Complete Interval Arithmetic and its

Implementation on the Computer

Ulrich W. Kulisch
Institut für Angewandte und Numerische Mathematik
Universität Karlsruhe

Abstract: Let IIR be the set of closed and bounded intervals of real numbers. Arithmetic
in IIR can be defined via the power set IPIR (the set of all subsets) of real numbers. If
divisors containing zero are excluded, arithmetic in IIR is an algebraically closed subset
of the arithmetic in IPIR, i.e., an operation in IIR performed in IPIR gives a result that
is in IIR. Arithmetic in IPIR also allows division by an interval that contains zero. Such
division results in closed intervals of real numbers which, however, are no longer bounded.
The union of the set IIR with these new intervals is denoted by (IIR).
The paper shows that arithmetic operations can be extended to all elements of the set
(IIR). On the computer, arithmetic in (IIR) is approximated by arithmetic in the subset
(IF ) of closed intervals over the floating-point numbers F ⊂ IR. The usual exceptions of
floating-point arithmetic like underflow, overflow, division by zero, or invalid operation do
not occur in (IF ).

Keywords: computer arithmetic, floating-point arithmetic, interval arithmetic, arith-

metic standards.

1 Introduction or a Vision of Future Computing

Computers are getting ever faster. The time can already be foreseen when the
P C will be a teraflops computer. With this tremendous computing power scientific
computing will experience a significant shift from floating-point arithmetic toward
increased use of interval arithmetic. With very little extra hardware, interval arith-
metic can be made as fast as simple floating-point arithmetic [3]. Nearly everything
that is needed for fast interval arithmetic is already available on most existing pro-
cessors, thanks to multimedia applications. What is still missing are the arithmetic
operations with the directed roundings. Properly developed, interval arithmetic is
a complete and exception-free calculus. The exceptions of floating-point arithmetic
like underflow, overflow, division by zero, or invalid operations do not occur in such
interval arithmetic. This will be shown in the following.
For interval evaluation of an algorithm (a sequence of arithmetic operations)
in the real number field a theorem by R. E. Moore [7] states that increasing the
precision by k digits reduces the error bounds by b−k , i.e., results can always be
guaranteed to a number of correct digits by using variable precision interval arith-
metic (for details see [1], [9]). Lengthy interval arithmetic can be made very fast
by an exact dot product and complete arithmetic [4]. By pipelining, an exact dot
product can be computed in the time the processor needs to read the data, i.e.,
it comes with extreme speed. Lengthy interval arithmetic fully benefits from this
speed. It can easily be applied by operator overloading.
The tremendous progress in computer technology should be accompanied by
extension of the mathematical capacity of the computer. A balanced standard for

1
computer arithmetic should require that the basic components of modern computing
(floating-point arithmetic, interval arithmetic, and an exact dot product) should be
provided by the computer’s hardware. See [5].

2 Remarks on Floating-Point Arithmetic

Computing is usually done in the set of real numbers IR. The real numbers can
be defined as a conditionally complete, linearly ordered field. Conditionally com-
plete means that every bounded subset has an infimum and a supremum. Ev-
ery conditionally ordered set can be completed by joining a least and a great-
est element. In case of the real numbers these are called −∞ and +∞. Then
IR∗ := IR ∪ {−∞} ∪ {+∞} is a complete lattice. The elements −∞ and +∞, how-
ever, are not real numbers, i.e., they are not elements of the field. The cancellation
law a + c = b + c ⇒ a = b, for instance, does not hold for c = ∞.
A real number consists of a sign, an integral, and a fractional part, for instance:
±345.789123 · · · ∈ IR. The point may be shifted to any other position if we com-
pensate for this shifting by a corresponding power of b (here b = 10). If the point
is shifted immediately to the left of the first nonzero digit: ±0.345789123 · · · · 103
the representation is called normalized. Zero is the only real number that has no
such representation. It needs a special encoding. Thus a normalized real number
consists of a signed fractional part m (mantissa) and an integer exponent e and we
have |m| < 1.
Only subsets of these numbers can be represented on the computer. If the
mantissa in truncated after the lth digit and the exponent is limited by emin < e <
emax one speaks of a floating-point number. The set F of all such floating-point
numbers is a finite subset of IR.
Arithmetic for floating-point numbers may cause exceptions. Well known such
exceptions are underflow, overflow, division by zero, or invalid operation. To
avoid interruption of program execution in case of an exception the so-called IEEE
floating-point arithmetic standard provides additional elements and defines oper-
ations for these, for instance, 4/0 =: ∞, −4/0 =: −∞, ∞ − ∞ =: N aN, 0 · ∞ =:
N aN, ∞/∞ =: N aN, 0/0 =: N aN, 1/(−∞) =: −0, (−0.3)/∞ =: −0. It should
be clear, however, that these artificial strategic objects −∞, +∞, N aN, −0,1 or +0
with their operations are not elements of the real number field and thus are not
floating-point numbers.

