Hyperreal Riemann Integration from the Bottom Up

John Starrett
Department of Mathematics, New Mexico Institute of Mining and Technology

Kevin Lenth
Department of Computer Science, Casper College

The standard Riemann integral can be recast in the language of non-standard calculus as an
infinite sum of areas of rectangles or trapezoids with infinitesimal base width. In the study of non-
standard analysis (NSA) the usual goal is to prove things about the standard universe more simply
by using non-standard methods. However, there is a rich and interesting world of mathematics on
hyperreal numbers that need not refer to the standard universe at all. Here we demonstrate some
techniques of Riemann-like integration over infinitesimal and infinite domains. While the goal is not
to prove or calculate standard things, some of these results do have standard interpretations, which
we highlight.

I. INTRODUCTION

A. Tools

In this paper we work freely with infinite and infinitesimal numbers from the hyperreal number system. We
have included an appendix containing a basic introduction to the hyperreals from the ultraproduct/superstructure
approach. Basically, the hyperreal numbers R are a conservative field extension of the real numbers R, containing
infinite, finite and infinitesimal numbers, which can be algebraically manipulated in the same way as real numbers.

B. Hyperreal Numbers

Robinson[10] showed that an enlargement R of the real number system R, containing the reals as a subfield, could
be constructed using model theory, and subsequently, with Zakon[11], showed that the hyperreals and the universe in
which they live could be constructed in a set theoretical manner by building a superstructure over an ultraproduct.
This method of building the hyperreals has become the most well known, although Nelsons internal set theory[9],
Hatchers algebraic approach[5] and Benci and DiNassos alpha theory[2] have adherents.

C. 2-Hyperreal Numbers and Integration

Once we have constructed R in the standard way, as countable sequences of real numbers mod equivalence on
a free ultrafilter, it is easy to see how to enlarge R to a field R containing R as a subfield. Our new numbers
are hypersequences (sequences indexed by hypernatural numbers) of elements of R filtered by a co-countable free
ultrafilter[8], an ultrafilter containing no countable sets.

II. INTEGRALS OVER R AND R IN A HYPERREAL SENSE

We take R as a subfield of R, use the symbol for some infinite hypernatural number,  for infinitesimal numbers,
Rb
and by mean infinitesimally close. We denote by I = f (x) dx and Iab = a f (x) dx the indefinite and definite
R

Riemann integrals, respectively, defined using the hyperreal numbers in the usual way: Let f be a real valued
function and fix some infinite hypernatural . Divide the interval [a, b] into subintervals of equal infinitesimal
 2
P
width dx = [b a]/. Assuming Riemann integrability, it can be shown that I k=1 f (xk )dx (where xk is drawn
arbitrarily from the kth subinterval).
We define integration of hyperreal functions over infinitesimal domains by dividing the interval [a, b](where a and
b are now hyperreal numbers, with b a infinitesimal) into pieces of second-order infinitesimal width, summing up
P
(first-order infinitely many) terms. We take as a definition I f = k=1 f (x)dx, where now dx = ba and xk is drawn
arbitrarily from the kth subinterval.
We may similarly define integrals over infinite domains of R by partitioning [a, b] into an infinite number of pieces
of hyperinfinitesimal width and adding areas of rectangles or trapezoids with these base widths.

III. INFINITESIMAL INTEGRATION IN ONE DIMENSIONAL SPACES

As an example, we integrate a few simple functions along the unit interval using a technique that extends to
integrating over a symmetric middle p/q Cantor set C.
C is constructed by scaling the interval C0 = [0, 1] by 0 < d < 1/2, taking its union with a copy of itself shifted
right by s = 1 d to obtain the generator C1 , then applying this dilation, shift and union operation to C1 to generate
C2 , and so forth. In the same manner, construct the unit interval [0, 1] by taking d = 12 and s = 12 and iterating as
with Ci . Continue this process times, producing [0, 1] as the union of 2 subintervals of length 2 .
Let aj be the left endpoint of the j th segment in this construction. Then the integral of x over the j th interval of
length 2 is
Z aj +2
x
x dx = 2 aj + 21 .


