Many best in class pseudo random generators grouped into one simple library.
PyRandLib is distributed under the MIT license for its largest use. If you decide to use this library, please add the copyright notice to your software as stated in the LICENSE file.
Copyright (c) 2016-2022 Philippe Schmouker, <ph.schmouker (at) gmail.com>
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
This library implements some of the best-in-class pseudo random generators as evaluated by Pierre L'Ecuyer and Richard Simard in their famous paper "TestU01: A C library for empirical testing of random number generators" (ACM Trans. Math. Softw. Vol. 33 N.4, August 2007 - see reference [1]. The reader will take benefit reading L'Ecuyer & Simard's paper.
Each of the Pseudo Random Generator (PRG) implemented in PyRandLib is self documented. Names of classes directly refer to the kind of PRG they implem- ent augmented with some number characterizing their periodicity. All of their randomness characteristics are explained in every related module.
The Mersenne twister PRG proposed by Matsumoto and Nishimura - see [5] - is the most widely used PRG. The Random class of module random in Python implements this PRG. It is also implemented in C++ and Java standard libraries for instance.
It offers a very good period (2^19937, i.e. about 4.3e6001). Unfortunately, this PRG is a little bit long to compute (up to 3 times than LCGs, 60% more than LFibs and a little bit less than MRGs, see below at section 'Architecture overview'). Moreover, it fails four of the hardest TestU01 tests. You can still use it as your preferred PRG but PyRandLib implements many other PRGs which are either far faster or far better in terms of generated pseudo-randomness than the Mersenne twister PRG.
Currently, the only way to install PyRandLib is to download the .zip or .tar.gz archive, then to directly put sub-directory 'PyRandLib' from archive into directory 'site-packages', in the main directory 'Lib' of your Python environment. See https://schmouk.github.io/PyRandLib/ for an easy access to download versions or click on tab releases on home page of GitHub repository.
A distribution version (to be installed via pip or easy-install in cmd tool or in console) is to come.
In [1], every known PRG at the time of the editing has been tested according to three different sets of tests:
- small crush is a small set of simple tests that quickly tests some of the expected characteristics for a pretty good PRG;
- crush is a bigger set of tests that test more deeply expected random characteristics;
- big crush is the ultimate set of difficult tests that any GOOD PRG should definitively pass.
We give you here below a copy of the resulting table for the PRGs that have been implemented in PyRandLib plus the Mersenne twister one which is not implemented in PyRabdLib, as provided in [1].
| PyRabndLib class | TU01 generator name | Memory Usage | Period | time-32bits | time-64 bits | SmallCrush fails | Crush fails | BigCrush fails |
|---|---|---|---|---|---|---|---|---|
| FastRand32 | LCG(2^32, 69069, 1) | 1 x 4-bytes | 2^32 | 3.20 | 0.67 | 11 | 106 | too many |
| FastRand63 | LCG(2^63, 9219741426499971445, 1) | 2 x 4-bytes | 2^63 | 4.20 | 0.75 | 0 | 5 | 7 |
| MRGRand287 | Marsa-LFIB4 | 256 x 4-bytes | 2^287 | 3.40 | 0.8 | 0 | 0 | 0 |
| MRGRand1457 | DX-47-3 | 47 x 4-bytes | 2^1457 | n.a. | 1.4 | 0 | 0 | 0 |
| MRGRand49507 | DX-1597-2-7 | 1,597 x 4-bytes | 2^49507 | n.a. | 1.4 | 0 | 0 | 0 |
| LFibRand78 | LFib(2^64, 17, 5, +) | 34 x 4-bytes | 2^78 | n.a. | 1.1 | 0 | 0 | 0 |
| LFibRand116 | LFib(2^64, 55, 24, +) | 110 x 4-bytes | 2^116 | n.a. | 1.0 | 0 | 0 | 0 |
| LFibRand668 | LFib(2^64, 607, 273, +) | 1,214 x 4-bytes | 2^668 | n.a. | 0.9 | 0 | 0 | 0 |
| LFibRand1340 | LFib(2^64, 1279, 861, +) | 2,558 x 4-bytes | 2^1340 | n.a. | 0.9 | 0 | 0 | 0 |
| Mersenne twister | MT19937 | 6 x 4-bytes | 2^19937 | 4.30 | 1.6 | 0 | 2 | 2 |
Current implementation of PyRandLib uses Python 3.x with no Cython version.
It has been tested with Python 3.8 but should run with all of Python 3.