3 Arithmetic for Intervals of IIR and IF

Interval arithmetic is another arithmetic tool. It solely deals with sets of real
numbers. Neither the exceptions of floating-point arithmetic mentioned above nor
the strategic objects to deal with them occur or are needed in interval arithmetic.
The symbol IIR usually denotes the set of closed and bounded intervals of IR. Arith-
metic in IIR can be interpreted as a systematic calculus to deal with inequalities.
We assume here that the basic rules for arithmetic in IIR with zero not in the divisor
are known to the reader. It is a fascinating result that, in contrast to floating-point
arithmetic, interval arithmetic even on computers can be further developed into a
well rounded, exception-free, closed calculus. We briefly sketch this development
here.
In floating-point arithmetic the crucial operation that leads to the exceptional
strategic objects mentioned above is division by zero. So we begin our study of
1 In IR, 0 is defined as the neutral element of addition. From the assumption that there are two

such elements 0 and 0′ it follows immediately that they are equal: 0 + 0′ = 0 = 0′ .

2
extended interval arithmetic by defining division by an interval that contains zero.
The set IIR is a subset of the power set IPIR (which is the set of all subsets) of
real numbers. For A, B ∈ IPIR arithmetic operations are defined by
^
A ◦ B := {a ◦ b | a ∈ A ∧ b ∈ B}, for all ◦ ∈ {+, −, ·, /}. (3.1)
A,B∈IPIR

The following properties are obvious and immediate consequences of this defini-
tion:
A ⊆ B ∧ C ⊆ D ⇒ A ◦ C ⊆ B ◦ D, for all A, B, C, D ∈ IPIR, (3.2)
and in particular

a ∈ A ∧ b ∈ B ⇒ a ◦ b ∈ A ◦ B, for all A, B ∈ IPIR. (3.3)

Property (3.2) is called inclusion-isotony (or inclusion-monotonicity). Property

(3.3) is called the inclusion property. (3.2) and (3.3) are the fundamental properties
of interval arithmetic. Under the assumption 0 6∈ B for division, the intervals of
IIR are an algebraically closed subset2 of the power set IPIR, i.e., an operation for
intervals of IIR performed in IPIR always delivers an interval of IIR.
On the computer, arithmetic in IIR is approximated by an arithmetic in IF .
An interval of IF represents a continuous set of real numbers with floating-point
bounds of F . Arithmetic operations in IF are defined by those in IIR with the lower
bound of the result rounded downwards and the upper bound rounded upwards.
In floating-point arithmetic, division by zero does not lead to a real number.
In contrast to this, in interval arithmetic division by an interval that contains zero
can be defined in a strict mathematical manner. The result again is a set of real
numbers.
In accordance with (3.1) division in IIR is defined by
^
A/B := {a/b | a ∈ A ∧ b ∈ B}. (3.4)
A,B∈IIR

The quotient a/b is defined as the inverse operation of multiplication, i.e., as the
solution of the equation b · x = a. Thus (3.4) can be written in the form
^
A/B := {x | bx = a ∧ a ∈ A ∧ b ∈ B}. (3.5)
A,B∈IIR

For 0 ∈/ B (3.4) and (3.5) are equivalent. While in IR division by zero is not
defined, the representation of A/B by (3.5) allows definition of the operation and
also interpretation of the result for 0 ∈ B.
By way of interpreting (3.5) for A = [a1 , a2 ] and B = [b1 , b2 ] ∈ IIR with 0 ∈ B
the following eight distinct cases can be set out:
1 0 ∈ A, 0 ∈ B.
2 0∈/ A, B = [0, 0].
3 a1 ≤ a2 < 0, b1 < b2 = 0.
4 a1 ≤ a2 < 0, b1 < 0 < b2 .
5 a1 ≤ a2 < 0, 0 = b1 < b2 .
6 0 < a1 ≤ a2 , b1 < b2 = 0.
7 0 < a1 ≤ a2 , b1 < 0 < b2 .
8 0 < a1 ≤ a2 , 0 = b1 < b2 .
The list distinguishes the cases 0 ∈ A (case 1) and 0 ∈/ A (cases 2 to 8). Since
it is generally assumed that 0 ∈ B, these eight cases indeed cover all possibilities.
2 as the integers are of the real numbers.

3
Since every x ∈ IR fulfills the equation 0 · x = 0 we obtain in case 1: A/B =
IR = (−∞, +∞). Here the parentheses indicate that the bounds are not included
in the set. In case 2 the set defined by (3.5) consists of all elements which fulfill the
equation 0 · x = a for a ∈ A. Since 0 ∈/ A, there is no real number which fulfills this
equation. Thus A/B is the empty set, i.e., A/B = ∅.

