Title: | Fast Pseudo Random Number Generators |
---|---|
Description: | Several fast random number generators are provided as C++ header only libraries: The PCG family by O'Neill (2014 <https://www.cs.hmc.edu/tr/hmc-cs-2014-0905.pdf>) as well as the Xoroshiro / Xoshiro family by Blackman and Vigna (2021 <doi:10.1145/3460772>). In addition fast functions for generating random numbers according to a uniform, normal and exponential distribution are included. The latter two use the Ziggurat algorithm originally proposed by Marsaglia and Tsang (2000, <doi:10.18637/jss.v005.i08>). The fast sampling methods support unweighted sampling both with and without replacement. These functions are exported to R and as a C++ interface and are enabled for use with the default 64 bit generator from the PCG family, Xoroshiro128+/++/** and Xoshiro256+/++/** as well as the 64 bit version of the 20 rounds Threefry engine (Salmon et al., 2011, <doi:10.1145/2063384.2063405>) as provided by the package 'sitmo'. |
Authors: | Ralf Stubner [aut, cre] |
Maintainer: | Ralf Stubner <[email protected]> |
License: | AGPL-3 |
Version: | 0.4.1.1 |
Built: | 2024-09-06 19:13:45 UTC |
Source: | https://github.com/daqana/dqrng |
Several fast random number generators are provided as C++ header only libraries: The PCG family by O'Neill (2014 https://www.cs.hmc.edu/tr/hmc-cs-2014-0905.pdf) as well as the Xoroshiro / Xoshiro family by Blackman and Vigna (2021 doi:10.1145/3460772). In addition fast functions for generating random numbers according to a uniform, normal and exponential distribution are included. The latter two use the Ziggurat algorithm originally proposed by Marsaglia and Tsang (2000, doi:10.18637/jss.v005.i08). The fast sampling methods support unweighted sampling both with and without replacement. These functions are exported to R and as a C++ interface and are enabled for use with the default 64 bit generator from the PCG family, Xoroshiro128+/++/** and Xoshiro256+/++/** as well as the 64 bit version of the 20 rounds Threefry engine (Salmon et al., 2011, doi:10.1145/2063384.2063405) as provided by the package 'sitmo'.
Maintainer: Ralf Stubner [email protected] (ORCID)
Other contributors:
daqana GmbH [copyright holder]
David Blackman (Xoroshiro / Xoshiro family) [copyright holder]
Melissa O'Neill [email protected] (PCG family) [copyright holder]
Sebastiano Vigna [email protected] (Xoroshiro / Xoshiro family) [copyright holder]
Aaron Lun [contributor]
Kyle Butts [email protected] [contributor]
Henrik Sloot [contributor]
Philippe Grosjean (ORCID) [contributor]
Useful links:
Report bugs at https://github.com/daqana/dqrng/issues
Multivariate Distributions
dqrmvnorm(n, ...)
dqrmvnorm(n, ...)
n |
number of observations |
... |
forwarded to |
numeric matrix of multivariate normal distributed variables
sigma <- matrix(c(4,2,2,3), ncol=2) x <- dqrmvnorm(n=500, mean=c(1,2), sigma=sigma) colMeans(x) var(x) plot(x)
sigma <- matrix(c(4,2,2,3), ncol=2) x <- dqrmvnorm(n=500, mean=c(1,2), sigma=sigma) colMeans(x) var(x) plot(x)
The dqrng
package provides several fast random number
generators together with fast functions for generating random numbers
according to a uniform, normal and exponential distribution. These
functions are modeled after the base
functions
set.seed
, RNGkind
, runif
,
rnorm
, and rexp
. However, note that the functions
provided here do not accept vector arguments for the number of observations
as well as the parameters describing the distribution functions. Please see
register_methods
if you need this functionality.
dqrrademacher
uses a fast algorithm to generate random
Rademacher variables (-1 and 1 with equal probability). To do so, it
generates a random 64 bit integer and then uses each bit to generate
a 0/1 variable. This generates 64 integers per random number generation.
