2 rand - subtractive 100 shuffle pseudo-random number generator
14 Generate a pseudo-random number using an subtractive 100 shuffle generator.
15 We return a pseudo-random number over the half closed interval [min,max).
16 By default, min is 0 and max is 2^64.
18 The shuffle method is fast and serves as a fairly good standard
19 pseudo-random generator. If you need a fast generator and do not
20 need a cryptographically strong one, this generator is likely to do
21 the job. Casual direct use of the shuffle generator may be
22 acceptable. For a much higher quality cryptographically strong
23 (but slower) generator use the Blum-Blum-Shub generator (see the
28 rand() Same as rand(0, 2^64)
29 rand(max) Same as rand(0, max)
31 The rand generator generates the highest order bit first. Thus:
35 will produce the save value as:
37 (rand(8) << 5) + rand(32)
39 when seeded with the same seed.
41 The rand generator has two distinct parts, the subtractive 100 method
42 and the shuffle method. The subtractive 100 method is described in:
44 "The Art of Computer Programming - Seminumerical Algorithms",
45 Vol 2, 3rd edition (1998), Section 3.6, page 186, formula (2).
47 The "use only the first 100 our of every 1009" is described in
48 Knuth's "The Art of Computer Programming - Seminumerical Algorithms",
49 Vol 2, 3rd edition (1998), Section 3.6, page 188".
51 The period and other properties of the subtractive 100 method
52 make it very useful to 'seed' other generators.
54 The shuffle method is feed values by the subtractive 100 method.
55 The shuffle method is described in:
57 "The Art of Computer Programming - Seminumerical Algorithms",
58 Vol 2, 3rd edition (1998), Section 3.2.2, page 34, Algorithm B.
60 The rand generator has a good period, and is fast. It is reasonable as
61 generators go, though there are better ones available. The shuffle
62 method has a very good period, and is fast. It is fairly good as
63 generators go, particularly when it is feed reasonably random
64 numbers. Because of this, we use feed values from the subtractive 100
65 method into the shuffle method.
67 The rand generator uses two internal tables:
69 additive table - 100 entries of 64 bits used by the subtractive
72 shuffle table - 256 entries of 64 bits used by the shuffle method
73 feed by the subtractive 100 method from the
76 The goals of this generator are:
78 * all magic numbers are explained
80 I (Landon Curt Noll) distrust systems with constants (magic
81 numbers) and tables that have no justification (e.g.,
82 DES). I believe that I have done my best to justify all of
83 the magic numbers used.
87 You have this source file, plus background publications,
88 what more could you ask?
90 * large selection of seeds
92 Seeds are not limited to a small number of bits. A seed
95 Most of the magic constants used by this generator ultimately are
96 based on the Rand book of random numbers. The Rand book contains
97 10^6 decimal digits, generated by a physical process. This book,
98 produced by the Rand corporation in the 1950's is considered
99 a standard against which other generators may be measured.
101 The Rand book of numbers was groups into groups of 20 digits. The
102 first 100 groups < 2^64 were used to initialize the default additive
103 table. The size of 20 digits was used because 2^64 is 20 digits
104 long. The restriction of < 2^64 was used to prevent modulus biasing.
106 The shuffle table size is longer than the 100 entries recommended
107 by Knuth. We use a power of 2 shuffle table length so that the
108 shuffle process can select a table entry from a new subtractive 100
109 value by extracting its low order bits. The value 256 is convenient
110 in that it is the size of a byte which allows for easy extraction.
112 We use the upper byte of the subtractive 100 value to select the
113 shuffle table entry because it allows all of 64 bits to play a part
114 in the entry selection. If we were to select a lower 8 bits in the
115 64 bit value, carries that propagate above our 8 bits would not
116 impact the subtractive 100 generator output.
118 It is 'nice' when a seed of "n" produces a 'significantly different'
119 sequence than a seed of "n+1". Generators, by convention, assign
120 special significance to the seed of '0'. It is an unfortunate that
121 people often pick small seed values, particularly when large seed
122 are of significance to the generators found in this file. An internal
123 process called randreseed64 will effectively eliminate the human
124 perceptions that are noted above.