case A = [a1 , a2 ] B = [b1 , b2 ] B′ A/B ′ A/B

1 0∈A 0∈B (−∞, +∞)
2 0∈/A B = [0, 0] ∅
3 a2 < 0 b1 < b2 = 0 [b1 , (−ǫ)] [a2 /b1 , a1 /(−ǫ)] [a2 /b1 , +∞)
4 a2 < 0 b1 < 0 < b2 [b1 , (−ǫ)] [a2 /b1 , a1 /(−ǫ)] (−∞, a2 /b2 ]
∪ [ǫ, b2 ] ∪ [a1 /ǫ, a2 /b2 ] ∪ [a2 /b1 , +∞)
5 a2 < 0 0 = b1 < b2 [ǫ, b2 ] [a1 /ǫ, a2 /b2 ] (−∞, a2 /b2 ]
6 a1 > 0 b1 < b2 = 0 [b1 , (−ǫ)] [a2 /(−ǫ), a1 /b1 ] (−∞, a1 /b1 ]
7 a1 > 0 b1 < 0 < b2 [b1 , (−ǫ)] [a2 /(−ǫ), a1 /b1 ] (−∞, a1 /b1 ]
∪ [ǫ, b2 ] ∪ [a1 /b2 , a2 /ǫ] ∪ [a1 /b2 , +∞)
8 a1 > 0 0 = b1 < b2 [ǫ, b2 ] [a1 /b2 , a2 /ǫ] [a1 /b2 , +∞)

Table 1: The eight cases of interval division A/B, with A, B ∈ IIR, and 0 ∈ B.

In all other cases 0 ∈ / A also. We have already observed under case 2 that
the element 0 in B does not contribute to the solution set. So it can be excluded
without changing the set A/B.
So the general rule for computing the set A/B by (3.5) is to remove its zero from
the interval B and replace it by a small positive or negative number ǫ as the case
may be. The resulting set is denoted by B ′ and represented in column 4 of Table 1.
With this B ′ the solution set A/B ′ can now easily be computed by applying the
rules for closed and bounded real intervals. The results are shown in column 5 of
Table 1. Now the desired result A/B as defined by (3.5) is obtained if in column 5
ǫ tends to zero.
Thus in the cases 3 to 8 the results are obtained by the limit process A/B =
lim A/B ′ . The solution set A/B is shown in the last column of Table 1 for all the
ǫ→0
eight cases. There, as usual in mathematics, parentheses indicate that the bound
is not included in the set. In contrast to this, brackets denote closed interval ends,
i.e., the bound is included.
The operands A and B of the division A/B in Table 1 are intervals of IIR. The
results of the division shown in the last column, however, are no longer intervals
of IIR. The result is now an element of the power set IPIR. With the exception of
case 2 the result is now a set which stretches continuously to −∞ or +∞ or both.
In two cases (rows 4 and 7 in Table 1) the result consists of the union of two
distinct sets of the form (−∞, c2 ] ∪ [c1 , +∞). These cases can easily be identified by
the signs of the bounds of the divisor before the division is executed. (For interval
multiplication and division a case selection has to be done before the operations
are performed). With existing processors only one interval can be delivered as the
result of an interval operation. In the cases 4 and 7 of Table 1 the result, however,
can be returned as an improper interval [c1 , c2 ] where the left hand bound is higher

4
than the right hand bound. Motivated by the extended interval Newton method3
it is reasonable to separate these results into the two distinct sets: (−∞, c2 ] and
[c1 , +∞). The fact that an arithmetic operation delivers two distinct results might
seem to be a totally new situation in computing. Evaluation of the square root,
however, also delivers two results and we have learned to live with it. Computing
certainly is able to deal with this situation.
In principle, a solution to the problem would be for the computer to provide a
flag for distinct intervals. The situation occurs if the divisor is an interval that
contains zero as an interior point. In cases 4 and 7 of Table 1 the flag would be
raised and signaled to the user. The user may then apply a routine of his choice to
deal with the situation as is appropriate for his application.4
If during a computation in the real number field zero appears as a divisor the
computation should be stopped immediately. In floating-point arithmetic the sit-
uation is different. Zero may be the result of an underflow. In such a case a cor-
responding interval computation would not deliver zero but a small interval with
zero as one bound and a tiny positive or negative number as the other bound. In
this case division is well defined by Table 1. The result is a closed interval which
stretches continuously to −∞ or +∞ as the case may be.
In the real number field zero as a divisor is an accident. So in interval arithmetic
division by an interval that contains zero as an interior point certainly will be a very
rare appearance. An exception is the interval Newton method. Here, however, it is
clear how the situation has to be handled. See, for instance, [4].
In the literature an improper interval [c1 , c2 ] with c1 > c2 occasionally is called
an ’exterior interval’. On the number circle an ’exterior interval’ is interpreted as an
interval with infinity as an interior point. We do not follow this line here. Interval
arithmetic is defined as an arithmetic for sets of real numbers. Operations for real
numbers which deliver ∞ as their result do not exist. Here and in the following the
symbols −∞ and +∞ are only used to describe sets of real numbers.
After the splitting of improper intervals into two distinct sets only four kinds of
result come from division by an interval of IIR which contains zero:

∅, (−∞, a], [b, +∞), and (−∞, +∞). (3.6)