Ij =
aj

The integral over [0, 1] is the infinite sum over all intervals [ak , ak + 2 ]

Z 1 2
X 2
X

Ijx aj + 21 .


x dx = =2 (1)
0 j=1 j=1

a P2k
The partial sums Sk j = j=1 aj of the aj , when taken in groups of 1, 2, 4, ... 2k , satisfy the recursion
k
2
aj a a
X
aj + 2k 2 = 2Sk j + 22k 2 ,


Sk+1 = Sk j +
j=1

whose solution is
a 1 2k
2 2k


Sk j = 2 (2)
2
when a0 = 0. Substituting (2) into (1) and letting k = , we find
Z 1
1
x dx = ,
0 2
as we would expect.
a2 a P2k k 2 a2j k+1 aj
2 Sk + 23k 22
j


In a similar fashion, by solving the recursion Sk+1 = Sk j + j=1 aj + 2 2 = 2Sk +2
a2
for the sum of squares of left end points to get Sk j = 22 1 k 1 k 1 k

2
3 8 2 4 + 6 2 , we can integrate x over the same
interval to obtain 13 .

A. Sets with Holes

To integrate over a set consisting of alternating infinitesimal intervals and holes, let J1 = [0, 2 ] and J2 =
J1 (J1 + 21 2 ), J3 = (J2 J2 + 22 2 ), ... Jk = Jk1 (Jk1 + 2k1 2 ). The sum of left end points aj of these
 3

intervals obeys the recursion

k
2
aj a a
X
Sk+1 = Sk j + (aj + 2k+1 2 ) = 2Sk j + 22k+1 2 ,
j=1

so the sum of the left end points is given by

a
Sk j = 2 22k 2k


(3)
a
when S0 j = 0. Now k ranges only to 1 so the integral of x over this set is
k 1
2 2X
X
1
Ijx k=1 aj + 21 = 22 222 21 + 221 21 = 1 2 .




=2
j=1 j=1
4

The standard part is 1/4, what we would intuitively expect, but in addition there is an infinitesimal part 22 .
The reason is not difficult to see: we summed over the odd set of indices j, but on each of the even-indexed intervals,
the integral exceeds that of its left partner by 21 .

FIG. 1: The sum of integrals of x over left alternating infinitesimal intervals in black, the complement in gray, and the deficit
in the left sum in dark gray.

In general, the integral of x over a set of uniformly spaced intervals of width 2 and period n spaces, beginning
1


at 0 in [0, 1], is 2n 1
12 2n 2 . There are similar formulas for other powers and functions of x.

FIG. 2: An increasing sequence of intervals and spaces.

Now consider the set consisting of alternating intervals and spaces where the length of each interval and following
space increase by one from one to the next. Number the left end points an as indicated in figure (III A). Then we
have an = n(n 1) and the integral of x2 is
Z an +n  
1 1
x2 dx = a2n n + an n2 2 + n3 3 = n3 3 n2 n + . (4)
an 3 3

If we continue the construction steps to complete the interval [0, 1], then  = 12 , and with the recur-
x2 2 2
= Skx + 3 k 5 + 4k 4 + 19 1
3 2


sion Sk+1 3 k + 5k + 2k + 3 for the sum through interval k + 1, the solution Skx =
 4

FIG. 3: The hyperreal Cantor set is not a dust at infinite magnification, but a set of infinitesimal length line segments. The
spacing between the kth successive sets of intervals is 2 3k .

P 2
3 1
3 5
3 16 k 6 + 10 k 16 k 3 + 30
1
k with k = is j=1 Ijx = 16 + 10 16 3 + 301 5


. The standard part is what we
would expect for a polynomial, as the intervals and spaces are evenly distributed. On the other hand, the values of
R 1
x dx when taken over the set and its compliment are quite different as the first is undefined and the second infinite.

IV. INTEGRATING ON THE HYPERREAL CANTOR SET

The hyperreal Cantor middle thirds set C consists of 2 closed line segments of length 3 . At the infinitesimal
level, we construct the set C by shifting C0 = [0, 3 ] to the right by 2 30 to obtain the generator C1 =


C0 C0 + 2 30 , then iterating on

Ck+1 = Ck Ck + 2 3k .


