in t and t
(0, t) = 0. For x = 0, (x, t) is differentiable in t and
Now we compute

t (x, t)

and then

xt2 (3x2 t2 )
(x, t) =
t
(x2 + t2 )3 .

(34)

Combining both cases (x = 0 and x = 0),

(x, t) =
t

xt2 (3x2 t2 )
(x2 +t2 )3 ,

if x = 0

0,

if x = 0

(35)



In particular, t
(x, t) = 0. Therefore at t = 0 the left side of equation (25) is F (0) = 1/2 and
1 t=0

the right side is
(x, t)dx = 0. The two sides are unequal!
t
t=0
0

The problem in this example is that t


(x, t) is not a continuous function of (x, t). Indeed, the
2
denominator in the formula in (35) is (x + t2 )3 , which has a problem near (0, 0). Specically, while this

(x, t) has
derivative vanishes at (0, 0), it we let (x, t) (0, 0) along the line x = t, then on this line t
the value 1/(4x), which does not tend to 0 as (x, t) (0, 0).
We posed the following theorem.

Theorem 4.1. The equation


d
dt

(x, t)dx =
a

(x, t)dx.
t

(36)

is valid at t = t0 , in the sense that both sides exist and are equal, provided the following two conditions
hold:

(x, t) and t
(x, t) are continuous functions of two variables when x is in the range of integration
and t is in some interval around t0 ,


there are upper bounds |(x, t)| A(x) and | t
(x, t)| B(x), both being independent of t, such
b
b
that a A(x)dx and a B(x)dx exist

To understand it better, we will quote verbatim Chapter XIII section 3 of Serge Lang Undergraduate
Analysis book.

4.1

Integration under integral sign: Riemann Sense

Theorem 4.2. Let f be a continuous function of two variables (t, x) defined for t a and x in some
compact set of numbers S. Assume that the integral

f (t, x)dt = lim

converges uniformly3 for x S. Let

f (t, x)dt
a

g(x) =

f (t, x)dt.
a

Then g is continuous.
Proof. For given x S we have


g(x + h) g(x) =
f (t, x + h)dt
f (t, x + h)dt by denition
a
a

=
f (t, x + h) f (t, x + h)dt since integrals is linear on continuous functions.
a

(37)
Given , select B such that for all y S we have





f (t, y) dt <

B

Then by triangle inequality






B





|g(x + h) g(x)|
f (t, x + h) f (t, x) +
f (t, x + h) +
f (t, x) .
a

B
B

(38)

We know that f is uniformly continuous on the compact set [a, B] S. Hence there exists such that
whenever |h| < we have
|f (t, x + h) f (t, x)| /B.
The rst integral on the right is then estimated by B/B = . The other two are estimated each by , so
we have a 3-proof for the theorem.
We shall now prove a special case of the theorem concerning dierentiation under the integral sign
which is sucient for many applications, in particular those of the next chapter. It may be called the
absolutely convergent case.
Theorem 4.3. Let f be a function of two variables (t, x) defined for t a and x in some interval
J = [c, d], c < d. Assume that D2 f exists, and that both f and D2 f are continuous. Assume that there
are functions (t) and (t) which are 0, such that |f (t, x)| (t) and
|D2 f (t, x)| (t),
3 A sequence of functions {f (x)} with domain D converges uniformly to a function f (x) if given any > 0 there is a
n
positive integer N such that |fn (x) f (x)| for all x D whenever n N . (Please note that the above inequality must
hold for all x in the domain, and that the integer N depends only on .

for all t, x and such that the integrals

(t) dt

and

(t) dt

converge. Let

f (t, x) dt.

g(x) =
a

Then g is differentiable, and

Dg(x) =

D2 f (t, x) dt.
a

Proof. We have




f (t, x + h) f (t, x)
g(x + h) g(x)





dt.