We call such elements extended intervals. The union of the set of closed and
bounded intervals of IIR with the set of extended intervals is denoted by (IIR). The
elements of the set (IIR) are themselves simply called intervals. (IIR) is the set of
closed intervals of IR. (A subset of IR is called closed if the complement is open.)
Intervals of IIR and of (IIR) are sets of real numbers. −∞ and +∞ are not
elements of these intervals. It is fascinating that arithmetic operations can be
introduced for all elements of the set (IIR) in an exception-free manner. This will
be shown in the next section.
On a computer only subsets of the real numbers are representable. We assume
now that F is the set of floating-point numbers of a given computer. An interval
between two floating-point bounds represents the continuous set of real numbers
between these bounds. Similarly, except for the empty set, extended intervals also
represent continuous sets of real numbers.
3 Newton’s method reaches its ultimate elegance and strength in the extended interval Newton

method. If division by an interval that contains zero delivers two distinct sets the computation
is continued along two separate paths, one for each interval. This is how the extended interval
Newton method separates different zeros from each other and finally computes all zeros in a given
domain. If the interval Newton method delivers the empty set, the method has proved that there
is no zero in the initial interval.
4 This routine could be: modify the operands and recompute, or continue the computation

with one of the sets and ignore the other one, or put one of the sets on a list and continue the
computation with the other one, or return the entire set of real numbers (−∞, +∞) as result and
continue the computation, or stop computing, or ignore the flag, or any other action.

5
 To transform the eight cases of division by an interval of IIR which contains
zero into computer executable operations we assume now that the operands A and
B are floating-point intervals of IF . To obtain a computer representable result
we round the result shown in the last column of Table 1 into the least computer
representable superset. That is, the lower bound of the result has to be computed
with rounding downwards and the upper bound with rounding upwards. Thus on
the computer the eight cases of division by an interval of IF which contains zero
have to be performed as shown in Table 2.

case A = [a1 , a2 ] B = [b1 , b2 ] A♦ / B

1 0∈A 0∈B (−∞, +∞)

2 0∈/A B = [0, 0] ∅
3 a2 < 0 b1 < b2 = 0 [a2 ▽
/
b1 , +∞)
4 a2 < 0 / b ] ∪ [a ▽
b1 < 0 < b2 (−∞, a2 △ /
b1 , +∞)
2 2
5 a2 < 0 0 = b1 < b2 (−∞, a2 △ / b ]
2
6 a1 > 0 b1 < b2 = 0 (−∞, a1 △ / b ]
1
7 a1 > 0 ▽
b1 < 0 < b2 (−∞, a1 △ b1 ] ∪ [a1 b2 , +∞)
/ /

8 a1 > 0 0 = b1 < b2 [a1 ▽

/
b2 , +∞)

Table 2: The eight cases of interval division with A, B ∈ IF , and 0 ∈ B.

Table 3 shows the same cases as Table 2 in another layout.

B = [0, 0] b1 < b2 = 0 b1 < 0 < b2 0 = b1 < b2

a2 < 0 ∅ [a2 ▽/
b1 , +∞) (−∞, a2 △ / b ]
2 (−∞, a2 △/ b ]
2
∪ [a2 ▽ b1 , +∞)
/

a1 ≤ 0 ≤ a2 (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

0 < a1 ∅ (−∞, a1 △
/ b ]
1 (−∞, a1 △ / b ]
1 [a1 ▽
/
b2 , +∞)
∪ [a1 ▽ b2 , +∞)
/

Table 3: The result of the interval division with A, B ∈ IF , and 0 ∈ B.

Table 2 and Table 3 display the eight distinct cases of interval division A ♦
/ B

with A, B ∈ IF and 0 ∈ B. On the computer the empty interval ∅ needs a

particular encoding. (+NaN, −NaN) may be such an encoding. We explicitly stress
that the symbols −∞, +∞, −NaN, and +NaN are used here only to represent the
resulting sets.
These symbols are not elements of these sets and no operations are defined for
them.
Division by an interval of IF which contains zero on the computer also leads to
extended intervals as shown in (3.6) with a, b ∈ F . The union of the set of closed
and bounded intervals of IF with such extended intervals is denoted by (IF ). (IF )

6
is the set of closed intervals of real numbers where all finite bounds are elements of
F.