(5)

The integral of a polynomial over this set requires evaluation of powers of left end points aj , and recursions for the
a2j 2 P2k
2
sums of these points can be obtained from (5); for instance, Sk+1 = S aj + j=1 aj + 2 3k .
The first few sums of powers of aj have recursions

a1 a1
j
Sk+1 =2Sk j + 2k (2 3k )1 ,
a2 a2 a1
j
Sk+1 =2Sk j + 2(2 3k )Sk j + 2k (2 3k )2 ,
a3 a3 a2 a1
j
Sk+1 =2Sk j + 3(2 3k )Sk j + 3(2 3k )2 Sk j + 2k (2 3k )3 ,
a4 a4 a3 a2 a1
j
Sk+1 =2Sk j + 4(2 3k )Sk j + 6(2 3k )2 Sk j + 4(2 3k )3 Sk j + 2k (2 3k )4
..
.

with general recursion

n1
an an X
n

i ani
n
j
Sk+1 = 2Sk j + Bi+1 2 3k Sk j + 2k 2 3k , (6)
i=1

n
where Bm is the mth binomial coefficient for binomials of degree n. The formulas for the first few sums of powers are
 
a1 1 k 1 k
Sk j = 3 6 2 ,
2 2
 
a 2
3 k 1 k 1 k
Sk j = 3 18 6 + 2 ,
8 2 8
 
a 3
5 k 9 3 1
Sk j = 3 54 18k + 6k + 2k ,
16 16 16 16
 
a 4
87 5 9 1 17 k
Sk j = 3 162k 54k + 18k + 6k 2 .
320 8 32 8 320
 5

For the hyperreal Cantor middle thirds set, the first few integrals of powers of x are
Z  
1 2
xdx = ,
C 2 3
Z    
3 2 1 2
x2 dx = ,
C 8 3 24 27
Z    
5 2 1 2
x3 dx = ,
C 16 3 16 27
Z      
87 2 3 2 7 2
x4 dx = .
C 320 3 32 27 320 243

The total length of the hyperreal Cantor set is (2/3) . Notice that multiplication by the inverse length of the set
2


3 and taking the standard part, indicated by a bar over the integral sign, gives us
Z Z Z Z
1 23 3 5 87
xdx = x dx = x dx = x4 dx = ,
C 2 C 8 C 16 C 320

which is what the calculations in Slomczynski et.al. ([12]) give for the standard Cantor middle thirds set.
We call this renormalized volume the h-content of f over the Cantor set. The h-content is a generalized p-volume
for non-integer p. The hyperreal scaling of a strictly self-similar fractal provides a scale of the renormalization, and
is the product of the (hypernatural) number of pieces in the object times the n-volume of the pieces. Non-fractal
objects have a hyperreal scaling of 1: for instance, it takes 81 cubes of volume 81 to make a unit cube. Therefore
the content of a non-fractal object is its length, area, volume, or in general its n-volume.
Finally, let us integrate two common functions that  do not have simple integrals over this set. The integral of
x
ex
R x aj 3
the exponential C e dx over interval j is e e 1 , and with S0e = e3 1, the recursion relation Sk+1 =
   
ex ex 23k 3 23k
Q1
Sk + Sk e gives S = e 1 k=0 1 + e 1.753279205 evaluated at k = 1.
Integration of the sine or cosine functions requires the solution of a system
" # " #" #
s
Sk+1 1 + cos 2 3k sin 2 3k Sks
c
= .
Sk+1 sin 2 3k 1 + cos 2 3k Skc

The solution is
" # 1
" #" #
Sks Y 1 + cos 2 3k sin 2 3k sin 3
=
Skc k=0
sin 2 3k 1 + cos 2 3k 1 cos 3

V. TWO DIMENSIONS: THE SIERPINSKI GASKET

The right Sierpinski gasket can be built recursively starting with a generator composed of three right triangles, as
in figure (4). Let T(a,b) be a triangle with vertices (a, b), (a + 2 , b), (a, b + 2 ) and parameterization

T(a,b) = (1 t)ha, b + 2 i + t((1 s)ha, bi + sha + 2 , bi) = ha + st2 , b + (1 t)2 i.

R1R1
An integral over T(a,b) of some function f = f (s, t) is 0 0 f dA where dA = 22 tdt ds. Because the limits of the
parameters are identical on each triangle of the gasket, we see that an integral over any particular triangle of a fixed
function f depends only on the point (a, b).
Let Vn be the set of right vertices at the nth level of construction. Then
X  X
Vn,j + (0, 2n ) Vn,j + (2n , 0) .