Note 1: PyRandLib version 1.1 and below should work with all versions of Python 3. In version 1.2, we have added underscores in numerical constants for the better readability of the code. This feature has been introduced in Python 3.6. If you want to use PyRandLib version 1.2 or above with Python 3.5 or below, removing these underscores should be sufficient to have the library running correctly.
Note 2: no version or PyRandLib will ever be provided for Python 2 which is a no more maintained version of the Python language.
Note 3: a Cython version of PyRandLib might be delivered in a next release. Up today, no date is planned for this.
This is available starting at version 1.2 of PyRandLib. The call operator
(i.e., '()') gets a new signature which is still backward compatible with
previous versions of this library. Its new use is described here below. The
implementation code can be found in class BaseRandom, in module
baserandom.py.
from fastrand63 import FastRand63
rand = FastRand63()
# prints a float random value ranging in [0.0, 1.0]
print( rand() )
# prints an integer random value ranging in [0, 5]
print( rand(5) )
# prints a float random value ranging in [0.0, 20.0]
print( rand(20.0) )
# prints a list of 10 integer values each ranging in [0, 5]
print( rand(5, 10) )
# prints a list of 10 float values each ranging in [0.0, 1.0]
print( rand(times=10) )
# prints a list of 4 random values ranging respectively in
# [0, 5], [0.0, 50.0], [0.0, 500.0] and [0, 5000]
print( rand(5, 50.0, 500.0, 5000) )
# a more complex call which prints something like:
# [ [3, 11.64307079016269, 127.65395855782158, 4206, [2, 0, 1, 4, 4, 1, 2, 0]],
# [2, 34.22526698212995, 242.54183578253426, 2204, [5, 3, 5, 4, 2, 0, 1, 3]],
# [0, 17.77303802057933, 417.70662295909983, 559, [4, 1, 5, 0, 5, 3, 0, 5]] ]
print( rand( (5, 50.0, 500.0, 5000, [5]*8), times=3 ) )
Each of the implemented PRG is described in an independent module. The name of the module is directly related to the name of the related class.
BaseRandom is the base class for every implemented PRG in library PyRandLib. It inherits from the Python built-in class random.Random. It aims at providing simple common behavior for all PRG classes of the library, the most noticeable one being the 'callable' nature of every implemented PRGs.
Inheriting from the Python built-in class random.Random, BaseRandom provides access to many useful distribution functions as described in later section Inherited Distribution Functions.
Furthermore, every inheriting class may override methods:
- random(),
- seed(),
- getrandbits(k),
- getstate() and
- setstate().
This lets inheriting classes implement the PRGs related core methods.
Notice: starting at PyRandLib 1.2.0 a new signature is available with this base class. See previous section 'New in release 1.2' for full explanations.
FastRand32 implements a Linear Congruential Generator dedicated to
32-bits calculations with very short period (about 4.3e+09) but very short
time computation.
LCG models evaluate pseudo-random numbers suites x(i) as a simple mathematical function of x(i-1):
x(i) = ( a * x(i-1) + c ) mod m
The implementation of FastRand32 is based on (a=69069, c=1) since these two values have evaluated to be the 'best' ones for LCGs within TestU01 while m = 2^32.
Results are nevertheless considered to be poor as stated in the evaluation
done by Pierre L'Ecuyer and Richard Simard. Therefore, it is not
recommended to use such pseudo-random numbers generators for serious
simulation applications.
See FastRand63 for a 2^63 (i.e. about 9.2e+18) period LC-Generator with low computation time and 'better' randomness characteristics.
FastRand63 implements a Linear Congruential Generator dedicated to
63-bits calculations with a short period (about 9.2e+18) and very short
time computation.
LCG models evaluate pseudo-random numbers suites x(i) as a simple mathematical function of x(i-1):
x(i) = ( a * x(i-1) + c ) mod m
The implementation of this LCG 63-bits model is based on (a=9219741426499971445, c=1) since these two values have evaluated to be the 'best' ones for LCGs within TestU01 while m = 2^63.
Results are nevertheless considered to be poor as stated in the evaluation
done by Pierre L'Ecuyer and Richard Simard. Therefore, it is not
recommended to use such pseudo-random numbers generators for serious
simulation applications, even if FastRandom63 fails on very far less tests
than does FastRandom32.
See FastRand32 for a 2^32 period (i.e. about 4.3e+09) LC-Generator with 25% lower computation time.