4 Arithmetic for Intervals of (IIR) and (IF )

For the sake of completeness, arithmetic operations now have to be defined for
all elements of (IIR) and (IF ). Since the development of arithmetic operations
follows an identical pattern in (IIR) and (IF ), we skip here the introduction of the
arithmetic in (IIR) and restrict the consideration to the development of arithmetic
in (IF ). This is the arithmetic that has to be provided on the computer.
First of all any operation with the empty set is again defined to be the empty
set.
The general procedure for defining all other operations follows a continuity prin-
ciple. Bounds like −∞ and +∞ in the operands A and B are replaced by a very
large negative and a very large positive number respectively. Then the basic rules
for the arithmetic operations in IIR and IF are applied. In the following tables
these rules are repeated and printed in bold letters.
In the resulting formulas the very large negative number is then shifted to −∞
and the very large positive number to +∞. Finally, very simple and well established
rules of real analysis like ∞ ∗ x = ∞ for x > 0, ∞ ∗ x = −∞ for x < 0, x/∞ =
x/ − ∞ = 0, ∞ ∗ ∞ = ∞, (−∞) ∗ ∞ = −∞ are applied together with variants
obtained by applying the sign rules and the law of commutativity.
Two situations have to be treated separately. These are the cases shown in rows
1 and 2 of Table 1.
If 0 ∈ A and 0 ∈ B (row 1 of Table 1), the result consists of all the real numbers,
i.e., A/B = (−∞, +∞). This applies to rows 2, 5, 6 and 8 of Table 8.
If 0 ∈
/ A and B = [0, 0] (row 2 of Table 1), the result of the division is the empty
set, i.e., A/B = ∅. This applies to rows 1, 3, 4 and 7 of column 1 of Table 8.
We outline the complete set of arithmetic operations for interval arithmetic in
(IF ) that should be provided on the computer in the next section. In summary it
can be said that after a possible splitting of an improper interval into two separate
intervals the result of arithmetic operations for intervals of (IF ) always leads to
intervals of (IF ) again. Exceptions or artificial strategic objects are not needed
for these operations. The reader should prove this assertion by referring to the
operations shown in the tables of the following section.
For the development in the preceding sections it was essential to distinguish
between parentheses and brackets. If a bound is adjacent to a parenthesis, the
bound is not included in the interval; if a bound is adjacent to a bracket, the bound
is included in the interval.
In the following tables an operator symbol with a ▽ upon it means an operation
performed with rounding downward. Correspondingly an operation with a △ upon
it means an operation performed with rounding upward.

7
5 Complete Arithmetic for Intervals of (IF )

Addition (−∞, b2 ] [b1 , b2 ] [b1 , +∞) (−∞, +∞)

(−∞, a2 ] (−∞, a2 △
+ b ]
2 (−∞, a2 △ + b ]
2 (−∞, +∞) (−∞, +∞)
[a1 , a2 ] + ▽
(−∞, a2 △ b2 ] [a 1 b 1 , a 2 △
+ + b ]
2 [a1 ▽+
b1 , +∞) (−∞, +∞)
[a1 , +∞) (−∞, +∞) [a1 ▽
+
b1 , +∞) [a1 ▽+
b1 , +∞) (−∞, +∞)
(−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

Table 4: Addition of extended intervals on the computer.

Subtraction (−∞, b2 ] [b1 , b2 ] [b1 , +∞) (−∞, +∞)

(−∞, a2 ] (−∞, +∞) (−∞, a2 △ b1 ]
− (−∞, a2 △ − b ]
1 (−∞, +∞)
[a1 , a2 ] [a1 ▽
−
b2 , +∞) [a 1 ▽ b 2 , a 2 △
− − b ]
1 (−∞, a2 △
− b ]
1 (−∞, +∞)
[a1 , +∞) ▽
[a1 b2 , +∞)
−
[a1 ▽
−
b2 , +∞) (−∞, +∞) (−∞, +∞)
(−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

Table 5: Subtraction of extended intervals on the computer.

8
 [b1 , b2 ] [b1 , b2 ] [b1 , b2 ] (−∞, b2 ] (−∞, b2 ] [b1 , +∞) [b1 , +∞)
Multiplication b2 ≤ 0 b1 < 0 < b 2 b1 ≥ 0 [0, 0] b2 ≤ 0 b2 ≥ 0 b1 ≤ 0 b1 ≥ 0 (−∞, +∞)
[a1 , a2 ], a2 ≤ 0 · b2, a 1 △
[a 2 ▽ · b1] [a 1 ▽· b2, a 1 △ · b1] · b2, a 2 △
[a 1 ▽ · b1] [0, 0] · b2 , +∞)
[a2 ▽ · b2 , +∞)
[a1 ▽ (−∞, a1 △ · b1 ] (−∞, a2 △ · b1 ] (−∞, +∞)
a1 < 0 < a 2 · b1, a 1 △
[a 2 ▽ · b1] [min(a 1 ▽ · b2, a 2 ▽ · b 1 ), · b2, a 2 △
[a 1 ▽ · b2] [0, 0] (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
max(a 1 △· b1, a 2 △
· b 2 )]
[a1 , a2 ], a1 ≥ 0 [a 2 ▽· b1, a 1 △ · b2] · b1, a 2 △
[a 2 ▽ · b2] · b1, a 2 △
[a 1 ▽ · b2] [0, 0] (−∞, a1 △ · b2 ] (−∞, a2 △ · b2 ] · b1 , +∞)
[a2 ▽ · b1 , +∞)
[a1 ▽ (−∞, +∞)
[0, 0] [0, 0] [0, 0] [0, 0] [0, 0] [0, 0] [0, 0] [0, 0] [0, 0] [0, 0]
9

(−∞, a2 ], a2 ≤ 0 [a2 ▽· b2 , +∞) (−∞, +∞) (−∞, a2 △ · b1 ] [0, 0] · b2 , +∞)

[a2 ▽ (−∞, +∞) (−∞, +∞) (−∞, a2 △ · b1 ] (−∞, +∞)
(−∞, a2 ], a2 ≥ 0 [a2 · b1 , +∞)
▽ (−∞, +∞) (−∞, a2 △ · b2 ] [0, 0] (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
[a1 , +∞), a1 ≤ 0 (−∞, a1 △ · b1 ] (−∞, +∞) · b2 , +∞)
[a1 ▽ [0, 0] (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
[a1 , +∞), a1 ≥ 0 (−∞, a1 △ · b2 ] (−∞, +∞) · b1 , +∞)
[a1 ▽ [0, 0] (−∞, a1 △ · b2 ] (−∞, +∞) (−∞, +∞) · b1 , +∞)
[a1 ▽ (−∞, +∞)
(−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) [0, 0] (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

Table 6: Multiplication of extended intervals on the computer.