D
f
(t,
x)dt

D
f
(t,
x)
2
2




h
h
a
a
But

(39)

f (t, x + h) f (t, x)
D2 f (t, x) = D2 f (t, ct,h ) D2 f (t, x).
h

Select B so large that

(t) dt < .
B

Then we estimate our expression by

=
a

+
a

.
B

Since D2 f is uniformly continuous on [a, B] [c, d], we can nd such that whenever |h| < ,
|D2 f (t, ct,h ) D2 f (t, x)| <

.
B

The integral between a and B is then bounded by . The integral between B and is bounded by 2
because


f (t, x + h) f (t, x)


D2 f (t, x) 2(t).

h
This proves our theorem.
Above theorem is the elementary calculus version, that is, in Riemann sense. We will give another
version in Lebesgue sense. Our source is [Swartz].

4.2

Integration under integral sign: Lebesgue Sense

First and foremost, we want to say something why people move out from Riemann to a powerful enough
integral like Lebegue. While the Riemann integral enjoys many desirable properties, it also has several
shortcomings. One of these shortcomings concerns the fact that a general form of the Fundamental
Theorem of Calculus does not hold for Riemann integrable functions. Another serious drawback is the
lack of good convergence theorems for the Riemann integral. A convergence theorem for an integral
concerns a sequence of integrable functions {fk }
k=1 which converge in some sense, such a pointwise, to a
limit function f and involves
sucient
conditions
for interchanging the limit and the integral, that is to

guarantee limk fk = limk fk . In modern integration theories, the standard convergence theorems are
the Monotone Convergence Theorem, in which the functions converge monotonically, and the Bounded
Convergence Theorem, in which the functions are uniformly bounded. These convergence theorems are
valid in the case of Lebesgue, which is also true for Henstock-Kurzweil integral.
Here, we will list as a recall of what we already know the convergence theorems like, monotone,
bounded and dominated in the sense of Lebesgue.
We denote the collection of measurable subsets of Rn by Mn .
We state the monotone convergence theorem.
10

Theorem 4.4 (Monotone Convergence Theorem). Let E Mn and {fk }


k=1 be an increasing sequence
of nonnegative, measurable functions defined on E. Set f (x) = limk fk (x). Then,

lim
fk =
f
k

For Lebesgue integrable functions, we can get an improvement of the Monotone Convergence Theorem
below.
Corollary 4.1. Let E Mn and {fk }
k=1 be an increasing sequence of nonnegative, Lebesgue integrable functions defined on E. Set f (x) = limk fk (x). Then, f is Lebesgue integrable if, and only if,
supk E fk < . In this case, f is finite a.e..
The Monotone Convergence Theorem is a very useful tool in analysis. However, in many situations, the
monotonicity condition is not satised by a convergent sequence and other conditions, which guarantee the
exchange of the limit and the integral are desirable. We next consider Lebesgues Dominated Convergence
Theorem. This result replaces the monotonicity condition of the Monotone Convergence Theorem by the
requirement that the convergent sequence of functions be bounded by a Lebesgue integrable function. As
a corollary of the Dominated Convergence Theorem, we will get the Bounded Convergence Theorem.
Theorem 4.5 (Dominated Convergence Theorem). Let {fk }
k=1 be a sequence of measurable functions
defined on a measurable set E. Suppose that {fk }
k=1 converges to f pointwise almost everywhere and
there is a Lebesgue integrable function g such that |fk (x)| |g(x)| for all k and almost every x E.
Then, f is Lebesgue integrable and

lim

fk =

Moreover,

f
E

|fk f | = 0

lim

If the measure of E is nite, then constant functions are Lebesgue integrable over E. From the
Dominated Convergence Theorem we get the Bounded Convergence Theorem.
Theorem 4.6 (Bounded Convergence Theorem). Let {fk }
k=1 be a sequence of measurable functions on
a set E of finite measure. Suppose there is a number M so that |fk (x)| M for all k and for almost all
x E. If f (x) = limk fk (x) almost everywhere, then

lim
fk =
f
k

Now, we the help of DCT, we will apply it to continuity of integrals and dierentiation under integral
sign.
We start with some settings.
Let S Rn , T Rm and f : S T R. If f (s, ) : T R is integrable for every s S and
F : S R is dened by F (s) = T f (s, t)dt, we say that the integral F (s) depends on the parameter s.
We consider how the properties of f are inherited by F.
We rst consider the continuity of integrals depending on parameters.
Theorem 4.7. Assume f (s, ) is integrable for every s S and f (, t) is continuous at s0 S for every
t T. If there exists an integrable function g : T R such that |f (s, t)| g(t) for all s S, t T , then
F (s) = T f (s, t)dt is continuous at s0 .
Proof. Let {sk }
k=1 be a sequence from S converging to s0 . By the continuity assumption, f (sk , t)
f (s0 , t). Since |f (sk , )| g(), the Dominated Convergence Theorem implies that

F (sk ) =
f (sk , t)dt F (s0 ) =
f (s0 , t)dt
T

and F is continuous at s0 , as we wished to show.