Vn+1 = Vn (7)
j j
 6

FIG. 4: The generator S1 of the Sierpinski triangle on the left, and a finite level of construction S4 on the right.

The integral over the entire gasket is the limit of the sequence of integrals over individual triangles parameterized by
pairs (a, b)j . Let us calculate the integral of x + y, where x = aj + st2 and y = bj + (1 t)2 , over the gasket.
The integral over each triangle is
Z 1Z 1
1 1
[aj + st2 + bj + (1 t)2 ]22 t dt ds = 23 + 22 (aj + bj ).
0 0 3 2

The integral over the k th stage of construction of the gasket is

k
3  
X 1 3 1 2
Ik = 2 + 2 (aj + bj ) . (8)
j=1
3 2

a ,bj P3k
The sums Sk j = j=0 (aj , bj ) of right vertex points (aj , bj ) obey the recursion

(a ,bj ) (aj ,bj )

j
Sk+1 = 3Sk + 3k 2k 2 ,

so we have
(aj ,bj ) 1 k k

Sk = 3 2 2 1 . (9)
3
Substituting (9) into (8), we obtain
k
3    
X 1 1 2 k k
23 + 22


Ik = 3 2 2 1
j=1
3 2 3

and with k = we get

 
1 1 1 3
I = 3 23 + 3 23 (2 1) = .
3 3 3 4
3


Upon renormalization by the volume fraction 4 and taking the standard part, we obtain the content
Z
1
x + y dA = .
S 3
 7

The integral of x2 over a triangle T(a,b) is 21 a2j 4 + 13 aj 8 + 12

1
16 , so we need a formula for the sum of squares
2 2
(a ,b ) 3k
= j=1 (a2j , b2j ) and (keeping only the aj ) it obeys the
P
of right vertices. The sum of squares of these points is Sk
recursion
k k
3 3
a2 2 X X
Sk+1 = Ska + (aj + 2 2k )2 + a2j (10)
j=1 j=1

from which we compute

 
a2 5 k 6 1
Sk j = 4 12 6k + 3k . (11)
27 27 27
The integral of x2 over the gasket is therefore

3  
x2
X 1 1 1
I = a2j 4 + aj 8 + 16 , (12)
j=1
2 3 12

5 3 5 3 1 1




which, upon substitution of the formulas for the sums of a2j and aj yields 54 4 54 16 + 12 16 . Renormal-
3 5


ization by 4 shows the result has standard part 54 . By similar means we can show the integral of xy over the
1
gasket has renormalized standard part 27 .
The general recursion formula for sums of powers of x or y over the gasket is
n1
an , bn an , bn X
j
Sk+1 j
= 3Sk j j
+ n
Bi+1 2i k Skni + 3k 2n k (13)
i=1
n
where Bm are the binomial coefficients as before.
am bn P3k
For the most general multivariate polynomial term Sk j j
= 2(m+n) j=1 am n
j bj we have the recursion

k
3
am n
j bj am bn (m+n)
X
(aj + 2k )m bnj + am k n


Sk+1 = Sk j j +2 j (bj + 2 ) ,
j=1

ap bqj am bn
which can be resolved into a weighted sum of terms Sk j of equal or lower degree and solved for Sk j j
as in the
other examples.

VI. THE THREE DIMENSIONAL CASE: THE MENGER SPONGE

As before, the integral over any cube will depend only on the coordinates of a vertex. Place the first unit cube
with an edge at the origin. At each stage of construction, the Menger sponge consists of 20 copies of the sponges at
the previous stage, 8 placed at the vertices of a cube of three times the width of the previous, and the remaining 12
placed to connect the vertices along the edges (see figure(VI)).
With generic coordinates of (aj , bj , cj ) = aij for a cubes vertex nearest the origin, the integral of x over one cube
ai P20k
with side length 3 is 33 aj + 21 3 . The sum of coordinates at any stage will be Sk j = j=1 aij . The sum

obeys the recursion

k
20
aij ai X
7aij + 4 aij + 3k + 8 aij + 2(3k )




Sk+1 =Sk j +
j=1
ai
=20Sk j + 20k+1 3k , (14)

so the sum over the entire sponge is

ai 1
60k 20k .


