Numbers

Douglas Crockford
Fingers
Tools
Accounting

Writing
 Writing

Fertile Crescent
China
America
Egypt
1
10
100
1,000
10,000
100,000
1,000,000
 Greece
Α 1 Ι 10 Ρ 100
Β 2 Κ 20 Σ 200
Γ 3 Λ 30 Τ 300
Δ 4 Μ 40 Υ 400
Ε 5 Ν 50 Φ 500
Ϛ 6 Ξ 60 Χ 600
Ζ 7 Ο 70 Ψ 700
Η 8 Π 80 Ω 800
Θ 9 Ϟ 90 Ϡ 900
 Rome
I 1 X 10 C 100
II 2 XX 20 CC 200
III 3 XXX 30 CCC 300
IV 4 XL 40 CD 400
V 5 L 50 D 500
VI 6 LX 60 DC 600
VII 7 LXX 70 DCC 700
VIII 8 LXXX 80 DCCC 800
IX 9 XC 90 CM 900
 Rome
I 1 X 10 C 100
II 2 XX 20 CC 200
III 3 XXX 30 CCC 300
IV 4 XL 40 CD 400
V 5 L 50 D 500
VI 6 LX 60 DC 600
VII 7 LXX 70 DCC 700
VIII 8 LXXX 80 DCCC 800
IX 9 XC 90 CM 900
 China
一 1 十 10
二 2 百 100
三 3 千 1,000
四 4
五 5 萬 10,000
六 6 億 100,000,000
七 7 兆 1,000,000,000,000
八 8
九 9
India
0 Zero
1 Positional
2
3
4
5
6
7
8
9
 ΧϠ
DCIX
六百九
609
 -Whole Numbers

-1
-2
-3
-4
4
3
2
1
0
.
.
.

.
.
.
-60998
.
(6 * 10↑2 + 0 * 10↑1 + 9 * 10↑0)
 Integers

-1
-2
-3
-4
4
3
2
1
0
.
.
.

.
.
.
-60998
(0 - 1) *
(6 * 10↑2 + 0 * 10↑1 + 9 * 10↑0)
 Real

-1
-2
-3
-4
4
3
2
1
0
.
.
.

.
.
.
-60902
.
(0 - 1) *
(6 * 10↑2 + 0 * 10↑1 + 9 * 10↑0)
+ 0 * 10↑-1 + 2 * 10↑-2)
Decimal Point
_
60902
60902
.
60902
,
Decimal Point
_
2,048
60902
60902
.
60902
,
Decimal Point
_
60902
60902
.
60902
,
Base

10
Base

10
20
Base

10
20
12
Base

10
20
12
60
 Binary

-1
-2
-3
-4
4
3
2
1
0
.
.
.

.
.
.
110098
.
(1 * 2↑3 + 1 * 2↑2 +
0 * 2↑1 + 0 * 2↑0)
 Sign bit
•  Signed magnitude
  Two zeros: 0 and -0
•  One’s complement
  -n = not n
  End around carry
  Two zeros: 0 and -0
•  Two’s complement
  -n = (not n) + 1
  The only zero is 0
  Extra negative
 indexOf

"abc".indexOf("z")

-1
null
 int

int32 + int32
 int

Java Correct
int32 + int32 int32 int33
 int

Java Correct
int32 + int32 int32 int33
int32 * int32
 int

Java Correct
int32 + int32 int32 int33
int32 * int32 int32 int63
 int

Overflow
Java Correct
int32 + int32 int32 int33
int32 * int32 int32 int63
 Overflow:
What should happen?

•  Store a null value.