11

Our second application applies to dierentiating under the integral sign.


Theorem 4.8 (Leibniz Rule). Let S = [a, b] and let I be a closed interval. Suppose that
1.

f
s

= D1 f exist for all s S and t I;

2. each function f (s, ) is integrable over I and there is an integrable function g : I R such that


f (s, t)


s = |D1 f (s, t)| g(t),
for all s S and t I.

Then, F (s) = I f (s, t) dt is differentiable on S and


F (s) =

f (s, t)
dt =
s

D1 f (s, t)dt

(40)

Proof. Fix s0 S and let {sk } be a sequence in S converging to s0 with sk = s0 . For t I, dene {hk }
by
f (sk , t) f (s0 , t)
hk (t) =
sk s0
and note that hk is integrable over I. By the Mean Value Theorem, for each pair (k, t) there is a z(k,t)
between sk and s0 so that
f (sk , t) f (s0 , t)
= D1 f (z(k,t) , t)
sk s0
which implies that



f (sk , t) f (s0 , t)

g(t).
|hk (t)| =

sk s0

The Dominated Convergence Theorem implies that


F (sk ) F (s0 )
sk s0

f (sk , t) f (s0 , t)
= lim
dt
k I
sk s0

= D1 f (s0 , t)dt

F (s0 ) = lim

as we wished to show.

4.3

Integration under integral sign: Henstock Sense

In this section, we will learn how Henstock dier with Lebesgue in terms of convergence theorems and
its applications to dierentiation under integral sign. In this case we follow the exposition by Swartz
[Charles].
It can be noted that the principal reason that the Lebesgue
integral
is favored over the Riemann
integral is the fact that convergence theorems of the form lim I fk = I (lim fk ) hold for the Lebesgue
integral under very general conditions. The major convergence theorems of this type are the Monotone
Convergence Theorem (MCT) and the Dominated Convergence Theorem (DCT). Here we found that
these same major convergence theorems hold for the gauge or the HK integral, showing that the HK
integral enjoys the same advantages over the Riemann integral as the Lebesgue integral.
We begin by introducing the concept of uniform integrability. This is at the center of the convergence
theorems for the HK integral. Lets do some notation: let I be a closed interval (bounded or unbounded)
in R4 and fk , f : I R for k N. Further, let I be the family of all closed subintervals of I.
4 This

means that R is extended by adding the two points at infinity,

12

Definition 4.1. {fk } is uniformly integrable over I if each fk is integrable over I and if for every > 0
there exists a gauge5 on I such that



S(fk , D) fk


I

for every k N whenever D << .6


The point of this Denition is, of course, that the same gauge works uniformly for all k. For uniformly
integrable sequences of integrable functions we have the following convergence theorem.
Theorem 4.9. Let {fk } be uniformly
over
) I and assume that fk f pointwise. Then f is
( integrable

integrable over I and lim I fk = I f = I (lim fk ) .


Now we straight to Monotone Convergence Theorem.
Theorem 4.10 (Monotone Convergence Theorem: MCT). Let fk : I R be integrable over R and
suppose that fk (t) f (t) R for every t R. If supk I fk < , then
1. {fk } is uniformly integrable over I,
2. f is integrable over I and
)

(
3. lim I fk = I f = I (lim fk ) .
The Monotone Convergence Theorem gives a very useful
andpowerful sucient condition guaranteeing
passage to the limit under the integral sign, i.e., lim I fk = I (lim fk ), but the monotone convergence
requirement is often not satised.Thus, another convergence result called the Dominated Convergence
Theorem, which relaxes this requirement.
Theorem 4.11 (Dominated Convergence Theorem: DCT). Let fk , f, g : I R and assume fk , g are
integrable over I with |fi fj | g for all i, j. If lim fk = f pointwise on I, then
1. {fk } is uniformly integrable over I,
2. f is integrable over I and
)