 Division [b1 , b2 ] [b1 , b2 ] (−∞, b2 ] [b1 , +∞)
0∈ /B b2 < 0 b1 > 0 b2 < 0 b1 > 0
[a1 , a2 ], a2 ≤ 0 [a 2 ▽
/
b 1, a 1 △ / b ]
2 [a 1 ▽
/
b 1, a 2 △ / b ]
2 [0, a1 △/ b ]
2 [a 1 ▽ b1 , 0]
/

[a1 , a2 ], a1 < 0 < a2 [a 2 ▽ b 2 , a 1 △

/ / b ]
2 [a 1 ▽ b 1 , a 2 △
/ / b ]
1 [a2 ▽ b2 , a1 △
/ / b ]
2 [a1 ▽ b1 , a2 △
/ / b ]
1
[a1 , a2 ], a1 ≥ 0 ▽
/ / ▽
[a 2 b 2 , a 1 △ b 1 ] [a 1 b 2 , a 2 △ b 1 ]
/ / ▽
[a2 b2 , 0]
/
[0, a2 △ b1 ]
/

[0, 0] [0, 0] [0, 0] [0, 0] [0, 0]

10

(−∞, a2 ], a2 ≤ 0 ▽
[a2 b1 , +∞)
/
(−∞, a2 △ / b ]
2 [0, +∞) (−∞, 0]
(−∞, a2 ], a2 ≥ 0 [a2 ▽ b2 , +∞)
/
(−∞, a2 △ / b ]
1 [a2 ▽/
b2 , +∞) (−∞, a2 △ / b ]
1
[a1 , +∞), a1 ≤ 0 (−∞, a1 △ / b ]
2 [a1 ▽
/
b1 , +∞) (−∞, a1 △ / b ]
2 ▽
[a1 b1 , +∞)
/

[a1 , +∞), a1 ≥ 0 (−∞, a1 △ / b ]

1 [a1 ▽
/
b2 , +∞) (−∞, 0] [0, +∞)
(−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

Table 7: Division of extended intervals with 0 6∈ B on the computer.

 Division B= [b1 , b2 ] [b1 , b2 ] [b1 , b2 ] (−∞, b2 ] (−∞, b2 ] [b1 , +∞) [b1 , +∞)
0∈B [0, 0] b1 < b 2 = 0 b1 < 0 < b 2 0 = b1 < b 2 b2 = 0 b2 > 0 b1 < 0 b1 = 0 (−∞, +∞)
[a1 , a2 ], a2 < 0 ∅ [a 2 ▽
/
b 1 , +∞) (−∞, a 2 △ / b2] (−∞, a 2 △
/ b2] [0, +∞) (−∞, a2 △ / b2 ] (−∞, 0] (−∞, 0] (−∞, +∞)
∪ [a 2 ▽
/
b 1 , +∞) ∪ [0, +∞) ∪ [a2 ▽
/
b1 , +∞)
[a1 , a2 ], a1 ≤ 0 ≤ a2 (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
[a1 , a2 ], a1 > 0 ∅ (−∞, a 1 △
/ b1] (−∞, a 1 △ / b1] [a 1 ▽
/
b 2 , +∞) (−∞, 0] (−∞, 0] (−∞, a1 △ / b1 ] [0, +∞) (−∞, +∞)
∪ [a 1 ▽
/
b 2 , +∞) ∪ [a1 ▽
/
b2 , +∞) ∪ [0, +∞)
(−∞, a2 ], a2 < 0 ∅ [a2 ▽
/
b1 , +∞) (−∞, a2 △ / b2 ] (−∞, a2 △
/ b2 ] [0, +∞) (−∞, a2 △ / b2 ] (−∞, 0] (−∞, 0] (−∞, +∞)
11

∪ [a2 ▽
/
b1 , +∞) ∪ [0, +∞) ∪ [a2 ▽
/
b1 , +∞)
(−∞, a2 ], a2 > 0 (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
[a1 , +∞), a1 < 0 (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)
[a1 , +∞), a1 > 0 ∅ (−∞, a1 △
/ b1 ] (−∞, a1 △ / b1 ] [a1 ▽
/
b2 , +∞) (−∞, 0] (−∞, 0] (−∞, a1 △ / b1 ] [0, +∞) (−∞, +∞)
∪ [a1 ▽
/
b2 , +∞) ∪ [a1 ▽
/
b2 , +∞) ∪ [0, +∞)
(−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞) (−∞, +∞)

Table 8: Division of extended intervals with 0 ∈ B on the computer.