•  Store the largest possible (saturation).
•  Fault.
•  The machine should halt.
To maximize creation of error:
•  Discard the most significant bits without
notification.
 Scaled Arithmetic
scaled value ← integer value * scale factor
Values can be added only if they have the same scale factor.
product ←
(multiplicand * scaled value) /
scale factor
quotient ←
(dividend * scale factor) /
scaled value
In some applications, it may be difficult to find an optimal
scale factor.
 Floating Point

Approximately real.
More convenient and accurate
than scaled arithmetic.
 Decimal Floating Point

significand * 10^log10(scale factor)

 Binary Floating Point
(sign * significand) * 2^exponent

biased
exponent significand

sign
 0.1 + 0.2
=
0.30000000000000004
 0.1

-10

-12
-13
-14
-15
-16
-17
-18
-19
-20
-21
-22
-23
-24
-25
-26
-27
-28
-29
-11
-1
-2
-3
-4
-5
-6
-7
-8
-9
...

...
4
3
2
1
0

. 000011001100110011001100110011…
too small!
. 0000110011001100110011001100110
1/16
too big!
. = 0.09375
+ 1/32
 0.1

-10

-12
-13
-14
-15
-16
-17
-18
-19
-20
-21
-22
-23
-24
-25
-26
-27
-28
-29
-11
-1
-2
-3
-4
-5
-6
-7
-8
-9
...

...
4
3
2
1
0

. 000011001100110011001100110011…
too small!
. 0000110011001100110011001100110
too big!
. 0000110011001100110011001101010
 0.1

-10

-12
-13
-14
-15
-16
-17
-18
-19
-20
-21
-22
-23
-24
-25
-26
-27
-28
-29
-11
-1
-2
-3
-4
-5
-6
-7
-8
-9
...

...
4
3
2
1
0

. 000011001100110011001100110011…
too small! . 0000110011001100110011001100110
too big!
. 0000110011001100110011001101010
 0.1

-10

-12
-13
-14
-15
-16
-17
-18
-19
-20
-21
-22
-23
-24
-25
-26
-27
-28
-29
-11
-1
-2
-3
-4
-5
-6
-7
-8
-9
...

...
4
3
2
1
0

. 000011001100110011001100110011…
too small! . 0000110011001100110011001100110
too big! . 0000110011001100110011010001010
The Associative Law holds only if
all of the inputs, outputs, and
intermediate results can all be
represented exactly.
(a + b) + c
a + (b + c)
 Business could not use
binary floating point.

Scienti c Business
Fortran Cobol
Binary Floating Point BCD
 Binary Coded Decimal
0000 0 0101 5
0001 1 0110 6
0010 2 0111 7
0011 3 1000 8
0100 4 1001 9
 Binary Floating Point
Text Conversion

•  Needs to be
  Correct
  Optimal
  Unsurprising
•  Computationally expensive
 Most modern programming
languages are a confusion of
faulty number types.

byte char short int long float

double
 The Ariane 5 Failure

bh = (short)dbHorizontalBias;
JavaScript has only one number type.

This avoids a large class of errors.

JavaScript has only one number type.

This avoids a large class of errors.

Unfortunately, it is the wrong type.
 DEC64
•  A modern decimal floating point type.
•  I recommend that DEC64 is the only number type
in well designed application programming
languages.
•  If there is only one number type, you cannot
make an error by choosing the wrong type.
•  In a hardware implementation, DEC64 integers
can be added in a single cycle, eliminating the
performance justification for int.
 DEC64
•  Numbers work the way humans think numbers
work.
•  Elimination numeric confusion reduces errors.
•  Conversion of DEC64 numbers to text and back
is simple, efficient, correct, and unsurprising.
•  DEC64 can exactly represent decimal fractions.
16 digits.
1.0E-127 3.6028797018963967E+143
 DEC64

56 8

Coefficient

Exponent
Number ←
Coefficient * 10↑Exponent
 dec64.com
56 8

Coefficient

Exponent
https://github.com/douglascrockford/DEC64/
 0/0

Unde ned.
Catch re!
null
0
1
2
0*n
0*n

0
NaN
0*n

0
NaN
 0*n
Code generators
Macro processors
Partial evaluators
Non-conditional idioms
 0/n
0*n
n*0
0 modulo n
0
Thanks
Leonardo Fibonacci of Pisa
Gottried Wilhelm Leibniz
George Boole
Claude Shannon
Tehuti