(
3. lim I fk = I f = I (lim fk ) .
Remark 4.1. The usual dominating hypothesis in the DCT for the Lebesgue integral is that there
exists an integrable function g such that |fj | g for all j. Since |fi fj | |fi | + |fj | , this hypothesis
clearly implies the one in Theorem 4.11 [see also Exercise 4.1]. On the other hand, the hypothesis
|fi fj | g allows the functions fj to be conditionally integrable (just take any conditionally integrable
function h and consider the sequence {fj + h}) and is more appropriate for a conditional integral like
the HK integral. Exercise 4.2 gives another equivalent way of phrasing the dominating hypothesis in
Theorem 4.11.
Exercise 4.1. Let fk : I R be integrable over I and fk f pointwise. Suppose there exists an
integrable function g such that |fk | g for all k. Show that in this case the conclusion of the DCT can
be improved to read I |fk f | 0.
Exercise 4.2. Let fk : I R be integrable over I. Show that there exists an integrable function
g : I R satisfying |fk fj | g for all k, j if and only if there exist integrable functions g and h
satisfying g fk h for all k.
function defined on I such that (t) is an open interval containing t for each t I is called a gauge on I
= {(ti , Ii )} is -fine, i.e., ti Ii (ti ).

5 Any
6D

13

We now give some examples illustrating the use of the convergence theorems. For this we derive some
results pertaining to integrals which depend upon parameters. Let I be a closed subinterval of R and S
a metric space (substitute R or Rn for S if necessary). If f : S I R, we write
f (s, ) for the function t f (s, t) when s S
and
f (, t) for the function s f (s, t) when t T

If f (s, ) is integrable over I for each s S, then F (s) = I f (s, t) dt denes a function F : S R.
Functions which depend upon parameters in this way arise in many areas of mathematics, and it
is often important to ask if properties of the function f (, t) are inherited by the function F. We rst
consider the property of continuity.
Theorem 4.12. If f (, t) is continuous at s0 S for each t I and f (s, ) is integrable over I for
each s S and there is an integrable function g : I R such that |f (s, t)| g(t) for all t I, then
F (s) = I f (s, t)dt is continuous at s0 .
Proof. Let sk s0 in S. Then f (sk , t) f (s0 , t) and |f (sk , t)| g(t) for every t I. The DCT implies
that F (sk ) F (s0 ) so F is continuous at s0 .
For our next application of DCT, we require a version of Leibniz Rule for dierentiating under the
integral sign. Note that when we say integrable, it means that it is HK integrable. There is no dierence
in the statement between the Leibniz Rule in Lebegue and Henstock. The only things that matter is
that, in Henstock, FTC holds in full generality, and no improper integral anymore.
Theorem 4.13 (Leibniz Rule). Let S = [a, b] and let I be a closed interval. Suppose that
1.

f
s

= D1 f exist for all s S and t I;

2. each function f (s, ) is integrable over I and there is an integrable function g : I R such that


f (s, t)


s = |D1 f (s, t)| g(t),
for all s S and t I.

Then, F (s) = I f (s, t) dt is differentiable on S and


F (s) =

f (s, t)
dt =
s

D1 f (s, t)dt

(41)

Proof. Fix s0 S and let {sk } be a sequence in S converging to s0 with sk = s0 . For t I,


lim
k

f (sk , t) f (s0 , t)
f
= D1 f (s0 , t) =
(s0 , t)
sk s0
s

and the function {hk } by


hk (t) =

f (sk , t) f (s0 , t)
sk s0

is integrable over I. By the Mean Value Theorem, for each pair (k, t) there is a z(k,t) lying between sk
and s0 so that
f (sk , t) f (s0 , t)
= D1 f (z(k,t) , t)
sk s0
which implies that



f (sk , t) f (s0 , t)

g(t).
|hk (t)| =

sk s0

14

The Dominated Convergence Theorem implies that


F (sk ) F (s0 )
sk s0

f (sk , t) f (s0 , t)
= lim
dt
k I
sk s0

= D1 f (s0 , t)dt,

F (s0 ) = lim

that is, F (s0 ) =

D1 f (s0 , t)dt as desired.