 The rules for the operations of extended intervals on the computer in Tables 4—8
look rather complicated. Their implementation seems to require many case distinc-
tions. The situation, however, can be greatly simplified as follows.
On the computer actually only the basic rules for addition, subtraction, multi-
plication, and division for closed and bounded intervals of IF including division by
an interval that includes zero are to be provided. In Tables 4—8 these rules are
printed in bold letters.
The remaining rules shown in the tables can automatically be produced out of
these basic rules by the computer itself if a few well established rules for computing
with −∞ and +∞ are formally applied. With x ∈ F these rules are

∞ + x = ∞, −∞ + x = −∞,
−∞ + (−∞) = (−∞) · ∞ = −∞, ∞ + ∞ = ∞ · ∞ = ∞,
∞ · x = ∞ for x > 0, ∞ · x = −∞ for x < 0,
x x
∞ = −∞ = 0,

together with variants obtained by applying the sign rules and the law of commu-
tativity. If in an interval operand a bound is −∞ or +∞ the multiplication with 0
is performed as if the following rules would hold
0 · (−∞) = 0 · (+∞) = (−∞) · 0 = (+∞) · 0 = 0.
These rules have no meaning otherwise.

6 Comparison Relations and Lattice Operations

Three comparison relations are important for intervals of IF and (IF ):
equality, less than or equal, and set inclusion. (6.1)
Let A and B be intervals of (IF ) with bounds a1 ≤ a2 and b1 ≤ b2 respectively.
Then the relations equality and less than or equal in (IF ) are defined by:
A=B :⇔ a1 = b1 ∧ a2 = b2 ,
A≤B :⇔ a1 ≤ b1 ∧ a2 ≤ b2 .
Since bounds for intervals of (IF ) may be −∞ or +∞ these comparison relations
are executed as if performed in the lattice {F ∗ , ≤} with F ∗ := F ∪ {−∞} ∪ {+∞}.
With the order relation ≤, {(IF ), ≤} is a lattice. The greatest lower bound (glb)
and the least upper bound (lub) of A, B ∈ (IF ) are the intervals
glb(A, B) := [min(a1 , b1 ), min(a2 , b2 )],
lub(A, B) := [max(a1 , b1 ), max(a2 , b2 )].
The greatest lower bound and the least upper bound of an interval with the empty
set are both the empty set.
The inclusion relation in (IF ) is defined by
A ⊆ B :⇔ b1 ≤ a1 ∧ a2 ≤ b2 . (6.2)
With the relation ⊆, {(IF ), ⊆} is also a lattice. The least element in {(IF ), ⊆} is
the empty set ∅ and the greatest element is the interval (−∞, +∞). The infimum
of two elements A, B ∈ (IF ) is the intersection and the supremum is the interval
hull (convex hull):
inf (A, B) := [max(a1 , b1 ), min(a2 , b2 )] or the empty set ∅,
sup(A, B) := [min(a1 , b1 ), max(a2 , b2 )].

12
The intersection of an interval with the empty set is the empty set. The interval
hull with the empty set is the other operand.
If in the formulas for glb(A, B), lub(A, B), inf (A, B), sup(A, B), a bound is −∞
or +∞ a parenthesis should be used at this interval bound to denote the resulting
interval. This bound is not an element of the interval.
If in any of the comparison relations defined here both operands are the empty
set, the result is true. If in (6.2) A is the empty set the result is true. Otherwise
the result is false if in any of the three comparison relations only one operand is the
empty set.5
A particular case of inclusion is the relation element of. It is defined by

a ∈ B :⇔ b1 ≤ a ∧ a ≤ b2 . (6.3)

Another useful check is for whether an interval [a1 , a2 ] is a proper interval, that
is, if a1 ≤ a2 .

7 Evaluation of Functions
Interval evaluation of real functions fits smoothly into complete interval arithmetic
as developed in the previous sections. Let f be a function and Df its domain of
definition. For an interval X ⊆ Df , the range f (X) of f is defined as the set of the
function’s values for all x ∈ X:

f (X) := {f (x)|x ∈ X}. (7.1)

A function f (x) = a/(x − b) with Df = IR \ {b} is sometimes called singular or

discontinuous at x = b. Both descriptions are meaningless in a strict mathematical
sense. Since x = b is not of the domain of f , the function cannot have any property
at x = b.
In this strict sense a division 2/[b1 , b2 ] by an interval [b1 , b2 ] that contains zero
as an interior point, b1 < 0 < b2 , means:
2/([b1 , 0) ∪ (0, b2 ]) = 2/[b1 , 0) ∪ 2/(0, b2 ] = (−∞, 2/b1 ] ∪ [2/b2 , +∞).