Sk j = 3 (15)
2
 8

The integral of x over the entire sponge is therefore

k
20  
X 1 ai 1
33 aij + 3 = 33 Sk j + 20k 34 . (16)
j=1
2 2

Substituting (15) into (16), we get

1 1 1
Ik = 33 3 (60k 20k ) + 20k 34 = 34 60k . (17)
2 2 2

20
When k = we obtain I = 21 34 60 = 12 20


27 . Upon division by the volume 27 , we get the normalized integral
1
2 .

VII. LINE INTEGRAL EXAMPLE: THE KOCH CURVE

For the Koch curve, the sums recursions are more complicated because the construction is not simply scale and
shift, but scale, shift and rotate. Therefore, the parameterization of each piece of the generator relies not just
on one distinguished point of the piece, but upon two, the initial and terminal point of a line segment. For the
standard Koch curve whose generator has endpoints (0, 0) and (3 , 0), the sum of functions over the entire curve is
 9

FIG. 5: Four stages in the construction of the infinitesimal von Koch curve.

P4+1
Skf = j=1 f (aj , bj ), and the recursion giving the sums of functions over successive generations is

4k !
f
X 1 3 1 k 3 1
Sk+1 =Skf + f aj bj + 3 , aj + bj
j=1
2 2 3 2 2
4 k !
X 1 3 1 k 3 1 3 k
+ f aj + bj + 3 , aj + bj + 3
j=1
2 2 2 2 2 2
4k  
X 2
+ f aj + 3k , bj . (18)
j=1
3

We will compute integrals of x, y, xy, x2 and y 2 over the Koch curve. Assume a parameterization K(s) by s so
that for the complete curve K(0) = (0, 0) and K(1) = (1, 0), and let the coordinates of the initial point of the j th
interval be denoted by aj = (aj , bj ) and that of the terminal point be aj = (aj , bj ). Then the integrals are
Z 1
1
t aj + (1 t)aj 31 dt = 3 (aj + aj ) (19)
0 6
Z 1
1
t bj + (1 t)bj 31 dt = 3 (bj + bj )
0 6
Z 1
1
(t aj + (1 t)aj )(t bj + (1 t)bj ) 31 dt = 3 (2aj bj + aj bj + aj bj + 2aj bj )
0 18
Z 1
1
(t aj + (1 t)aj )2 31 dt = 3 (a2j + aj aj + a2j )
0 9
Z 1
2 1
t bj + (1 t)bj 31 dt = 3 (b2j + bj bj + b2j ).
0 9

To sum these integrals over the entire fractal, we need the sums of all the terms on the right hand sides of (19).
Recursions can be calculated using (18), and these recursions have solutions
 10

a 1 1
Sk = 3 3k + 12k Ska = 3 3k + 12k



6
6

b 3 3
12k 3k Skb = 12k 3k



Sk = 3 3
54   54  
a2 1 k 31 k 19 k a 2 1 k 31 k 19 k
Sk = 9 1 + 4 9 + 36 Sk = 9 1 + 4 9 + 36
90 6 6 15 34 34
   
b2 1 k 1 k 1 k b2 1 k 1 k 1 k
Sk = 9 1 + 4 9 + 36 Sk = 9 1 + 4 9 + 36
90 6 6 15 34 34
   
aa 1 k 1 k 19 k bb 1 k 1 k 1 k
Sk = 9 1 4 9 + 36 Sk = 9 1 4 9 + 36
60 9 9 60 9 9
   
ab 3 1 k 7 k 1 k ab 3 1 k 1 k 1 k
Sk = 9 4 9 + 36 Sk = 9 4 + 9 + 36
18 10 45 18 9 20 45 36

ab 3 3
9 9k + 36k Skab = 9 9k + 36k



Sk =
324 324
Inserting (20) into (19), we obtain

Z  
2 4
x ds =
K 3 3

Z    
2 3 4 3 1
y ds =
K 27 3 54 3

Z    
3 4 3 1
xy ds =
K 27 3 27 3

Z        
2 19 4 7 1 11 1 11 4
x ds = + +
K 45 3 60 3 1620 27 405 27

Z        
1 4 1 1 11 1 11 4
y 2 ds = + ,
K 45 3 180 3 1620 27 405 27
4


which, when renormalized by 3 , have standard parts

Z Z Z
2 2 3 3
x ds = y ds = xy ds =
K 3 K 27 K 27
Z Z
19 1
x2 ds = y 2 ds = .
K 45 K 45