If I is a bounded interval and D1 f is continuous over S I, then D1 f is bounded and the function g
in Theorem 4.13 can be taken to be a constant. We use this observation in the next example.

2
Example 4.4. We show A = 0 et dt = 2 . For this we introduce

F (s) =
0

ex(1+t )
dt.
(1 + t2 )

Note that F (0) = arctan 1 = /4, and if x > 0


1
ex
0 F (x)
0

dt = ex .
(1 + t2 )
4

so
F () = lim F (x) = 0.
x

Fix x. We have, by Leibniz Rule,


( x )
( x ) x
1

e
e

x
u2
xt2
F (x) = e
e du =
g( x),
e
dt =
x
x
0
0
z u2
where g(z) = 0 e
du. Integrating F from 0 to gives
( x )

=
g( x)dx
4
x
0

2
= 2
ez g(z)dz
0
= 2
g (z)g(z)dz

(42)

= g()2
= A2 .
Hence, A =

2 .

Justification of Conservation law term

In this section, we shall go back to the topic in conservation laws. Specically, we want to justify the
term conservation law. So we recall the following conservation law:

x2
d x2
(x, t)dx = q(x1 , t) q(x2 , t) +
(x, t)dx.
(43)
dt x1
x1
which is the variation in time of some macroscopic quantity given by (x, t), and which has the following
partial dierential equation in divergence form:
t + qx =
15

(44)

where , q and are regular enough.


Suppose we are observing a physical phenomena where a matter is situated on a bounded set where
no ux coming in and out of the this set. For simplicity, we also suppose that there is no sink or source,
that is, = 0. See gure below.

Figure 2: Matter on a bounded set


Thus, for x , q(x, t) 0 similary, x , q(x, t) 0, for all time t 0. This means that

d
dt

z
(x, t)dx = q(, t) q(, t) +
| {z } | {z }
0

Therefore,
d
dt

}|

(x, t)dx

(x, t)dx = 0

This means that

(x, t)dx = C
R

for some constant C, for all time t 0. This means that mass is constant for all time, that is, mass is
conserve.

5.1

Some notes regarding the conservation law equation in divergence form

We remark that equation (44) a pde in divergence form is a 1-equation in 1-variable if q = (, x, t).

Now, if q = ()
where space varible x and time variable t are not explicit, it would still imply and

x, t) since is a function of x and t.


understood as q = ()
= (,
Moreover, if
q = (x , , x, t)
or
q = (T, x , , x, t)
where T is a xed but arbitrary temperature, then we could write equation (44) as
t + ()x = .

(45)

Now, suppose that q = (, x, t). Then by chain rule, we can write


((, x, t))x =


+
1 + 0 = x + x .
x
x

Thus, we can write equation (44) as


t + x + x =
16

(46)

equivalently
t + c(, x, t)x = g(, x, t)

(47)

where
c(, x, t) = (, x, t)
which can be thought of signal speed of ow and
g(, x, t) = (, x, t) x (, x, t).

(48)

Now, this time we note that ow of speed


v(x, t) =

q(x, t)
(x, t)

is not equal or similar to the signal speed of ow.


So, if in equation (47) g 0, then
t + c(, x, t)x = 0.

(49)

Also, since g 0 we can have


(, x, t) = x (, x, t).
What does this means physically (source/sink= change in ux wrt x)?
Suppose in (48), there is no source/sink. Then 0 g x (, x, t). So we can write (47) as
t + cx = 0

(50)

where
c = (, x, t),
note that c is dependent on spatial variable x. The presence of x in equation (50) has some relevance.
This implies that equation (50) can be of the form
t + h(, x, t)x = 0

for some h(, x, t) regular enough

(51)

for some h(x, t) regular enough

(52)

which is a quasi-linear pde or


t + h(x, t)x = 0

which is a linear pde. These pdes are not conservation laws (Am I getting this right?). We cannot prove
that

d
(x, t)dx = 0.
dt

5.2

Dispersive-Diffusive Conservation law

Here we go back to the previous subsection where we assume that


q = (x , , x, t)
and as a consequence, we could write equation (44) as
t + ()x = .