126 It should be noted that the purpose of randreseed64 is to scramble a
127 seed ONLY. We do not care if these generators produce good random
128 numbers. We only want to help eliminate the human factors & perceptions
131 The randreseed64 process scrambles all 64 bit chunks of a seed, by
132 mapping [0,2^64) into [0,2^64). This map is one-to-one and onto.
133 Mapping is performed using a linear congruence generator of the form:
135 X1 <-- (a*X0 + c) % m
137 with the exception that:
139 0 ==> 0 (so that srand(0) acts as default)
141 while maintaining a 1-to-1 and onto map.
143 The randreseed64 constants 'a' and 'c' based on the linear
144 congruential generators found in:
146 "The Art of Computer Programming - Seminumerical Algorithms"
147 by Knuth, Vol 2, 2nd edition (1981), Section 3.6, pages 170-171.
149 We will select the randreseed64 multiplier 'a' such that:
151 a mod 8 == 5 (based on note iii)
152 0.01*m < a < 0.99*m (based on note iv)
153 0.01*2^64 < a < 0.99*2^64
154 a is prime (help keep the generators independent)
156 The choice of the randreseed64 adder 'c' is considered immaterial
157 according (based in note v). Knuth suggests 'c==1' or 'c==a'. We
158 elect to select 'c' using the same process as we used to select
159 'a'. The choice is 'immaterial' after all, and as long as:
161 gcd(c, m) == 1 (based on note v)
163 gcd(a, c) == 1 (adders & multipliers will be more independent)
165 The values 'a' and 'c for randreseed64 are taken from the Rand book
166 of numbers. Because m=2^64 is 20 decimal digits long, we will
167 search the Rand book of numbers 20 at a time. We will skip any of
168 the 100 values that were used to initialize the subtractive 100
169 generators. The values obtained from the Rand book are:
171 a = 6316878969928993981
172 c = 1363042948800878693
174 As we stated before, we must map 0 ==> 0 so that srand(0) does the
175 default thing. The randreseed64 would normally map as follows:
177 0 ==> 1363042948800878693 (0 ==> c)
179 To overcome this, and preserve the 1-to-1 and onto map, we force:
182 10239951819489363767 ==> 1363042948800878693
184 One might object to the complexity of the seed scramble/mapping via
185 the randreseed64 process. But Calling srand(0) with the randreseed64
186 process would be the same as calling srand(10239951819489363767)
187 without it. No extra security is gained or reduced by using the
188 randreseed64 process. The meaning of seeds are exchanged, but not
189 lost or favored (used by more than one input seed).
191 The randreseed64 process does not reduce the security of the rand
192 generator. Every seed is converted into a different unique seed.
193 No seed is ignored or favored.
195 The truly paranoid might suggest that my claims in the MAGIC NUMBERS
196 section are a lie intended to entrap people. Well they are not, but
197 if you that paranoid why would you use a non-cryprographically strong
198 pseudo-random number generator in the first place? You would be
199 better off using the random() builtin function.
201 The two constants that were picked from the Rand Book of Random Numbers
202 The random numbers from the Rand Book of Random Numbers can be
203 verified by anyone who obtains the book. As these numbers were
204 created before I (Landon Curt Noll) was born (you can look up
205 my birth record if you want), I claim to have no possible influence
208 There is a very slight chance that the electronic copy of the
209 Rand Book of Random Numbers that I was given access to differs
210 from the printed text. I am willing to provide access to this
211 electronic copy should anyone wants to compare it to the printed text.
213 When using the s100 generator, one may select your own 100 subtractive
218 and avoid using my magic numbers. The randreseed64 process is NOT
219 applied to the matrix values. Of course, you must pick good subtractive
223 > print srand(0), rand(), rand(), rand()
224 RAND state 2298441576805697181 3498508396312845423 5031615567549397476
226 > print rand(123), rand(123), rand(123), rand(123), rand(123), rand(123)
229 > print rand(2,12), rand(2^50,3^50), rand(0,2), rand(-400000, 120000)
230 2 658186291252503497642116 1 -324097
236 void zrand(long cnt, ZVALUE *res)
237 void zrandrange(ZVALUE low, ZVALUE high, ZVALUE *res)
241 seed, srand, randbit, isrand, random, srandom, israndom