MRGRand287 implements a fast 32-bits Multiple Recursive Generator (MRG) with a long period (2^287, i.e. 2.49e+86) and low computation time (about twice the computation time of above LCGs) but 256 integers memory consumption.
Multiple Recursive Generators (MRGs) use recurrence to evaluate
pseudo-random numbers suites. For 2 to more different values of k,
recurrence is of the form:
x(i) = A * SUM[ x(i-k) ] mod M
MRGs offer very large periods with the best known results in the evaluation
of their randomness, as evaluated by Pierre L'Ecuyer and Richard Simard. It
is therefore strongly recommended to use such pseudo-random numbers
generators rather than LCG ones for serious simulation applications.
The implementation of this MRG 32-bits model is finally based on a Lagged
Fibonacci generator (LFIB), the Marsa-LFIB4 one.
Lagged Fibonacci generators LFib( m, r, k, op) use the recurrence
x(i) = ( x(i-r) op (x(i-k) ) mod m
where op is an operation that can be + (addition), - (substraction), * (multiplication), ^(bitwise exclusive-or).
With the + or - operation, such generators are true MRGs. They offer very large periods with the best known results in the evaluation of their randomness, as evaluated by Pierre L'Ecuyer and Richard Simard in their paper.
The Marsa-LIBF4 version, i.e. MRGRand287 implementation, uses the recurrence:
x(i) = ( x(i-55) + x(i-119) + x(i-179) + x(i-256) ) mod 2^32
MRGRand1457 implements a fast 31-bits Multiple Recursive Generator with a longer period than MRGRan287 (2^1457 vs. 2^287, i.e. 4.0e+438 vs. 2.5e+86) and 80 % more computation time but with much less memory space consumption (47 vs. 256 integers).
The implementation of this MRG 31-bits model is based on DX-47-3
pseudo-random generator proposed by Deng and Lin, see [2]. The DX-47-3
version uses the recurrence:
x(i) = (2^26+2^19) * ( x(i-1) + x(i-24) + x(i-47) ) mod (2^31-1)
MRGRand49507 implements a fast 31-bits Multiple Recursive Generator with the longer period of all of the PRGs that are implemented if PyRandLib (2^49507, i.e. 1.2e+14903) with low computation time also (same as for MRGRand287) but use of much more memory space (1597 integers).
The implementation of this MRG 31-bits model is based on the 'DX-1597-2-7' MRG proposed by Deng, see [3]. It uses the recurrence:
x(i) = (-2^25-2^7) * ( x(i-7) + x(i-1597) ) mod (2^31-1)
LFibRand78 implements a fast 64-bits Lagged Fibonacci generator (LFib). Lagged Fibonacci generators LFib( m, r, k, op) use the recurrence
x(i) = ( x(i-r) op (x(i-k) ) mod m
where op is an operation that can be + (addition), - (substraction), * (multiplication), ^(bitwise exclusive-or).
With the + or - operation, such generators are MRGs. They offer very large periods with the best known results in the evaluation of their randomness, as stated in the evaluation done by Pierre L'Ecuyer and Richard Simard while offering very low computation times.
The implementation of LFibRand78 is based on a Lagged Fibonacci generator (LFib) which uses the recurrence:
x(i) = ( x(i-5) + x(i-17) ) mod 2^64
It offers a period of about 2^78 - i.e. 3.0e+23 - with low computation time due to the use of a 2^64 modulo (less than twice the computation time of LCGs) and low memory consumption (17 integers).
Please notice that the TestUO1 article states that the operator should be '*' while George Marsaglia in its original article [4] used the operator '+'. We've implemented in PyRandLib the original operator '+'.
LFibRand116 implements an LFib 64-bits generator proposed by George Marsaglia in [4]. This PRG uses the recurrence
x(i) = ( x(i-24) + x(i-55) ) mod 2^64
It offers a period of about 2^116 - i.e. 8.3e+34 - with low computation time due to the use of a 2^64 modulo (less than twice the computation time of LCGs) and some memory consumption (55 integers).
Please notice that the TestUO1 article states that the operator should be '*' while George Marsaglia in its original article [4] used the operator '+'. We've implemented in PyRandLib the original operator '+'.
LFibRand668 implements an LFib 64-bits generator proposed by George Marsaglia in [4]. This PRG uses the recurrence
x(i) = ( x(i-273) + x(i-607) ) mod 2^64
It offers a period of about 2^668 - i.e. 1.2e+201 - with low computation time due to the use of a 2^64 modulo (less than twice the computation time of LCGs) and much memory consumption (607 integers).