(53)

By chain rule
1

z}|{
()x = ((x , , x, t))x = xx + x + x xx .
Then
t + cx = g
17

(54)

where
c =

and g = xx x

or equivalently
t + cx = g axx

(55)

where
c = ,

g = x ,

and a = .

Equation (55) is a parabolic conservation law.


Suppose
q = (x , , x, t) := (, x, t) x .
Then
qx = [(, x, t)]x [x ]x = x + x xx .
Then equation (44) becomes
t + [ x + x xx ] = .

(56)

t + x = x + xx .

(57)

t + cx = g + xx .

(58)

that is,
which we can write as
where
c = ,

and g = x .

If g 0, then equation (58) is a nonlinear diusion equation. Its a conservation law but not hyperbolic.
Moreover, if we have the following pde:
t + cx = g + xx + xxx .

(59)

Then equation (59) is a dissipative-dispersive equation, where


1. second term cx is transport
2. xx is a dissipation term also called diusive term
3. xxx dispersive term or oscillation.

Traffic flow

In this section, we shall begin with an intuition on the units of the quantities we are studying.
To understand the unit of quantities we are studying, we will give an explanation below. See table 1
for summary and intuition.
1. Mass as the integral of density on space:
To understand, consider a rod whose density is equal to 3, and we have to integrate over interval
[3, 19]. Hence we have
19
1
3dx = 3x 3 9 = 3(19 3) = 3 length = mass.
3

Note that length of a rod can be regarded as a 1-dimensional volume7 . Hence, on the unit, we are
multiplying the unit of the density, mass V ol1 by another unit, V ol. Thus, whats left out is the
unit of mass. So that means that when we integrate over the spatial variable say x, in this case, the
length or volume, we are adding 1 to the power of V ol1 , just like the usual when we integrate
a function. In this case, (V ol)1+1 = (V ol)0 = 1.
7 Similary,

area is a 2-dimensional volume and so the 3-dimensional volume is the Volume that we know.

18

Table 1: Summary of Quantities and thier units


Quantity Name
Symbol
Unit
density
mass
density change wrt space x
density change wrt time t
Variation in time of the mass
Flux
??
Source/sink

dx
x
t

dx
t
q
dx

mass V ol
mass
mass V ol2
mass V ol1 time1
mass time1
mass time1
mass time1
mass V ol1 time1

Sample Unit
kg m1
kg
kg m2
kg m1 hr1
kg hr1
kg hr1
kg hr1
kg hr1

2. Change of density in space, x :


Note that by denition,
mass V ol1
(a + x, t) (a, t)
=
,
x0
x
V ol

x (a, t) = lim

since the numerator is subtraction of densities, so it has a density unit, while the numerator is a
length, hence it has a unit of volume by denition. Therefore, the unit of x is mass V ol2 .

3. To get the unit of ux q and integral of source or sink dx, we can observe that

x2
dx = [q]x1 + dx.
t

This means that whatever the unit of t


dx must also be the unit of the ux and the integral of
source or sink. The unit is mass time1 .
Therefore, it will now be easy to see that the unit for source or sink is mass time1 V ol1 .
4. Suppose q(x, t) = (, x, t). Then the unit for is
(
)1
mass time1 mass V ol1
= V ol time1 .
Hence, as the variation of ux in density can therefore be regarded as speed or velocity.
Also consider the ow speed:
v(t, x) =

q(x, t)
mass time1

= V ol/time
(x, t)
mass V ol1

which is speed.

6.1

Settings: Assumptions

In our trac ow example, we shall made the following assumptions.


For simplicity, we suppose a one-lane road. This will serve our x-axis with the positive orientation
going to the right direction.
We assume that on this one-lane road, there were no on/o ramps, i.e., = 0. It means that one
car cannot go in or out of the lane.
We assume that X has a unit distance (dist) kilometer, km. Recall that distance or length is
synonymous with volume.
19

We assume T has a time unit of hour, hr.