VIII. SUMMARY

We have demonstrated a method for calculating integrals over strictly self-similar fractals which can be built by
translation, scaling and rotations. Polynomials have simple integrals, and were chosen because the method requires
functions that are linear in their arguments in order to find explicit recursion equations. One could build polynomial
approximations of arbitrary degree to analytic functions, and perhaps find patterns hinting at closed form solutions.
The method should generalize easily to complex polynomials.
 11

IX. APPENDIX

A. Superstructures

In order to define a mathematical universe in which we can do calculus, we build a set of subsets of numbers big
enough to contain any object we might need. Let S be a set of individuals, each of which is assumed to have no
internal structure of its own; that is, x S, @y x. We form an iterated structure

V0 (S) = S
Vn+1 (S) = Vn (S) P(Vn (S)),

where P() is the power set. We then define the superstructure V (S) over S to be
[
V (S) = Vn (S).
nN

For instance, if we choose the real numbers R as our base set S, we can easily build real analysis.[14]

B. R and its universe

Let S = R. Then

V0 = {x : x R}
V1 = {x : x R, P({x : x R})}
V2 = {x : x R, P({x : x R}), P({x : x R, P({x : x R})})}
..
..

Thus, V0 contains the real numbers, and because V1 contains all pairs of real numbers, V2 contains all ordered pairs
(x, y) {x, {x, y}}, x, y R, and V3 contains all sets of ordered pairs of real numbers, and therefore contains the
functions. Similarly, functionals can be shown to belong to V6 (S) as the set of all ordered pairs, the first element
of which is a function and the second a real number. Continuing on up the ladder, we get a huge collection of
set-theoretical structures, subsets of which correspond to any of the structures of classical analysis.

X. THE HYPERREALS R

In order to build a nonstandard universe with the standard universe embedded in it, we take our base set T to
consist of all mappings f : I S, where I is some index set. This set of mappings is called ultraproduct. For the
construction of the standard model of the hyperreals, we use a countably infinite index set, usually N. We then form
a superstructure as before

V0 (T ) = T (20)
Vn+1 (T ) = Vn (T ) P(Vn (T )) (21)
[
V (T ) = Vn (T ), (22)
nN

and get the same sort of hierarchy of structures. Usually we think of the set of mappings as being the set of (countably)
infinite sequences {hs1 , s2 , ..., sm , ...ii }, m N, i R of elements our base set S, where R is an uncountable index
set. This superstructure V (T ) now contains functions, functionals, and all the other structures of analysis, but with
infinite sequences of real numbers as arguments rather than real numbers.
 12

The superstructure V (S) has built into it the structure of a field if S does, because the true relations in V (S)
are those that respect the field structure of S. Likewise, the superstructure V (T ), properly interpreted[15], has the
structure of a field if S does because of the way S is mapped to T .

A. Basic properties of R and

R

Interpreting T as the set of all infinite sequences hr1 , r2 , . . .i = hri i, i N of real numbers, we may identify a
real number r with the constant infinite sequence r = hr, r, r, . . .i. In the standard model, r is called the standard
copy of r in R. In anticipation of our multiple extensions, we call r = hr, r, r, . . .i the first extension of r. If we
interpret algebraic operations, functions, etc. on these sequences as acting term by term, the set of numbers of the
form r = hr, r, r, . . .i are isomorphic to the usual real numbers, and so form a totally ordered field. Thus, we can
naturally write hr, r, r, ...i + hs, s, s, ...i = hr + s, r + s, r + s, ...i, f (hx, x, x, ...i) = hf (x), f (x), f (x), ...i and so on.
When we consider nonconstant sequences of reals hr1 , r2 , . . .i = hri i, such as h1, 2/3, , 0, ...i, h1, 1, 2, 3, ...i,
h3, 1, 4, 1, 5, 9, ...i, we find we have what seem to be candidates for infinite and infinitesimal numbers, such as =
h1, 2, 3, 4, ...i and  = 1/ = h 11 , 12 , 31 , 14 , ...i. We have given these two archetypes the names by which they are
commonly known, as we they are used frequently in examples. There is more to do, though. We need to classify
numbers such as h0, 1, 1, 1, 1, ...i and h3, 1, 4, 1, 5, 9, ...i, and we have divisors of zero to take care of, such as h1, 0, 1, 0, ...i
and h0, 1, 0, 1, ...i, neither of which is obviously zero, but whose product h1, 0, 1, 0, ...ih0, 1, 0, 1, ...i = h10, 10, 10, ...i =
h0, 0, 0, ...i is zero. To sort non-constant sequences into useful equivalence classes, we will say r = hr1 , r2 , . . . , rj , . . .i =
s = hs1 , s2 , . . . , sk , . . .i if the set of indices on which the terms agree form a large set. The sort of large sets we are
looking for are the elements of a free ultrafilter over N. Recall that a filter F over N satisfies the following