Please notice that the TestUO1 article states that the operator should be '*' while George Marsaglia in its original article [4] used the operator '+'. We've implemented in PyRandLib the original operator '+'.
LFibRand1340 implements an LFib 64-bits generator proposed by George Marsaglia in [4]. This PRG uses the recurrence
x(i) = ( x(i-861) + x(i-1279) ) mod 2^64
It offers a period of about 2^1340 - i.e. 2.4e+403 - with low computation time due to the use of a 2^64 modulo (less than twice the computation time of LCGs) and much more memory consumption (1279 integers).
Please notice that the TestUO1 article states that the operator should be '*' while George Marsaglia in its original article [4] used the operator '+'. We've implemented in PyRandLib the original operator '+'.
(some of next explanation may be free to exact copy of Python 3.6 documentation. See https://docs.python.org/3.6/library/random.html?highlight=random#module-random)
Since the base class BaseRandom inherits from the built-in class random.Random, every PRG class of PyRandLib gets automatic access to the next distribution and generic methods:
betavariate(self, alpha, beta)
Beta distribution.
Conditions on the parameters are alpha > 0 and beta > 0. Returned values range between 0 and 1.
choice(self, seq)
Chooses a random element from a non-empty sequence. 'seq' has to be non empty.
choices(population, weights=None, *, cum_weights=None, k=1)
Returns a k sized list of elements chosen from the population with
replacement. If the population is empty, raises IndexError.
If a weights sequence is specified, selections are made according to the relative weights. Alternatively, if a cum_weights sequence is given, the selections are made according to the cumulative weights (perhaps computed using itertools.accumulate()). For example, the relative weights [10, 5, 30, 5] are equivalent to the cumulative weights [10, 15, 45, 50]. Internally, the relative weights are converted to cumulative weights before making selections, so supplying the cumulative weights saves work.
If neither weights nor cum_weights are specified, selections are made with equal probability. If a weights sequence is supplied, it must be the same length as the population sequence. It is a TypeError to specify both weights and cum_weights.
The weights or cum_weights can use any numeric type that interoperates with the float values returned by random() (that includes integers, floats, and fractions but excludes decimals).
Notice: 'choices' has been provided since Python 3.6. It should be implemented for older versions.
expovariate(self, lambd)
Exponential distribution.
lambd is 1.0 divided by the desired mean. It should be nonzero. (The parameter should be called "lambda", but this is a reserved word in Python.) Returned values range from 0 to positive infinity if lambd is positive, and from negative infinity to 0 if lambd is negative.
gammavariate(self, alpha, beta)
Gamma distribution. Not the gamma function!
Conditions on the parameters are alpha > 0 and beta > 0.
gauss(self, mu, sigma)
Gaussian distribution.
mu is the mean, and sigma is the standard deviation. This is slightly faster than the normalvariate() function.
Not thread-safe without a lock around calls.
getrandbits(self, k)
Returns a Python integer with k random bits. Inheriting generators may also provide it as an optional part of their API. When available, getrandbits() enables randrange() to handle arbitrarily large ranges.
getstate(self)
Returns internal state; can be passed to setstate() later.
lognormvariate(self, mu, sigma)
Log normal distribution.
If you take the natural logarithm of this distribution, you'll get a normal distribution with mean mu and standard deviation sigma. mu can have any value, and sigma must be greater than zero.
normalvariate(self, mu, sigma)
Normal distribution.
mu is the mean, and sigma is the standard deviation. See method gauss() for a faster but not thread-safe equivalent.
paretovariate(self, alpha)
Pareto distribution. alpha is the shape parameter.
randint(self, a, b)
Returns a random integer in range [a, b], including both end points.
randrange(self, stop)
randrange(self, start, stop=None, step=1)
Returns a randomly selected element from range(start, stop, step). This is equivalent to choice( range(start, stop, step) ) without building a range object.
The positional argument pattern matches that of range(). Keyword arguments should not be used because the function may use them in unexpected ways.
sample(self, population, k)
Chooses k unique random elements from a population sequence or set.
Returns a new list containing elements from the population while leaving the original population unchanged. The resulting list is in selection order so that all sub-slices will also be valid random samples. This allows raffle winners (the sample) to be partitioned into grand prize and second place winners (the subslices).
Members of the population need not be hashable or unique. If the population contains repeats, then each occurrence is a possible selection in the sample.