For trac density , a unit is: n cars/km.
Now, consider the ux as a function of density (X, T ) which is interpreted as how many numbers of
cars in a certain distance, say, kilometers (#cars/km), ux which is also a function of X, this can be
interpreted as curves, ups or downs of the road. Flux is also a function of T , this can be interpreted as
time. So when you drive your car, it does matter when you drive during the night or day or your speed
could be limited because of the topography of the road (curvy, zigzag, etc) or even in trac, that is the
density ( no. of cars in a distance).
Thus, in reality ux can expressed as
q(X, T ) = ((X, T ), X, T ).
However, for simplicity, we forget about the intricacy of our ux system and we consider that we may
be driving on a straight road during a sunny day free from trac whatsoever. Here, we can expressed
ux by an empirical function:
q(X, T ) = () = R ( max ) ,
where R > 0. So the ux function is not explicit of X and T .
Now, ux could also be of the following form, which is an reformulation of the above equation:
(X , ) = () kX = R ( max ) kX ,
where k > 0.
We can combine these two formulation of ux into one function and we have
{
() = R ( max )
R>0
q(X, T ) =
(X , ) = () kX k > 0

(60)

which can be viewed graphically roughly as a parabola.


As to the question on how and why we arrive such formulation(?) we will defer for it for the moment
and pick it up later in our discussion.

Figure 3: Trac ux
In light with Figure 3: when = 0, it means theres no car, hence no ux, i.e., q = 0. Also, if
= max , that is, too much cars in a given distance, then theres trac that cars dont move, hence no
ux, i.e., q = 0.

6.2

Into a dimensionless PDE

Now, our aim in this section is to transform the conservation law in divergence form below into a dimensionless pde:
T + qX = 0.

20

(61)

By plugging (X , ) into (61), we have


T + (R ( max ) kX )X = 0.

(62)

To understand whats going on in (62), we will inspect the unit of each quantities involved in (62). Note
that the unit of T is n cars dist1 time1 . Thus, the second term should be the same unit as T in
order to compatibly add them.
Now, to know the unit of R, bearing in mind that we are multiplying density to a density and
afterwards applying derivative wrt X, hence R unit would make sense if
n cars1 dist2 time1 .
To know the unit of k, remember that X has unit n cars dist2 , so after applying derivative wrt X,
the unit becomes n cars dist3 . Thus, in order for the unit of the third term kXX to be compatible
with the unit of T , the unit of k should be
dist2 time1 .
6.2.1

Characteristic scale

One important procedure while formulating a mathematical model of a physical situation is scaling.
Roughly speaking, scaling deals with choosing new, usually dimensionless variables and reformulate the
problem expressed in these variables. This is what we were going to do with (62)
So, we begin to normalize variables x, t and u by introducing a characteristic length8 denoted by L0 .
We have the following

x=
t=
u=

X
L0
T
L0 / (Rmax )
max
e
2

(63)

max
2

where e(x, t) = (X, T ), X and T are given by scaling above, i.e., X = xL0 and T = t (L0 / (Rmax )).
Therefore, u = u(x, t) that is, u is a function now of x and t.
It would be useful to know that given X = xL0 we have
Xx = L0 xx = L0 1 = L0 .
(
and given T = t

L0
Rmax

(64)

)
we obtain
(
Tt =

L0
Rmax

)
tt =

L0
Rmax

since tt = 1

(65)

Also, given
e(x, t) = (X, T )

(66)

we have
T
et (x, t) =
= T Tt = T
T t

L0
Rmax

)
(67)

8 There is no unique definition for characteristic length. One can define the length scale as the square root of the surface
area, as the third root of the volume, as the volume-to-surface ratio, etc. It all depends what length scale you want identify.
For our problem (traffic flow on a straight road) the characteristic length is the length of a road (from X = 0 to X = L)
and we denote by L0 . Is this right, Joaquim?