Definition X.1 A filter F on N is any family of nonempty subsets of N with the following properties:

1. N F and
/ F.

2. A F and B F imply A B F.

3. A F and A B N imply B F.

Definition X.2 A filter F is said to be free (or non-principal) if the intersection of all sets belonging to F is empty.

One particularly important filter is a Frechet filter, which is also known as a cofinite filter because it consists of all
cofinite subsets of N. A Frechet filter is a free filter.

Definition X.3 An ultrafilter is a filter that cannot be enlarged.

Theorem X.4 Let U be an ultrafilter on N. Then for any subset A of N, either A or N A belongs to U.
S
Corollary X.5 Let U be an ultrafilter on N. Then for any finite disjoint set of subsets {Ai } of N such that i Ai = N,
exactly one of the Ai belongs to U.

Ultrafilter Hypothesis There exists an ultrafilter U on N that contains the Frechet filter. Any ultrafilter containing
the Frechet filter is free. We will henceforth be referring to free ultrafilters when we use the term ultrafilter.
There are an infinite number of possible ultrafilters, but they all give the same structure to sequences of reals: the
only difference is which sequences stand for which hyperreal. Therefore, we assume an ultrafilter has been fixed, and
do not mention its particular elements unless we need a particular version for an example.

Definition X.6 We say that two infinite sequences han i and hbn i are equal almost everywhere (a.e.) if the set of
indices on which the sequences agree is in U.
 13

Definition X.7 The number a is called standard if it is in the equivalence class of the first extension (usually called
the standard copy) of a real number 1 x = hx, x, x, ...i. A real number y not in the equivalence class of a standard
/ [hxi i] for some standard 1 x = hxi i is called nonstandard.
number, that is, y

Definition X.8 We say that r = s if r is in the equivalence class of s, that is, [s] = {r = hr1 , r2 , r3 , ...i : ri =
si for i u U}.

For example, x = h1, 1, 1/3, 1/4, 1/5, ...i [] where  = h1, 1/2, 1/3, 1/4, ...i. Additionally, if u = h1, 3, 5, 7, ...i is in
the ultrafilter, then y = h1, , 1/3, , 1/5, ...i [] also.
Now with filtration by a free ultrafilter, V0 (T ) = T has the structure of a completely ordered field, and the
superstructure V (T ) contains the formal representation of analysis on R.
In order to reason about R using R, the transfer principle must be proved, and we would need to talk about
entities like internal and external sets. We do not need transer for this particular work but we refer the reader to the
references for detailed proofs using formal languages. [6] [13] [3]
Two hyperreal numbers x and y are said to be infinitely close, written x y, if their difference is infinitesimal.
Around every real number c there is an uncountable set of hyperreal numbers infinitely close to c; in the literature
this is typically referred to as the halo (or monad) of c. Conversely, every finite hyperreal number is infinitely close to
exactly one real number, its shadow (or standard part); the shadow of any infinitesimal number is 0. There is also a
set of hypernatural numbers N R analogous to the natural numbers and sharing many properties (the only finite
hypernaturals are natural numbers, consecutive hypernaturals are separated by unity, etc.). An important role is also
played by the hyperfinite sets, finite or uncountable sets of hyperreal numbers that share many of the properties of
finite sets, and functions on those sets.
If f : R 7 R is a real-valued function, there exists a uniquely-defined function f : R 7 R, f s -transform, that
agrees with f on R and obeys the transfer principle (loosely speaking, any statement about f on R is true iff the
equivalent statement about f on R holds; e.g., f (x) < L for all real x and some real L iff f (x) < L for all hyperreal
x and some hyperreal L). Since f and f are for all practical purposes the same function, we typically drop the
prefix and simply write f .