To choose a sample in a range of integers, use range as an argument. This is especially fast and space efficient for sampling from a large population: sample(range(10000000), 60)
seed(self, a=None, version=2)
Initialize internal state from hashable object.
None or no argument seeds from current time or from an operating system specific randomness source if available.
For version 2 (the default), all of the bits are used if a is a str, bytes, or bytearray. For version 1, the hash() of a is used instead.
If a is an int, all bits are used.
setstate(self, state)
Restores internal state from object returned by getstate().
shuffle(self, x, random=None)
Shuffle the sequence x in place. Returns None.
The optional argument random is a 0-argument function returning a random float in [0.0, 1.0); by default, this is the function random().
To shuffle an immutable sequence and return a new shuffled list, use sample(x, k=len(x)) instead.
Note that even for small len(x), the total number of permutations of x can quickly grow larger than the period of most random number generators. This implies that most permutations of a long sequence can never be generated. For example, a sequence of length 2080 is the largest that can fit within the period of the Mersenne Twister random number generator.
triangular(self, low=0.0, high=1.0, mode=None)
Triangular distribution.
Continuous distribution bounded by given lower and upper limits, and having a given mode value in-between. Returns a random floating point number N such that low <= N <= high and with the specified mode between those bounds. The low and high bounds default to zero and one. The mode argument defaults to the midpoint between the bounds, giving a symmetric distribution.
http://en.wikipedia.org/wiki/Triangular_distribution
uniform(self, a, b)
Gets a random number in the range [a, b) or [a, b] depending on rounding.
vonmisesvariate(self, mu, kappa)
Circular data distribution.
mu is the mean angle, expressed in radians between 0 and 2pi, and kappa is the concentration parameter, which must be greater than or equal to zero. If kappa is equal to zero, this distribution reduces to a uniform random angle over the range 0 to 2pi.
weibullvariate(self, alpha, beta)
Weibull distribution.
alpha is the scale parameter and beta is the shape parameter.
[1] Pierre L'Ecuyer and Richard Simard. 2007. TestU01: A C library for empirical testing of random number generators. ACM Transaction on Mathematical Software, Vol.33 N.4, Article 22 (August 2007), 40 pages. DOI: http://dx.doi.org/10.1145/1268776.1268777
BibTex: @article{L'Ecuyer:2007:TCL:1268776.1268777, author = {L'Ecuyer, Pierre and Simard, Richard}, title = {TestU01: A C Library for Empirical Testing of Random Number Generators}, journal = {ACM Trans. Math. Softw.}, issue_date = {August 2007}, volume = {33}, number = {4}, month = aug, year = {2007}, issn = {0098-3500}, pages = {22:1--22:40}, articleno = {22}, numpages = {40}, url = {http://doi.acm.org/10.1145/1268776.1268777}, doi = {10.1145/1268776.1268777}, acmid = {1268777}, publisher = {ACM}, address = {New York, NY, USA}, keywords = {Statistical software, random number generators, random number tests, statistical test}, }
[2] Lih-Yuan Deng & Dennis K. J. Lin. 2000. Random number generation for the new century. The American Statistician Vol.54, N.2, pp. 145–150.
BibTex: @article{doi:10.1080/00031305.2000.10474528, author = { Lih-Yuan Deng and Dennis K. J. Lin }, title = {Random Number Generation for the New Century}, journal = {The American Statistician}, volume = {54}, number = {2}, pages = {145-150}, year = {2000}, doi = {10.1080/00031305.2000.10474528}, URL = {ttp://amstat.tandfonline.com/doi/abs/10.1080/00031305.2000.10474528}, eprint = {http://amstat.tandfonline.com/doi/pdf/10.1080/00031305.2000.10474528} }
[3] Lih-Yuan Deng. 2005. Efficient and portable multiple recursive generators of large order. ACM Transactions on Modeling and Computer. Simulation 15:1.
[4] Georges Marsaglia. 1985. A current view of random number generators. In Computer Science and Statistics, Sixteenth Symposium on the Interface. Elsevier Science Publishers, North-Holland, Amsterdam, 1985, The Netherlands. pp. 3–10.
[5] Makoto Matsumoto and Takuji Nishimura. 1998.
Mersenne twister: A 623-dimensionally equidistributed uniform pseudo-random number generator.
In ACM Transactions on Modeling and Computer Simulation (TOMACS) - Special issue on uniform random number generation.
Vol.8 N.1, Jan. 1998, pp. 3-30.