21

and

X
= X Xx = X L0 .
(68)
X x
So to expressed (61) or (62) into dimensionless equation, we will solve what is ut , ux , and uxx because
these quantities involves T , X , and XX as we will see.
Take note that u can be written as
ex (x, t) =

u=1

e
max
2

(69)

Then by chain rule and equation (65), the variation in time of u is


ut =

1
max /2

T Tt =

2L0
L0
T =
T .
max Rmax
R (2max )

(70)

The variation in space of u


ux =

2
2
2L0
X Xx =
X L0 =
X .
max
max
max

The variation in space of ux (or the diusion)


[
]
2L0
X (X, T )
uxx =
max
x
2L0
=
[X (X, T )]x
max
2L0
=
(X )X Xx xx
max
2L0
=
XX L0 1
max
Hence,
uxx =

2L20
XX .
max

(71)

(72)

(73)

Now, we can write


T + (() kX )X = 0

where () = R ( max )

into
T + ()X kXX = 0,

(74)

and by using our dimensionless quantity (70) to (73), we can transform (74) into
R2max
e x 1 + kmax uxx = 0
ut +
2L0
L0
2L20

(75)

e t) = (X(x), T (t))
(x,

(76)

where

from whence,

e x = X Xx = X L0

which means that


X =
(we substitute it in the second term of (74)).

22

ex

,
L0

Now, we know that


e t) = (X(x), T (t)) = R(X(x), T (t)) ((X(x), T (t)) max ) ,
(x,
and by our assumption in (66), we can write it as
e t) = Re
(x,
(x, t) (e
(x, t) max ) .

(77)

e x in (75), by substituting
We are halfway already to solving
e =

max
(1 u)
2

solved in (69) to be plug into (77).


Therefore,
(
)
e t) = R max (1 u) max (1 u) max
(x,
2
( 2
)
max
max
=R
(1 u)
(1 + u)
2
2
2
= R max (1 u)(1 + u)
4

(78)

Therefore,
2

e t) = R max (1 u2 ),
(x,
4
and by taking the variation in space, it follows that

(79)

2
e x (x, t) = Rmax uux .

(80)

R2max
R2max
1
kmax
uxx = 0.
ut
uux
+
2L0
2
L0
2L20

(81)

So, we can now write (75) into

Multiplying (81) by 2L0 , we have


R2max ut + R2max uux

kmax
uxx = 0,
L0

(82)

kmax
uxx .
L0

(83)

and equivalently
R2max ut + R2max uux =
Therefore, our dimensionless equation
ut + uux =
which nally we can write as

(
ut +

u2
2

k
uxx ,
L0 Rmax

(84)

)
= uxx ,

(85)

where = L0 Rk max is known as the Burgers equation.


If k = 0, then (86) becomes
( 2)
u
=0
ut +
2 x
which we know as the inviscid Burgers equation.
23

(86)

6.2.2

Remarks on inviscid Burgers and the Burgers Equation

References
[Flanders]

Harley Flanders. Dierentiation under the Integral Sign. American Mathematical Monthly,
vol. 80 (June-July 1973), p. 615-627.

[Folland]

Gerald B. Folland. Real Analysis: Modern Techniques and Their Applications, second ed.
Wiley-Interscience, 1999.

[Conrad]

K. Conrad. Differentiating under integral sign, http://www.math.uconn.edu/kconrad/blurbs


/analysis/diffunderint.pdf

[Goel]

S. K. Goel and A. J. Zajta. Parametric Integration Techniques. Math. Mag. 62 (1989),


318322.

[Charles]

Swartz, Charles. Introduction to gauge integrals. Singapore: World Scientic, 2001.

[Swartz]

Swartz et al. Theories of integration: The integrals of Riemann, Lebesgue, Henstock-Kurzweil


and Mcshane. Second edition, World Scientic.

[McCormmach] Jungnickel, C., & McCormmach, R. (1990). Intellectual Mastery of Nature. Theoretical
Physics from Ohm to Einstein, Volume 2: The Now Mighty Theoretical Physics, 1870 to
1925 (Vol. 2). University of Chicago Press.
[Talvila2]

E. Talvila. Some Divergent Trigonometric Integrals, Amer. Math. Monthly 108 (2001), 432
436.

[Talvila]

Erik Talvila. Necessary and Sufficient Conditions for Differentiating Under the Integral
Signhttp://www.math.ualberta.ca/etalvila/papers/dinal.pdf. American Mathematical
Monthly, vol. 108 (June-July 2001), p. 544-548.

The author of this entry has also written an exposition, Differentiation under the Integral Sign using
Weak Derivativeshttp://gold-saucer.afraid.org/math/di-int/di-int.pdf, containing a proof of Theorem
4 along with detailed computational examples.

24

You might also like