B. The HyperHyperReals R

We have seen that the hyperreal number system R is a proper field extension of R containing infinite and infinites-
imal numbers (being larger and smaller respectively in absolute value than any positive real number) both positive
and negative.
In a similar manner to the construction of R from R, we may create R the second-order hyperreal numbers, that
has second-order infinite and infinitesimal numbers (larger and smaller respectively in absolute value than any positive
hyperreal number) [7][1][4]. We construct R as an ultrapower of R, i N Ri /U , U a co-countable ultrafilter. All
properties of R relative to R are exactly analogous to those of R relative to R; e.g., each finite (relative to R)
R is second-order infinitely close to exactly one hyperreal number x (and therefore infinitely close to exactly one
real number provided x itself is finite). Finally, we may extend any hyperreal function (and thus any real function)
to an equivalent second-order hyperreal one.
We note that while it is possible to build new number classes between the infinitesimal and appreciable (non-
infinitesimal) numbers, and between appreciables and infinite numbers using only countably infinite sequences of
hyperreals, one cannot build hyperinfinitesimals and hyperinfinites this way because of this

Theorem X.9 (Lenth) Every sequence of infinitesimals has an infinitesimal lower bound. Proof: Let aj = aij > 0
V
be a sequence of infinitesimals. Let bk < min{|ai,j | : i = k, j = 1, ..., k j = k, i = 1, ..., k}. Then hbk im with
hbk im = bk m is smaller than aj for all j.
 14

The proof that every sequence of infinite numbers has an infinite upper bound is similar. Therefore, we require
uncountably infinite sequences of infinitesimals and infinite numbers, and a co-countable ultrafilter.

[1] V. Benci, M di Nasso, and M. Forti. The eightfold path to non standard analysis. In Nonstandard Methods and Applications
in Mathematics. Ass. Symbol. Logic, La Jolla, California, 2006.
[2] V. Benci and M. Di Nasso. Alpha-theory: An elementary axiomatics for nonstandard analysis. Expositiones Mathematicae,
21(4):355386, 2003.
[3] Martin Davis. Applied Nonstandard Analysis. John Wiley and Sons, New York London Sydney Toronto, 1977.
[4] M. Forti. A simple algebraic characterization of nonstandard extensions. Proc. Amer. Math. Soc., 140:29032912, 2011.
[5] W. S. Hatcher. Calculus is algebra. The American Mathematical Monthly, 89(6):362370, 1982.
[6] A. E. Hurd and Peter A. Loeb. An introduction to nonstandard real analysis. Academic Press, Orlando, 1968.
[7] Kevin Lenth. Extension of the hyperreal number system. Masters thesis, New Mexico Institute of Mining and Technology,
Socorro, NM, August 2009.
[8] Kevin Lenth. Integration over infinite and infinitesimal domains. Masters thesis, New Mexico Institute of Mining and
Technology, May 2009.
[9] E. Nelson. Internal set theory, a new approach to nsa. Bull. Amer. Math. Soc., 83:11651198, 1977.
[10] A. Robinson. Non-Standard Analysis, revised edition. Princeton University Press, 41 William Street, Princeton, New
Jersey 08540, 1996.
[11] A. Robinson and E. Zakon. Applications of model theory to algebra, analysis, probability. Holt, Rinehart and Winston,
1969.
[12] W. Slomczynski, J. Kwapien, and K. Zyczkowski. Entropy computing via integration over fractal measures. Chaos,
10(1):180188, 2000.
[13] K. D. Stroyan and W. A. J. Luxemburg. Non-Standard Analysis. Academic Press, 111 Fifth Avenue, New York, New
York, 10003, 1976.
[14] If we wish, we can get real or complex analysis, point set topology, and a number of other structures directly from V (R)
or even V (N), but for simplicity we ignore the slightly more involved approach. It is enough that we get real analysis with
S = R, complex analysis with S = C, and so on.
[15] We must take the infinite sequences of real numbers mod an ultrafilter