We give two examples:

f (x) = 4/(x − 2)2 , Df = IR \ {2}, X = [1, 4],

f ([1, 2) ∪ (2, 4]) = f ([1, 2)) ∪ f ((2, 4]) = [4, +∞) ∪ [1, +∞) = [1, +∞).

g(x) = 2/(x − 2), Dg = IR \ {2}, X = [1, 3],

g([1, 2) ∪ (2, 3]) = g([1, 2)) ∪ g((2, 3]) = (−∞, −2] ∪ [2, +∞),
Here the flag distinct intervals should be raised and signaled to the user. The user
may then choose a routine to apply which is appropriate for the application.

It has been suggested in the literature that the entire set of real numbers (−∞, +∞)
be returned as result in this case. However, this may be a large overestimation of the
true result and there are applications (Newton’s method) which need the accurate
answer. To return the entire set of real numbers is also against a basic principle
of interval arithmetic—to keep the sets as small as possible. So a standard should
have the most accurate answer returned.
5 A convenient encoding of the empty set may be ∅ = [+N aN, −N aN ]. Then most comparison

relations and lattice operations considered in this section would deliver the correct answer if
conventional rules for N aN are applied. However, if A = ∅ then set inclusion (6.2) and computing
the interval hull do not follow this rule. So in these two cases whether A = ∅ must be checked
before the operations can be executed.

13
 A somewhat natural solution would be to continue the computation on different
processors, one for each interval. But the situation can occur repeatedly. How many
processors would we need? Future multicore units will provide a large number of
processors. They will suffice for a quite while. A similar situation occurs in global
optimization using subdivision. After a certain test several candidates may be left
for further investigation.
On the computer, interval evaluation of a real function f (x) for X ⊆ Df should
deliver a highly accurate enclosure of the range f (X) of the function.
Evaluation of a function f (x) for an interval X with X ∩ Df = ∅, of course,
does not make sense, since f (x) is not defined for values outside of its domain Df .
The computation should be terminated and an error message given to the user.
There are, however, applications in interval arithmetic where information about
a function f is useful when X exceeds the domain Df of f . The interval X may
also be the result of overestimation during an earlier interval computation.
In such cases the range of f can only be computed for the intersection
X ′ := X ∩ Df :
f (X ′ ) := f (X ∩ Df ) := {f (x)|x ∈ X ∩ Df }. (7.2)
To prevent the wrong conclusions being drawn, the user must be informed that the
interval X had to be reduced to X ′ := X ∩ Df to compute the delivered range. A
particular flag for domain overflow may serve this purpose. An appropriate routine
can be chosen and applied if this flag is raised.
We give a few examples:

l(x) := log(x), Dlog = (0, +∞),

log((0, 2]) = (−∞, log(2)].
But also
log([−5, 2]′ ) = log((0, 2]) = (−∞, log(2)].
The flag domain overflow should be set. It informs the user that the function has
been evaluated for the intersection X ′ := X ∩ Df = [−5, 2] ∩ (0, +∞) = (0, 2].

h(x) := sqrt(x), Dsqrt = [0, +∞),

sqrt([1, 4]) = [1, 2],
sqrt([4, +∞)) = [2, +∞).
sqrt([−5, −1]) terminates the program with the error message sqrt not defined for
[−5, −1],
sqrt([−5, 4]′ ) = sqrt([0, 4]) = [0, 2].
The flag domain overflow should be set. It informs the user that the function has
been evaluated for the intersection X ′ := X ∩ Df = [−5, 4] ∩ [0, +∞) = [0, 4].

k(x) := sqrt(x) − 1, Dk = [0, +∞),

k([−4, 1]′ ) = k([0, 1]) = sqrt([0, 1]) − 1 = [−1, 0].
The flag domain overflow should be set. It informs the user that the function has
been evaluated for the intersection X ′ := X ∩ Df = [−4, 1] ∩ [0, +∞) = [0, 1].

References
[1] Alefeld, G., Herzberger, J.: Introduction to Interval Computations. Academic
Press, New York, 1983.

[2] Kahan, W.: A More Complete Interval Arithmetic. Lecture Notes prepared for
a summer course at the University of Michigan, June 17-21, 1968.

14
[3] Kirchner, R., Kulisch, U.: Hardware support for interval arithmetic. Reliable
Computing 12:3, 225–237, 2006.

[4] Kulisch, U.,W.: Computer Arithmetic and Validity – Theory, Implementation

and Applications. De Gruyter, Berlin, New York, 2008.

[5] IFIPWG-IEEE754R: Letter of the IFIP WG 2.5 to the IEEE Computer Arith-
metic Revision Group, 2007.6

[6] Moore, R. E.: Interval Analysis. Prentice Hall Inc., Englewood Cliffs, New
Jersey, 1966.

[7] Moore, R. E.: Methods and Applications of Interval Analysis. SIAM, Philadel-
phia, Pennsylvania, 1979.

[8] Ratz, D.: On Extended Interval Arithmetic and Inclusion Isotony. Preprint,
Institut für Angewandte Mathematik, Universität Karlsruhe, 1999.

[9] Rump, S.M.: Kleine Fehlerschranken bei Matrixproblemen. Dissertation, Uni-

versität Karlsruhe, 1980.

6 See the author’s homepage, http://www.mathematik.uni-karlsruhe.de/ianm2/∼kulisch.

15