dqrng_get_state
and dqrng_set_state
can be used to get and set
the RNG's internal state. The character vector should not be manipulated directly.
dqRNGkind(kind, normal_kind = "ignored") dqrng_get_state() dqrng_set_state(state) dqrunif(n, min = 0, max = 1) dqrnorm(n, mean = 0, sd = 1) dqrexp(n, rate = 1) dqrrademacher(n) dqset.seed(seed, stream = NULL)
dqRNGkind(kind, normal_kind = "ignored") dqrng_get_state() dqrng_set_state(state) dqrunif(n, min = 0, max = 1) dqrnorm(n, mean = 0, sd = 1) dqrexp(n, rate = 1) dqrrademacher(n) dqset.seed(seed, stream = NULL)
kind |
string specifying the RNG (see details) |
normal_kind |
ignored; included for compatibility with |
state |
character vector representation of the RNG's internal state |
n |
number of observations |
min |
lower limit of the uniform distribution |
max |
upper limit of the uniform distribution |
mean |
mean value of the normal distribution |
sd |
standard deviation of the normal distribution |
rate |
rate of the exponential distribution |
seed |
integer scalar to seed the random number generator, or an integer vector of length 2 representing a 64-bit seed. Maybe |
stream |
integer used for selecting the RNG stream; either a scalar or a vector of length 2 |
Supported RNG kinds:
The default 64 bit variant from the PCG family developed by Melissa O'Neill. See https://www.pcg-random.org/ for more details.
RNGs developed by David Blackman and Sebastiano Vigna. See https://prng.di.unimi.it/ for more details. The older generators Xoroshiro128+ and Xoshiro256+ should be used only for backwards compatibility.
The 64 bit version of the 20 rounds Threefry engine as
provided by sitmo-package
Xoroshiro128++ is the default since it is fast, small and has good statistical properties.
The functions dqrnorm
and dqrexp
use the Ziggurat algorithm as
provided by boost.random
.
See generateSeedVectors
for rapid generation of integer-vector
seeds that provide 64 bits of entropy. These allow full exploration of
the state space of the 64-bit RNGs provided in this package.
If the provided seed
is NULL
, a seed is generated from R's RNG
without state alteration.
dqrunif
, dqrnorm
, and dqrexp
return a numeric vector
of length n
. dqrrademacher
returns an integer vector of length n
.
dqrng_get_state
returns a character vector representation of the RNG's internal state.
set.seed
, RNGkind
, runif
,
rnorm
, and rexp
library(dqrng) # Set custom RNG. dqRNGkind("Xoshiro256++") # Use an integer scalar to set a seed. dqset.seed(42) # Use integer scalars to set a seed and the stream. dqset.seed(42, 123) # Use an integer vector to set a seed. dqset.seed(c(31311L, 24123423L)) # Use an integer vector to set a seed and a scalar to select the stream. dqset.seed(c(31311L, 24123423L), 123) # Random sampling from distributions. dqrunif(5, min = 2, max = 10) dqrexp(5, rate = 4) dqrnorm(5, mean = 5, sd = 3) # get and restore the state (state <- dqrng_get_state()) dqrunif(5) dqrng_set_state(state) dqrunif(5)
library(dqrng) # Set custom RNG. dqRNGkind("Xoshiro256++") # Use an integer scalar to set a seed. dqset.seed(42) # Use integer scalars to set a seed and the stream. dqset.seed(42, 123) # Use an integer vector to set a seed. dqset.seed(c(31311L, 24123423L)) # Use an integer vector to set a seed and a scalar to select the stream. dqset.seed(c(31311L, 24123423L), 123) # Random sampling from distributions. dqrunif(5, min = 2, max = 10) dqrexp(5, rate = 4) dqrnorm(5, mean = 5, sd = 3) # get and restore the state (state <- dqrng_get_state()) dqrunif(5) dqrng_set_state(state) dqrunif(5)
Unbiased Random Samples and Permutations
dqsample(x, size, replace = FALSE, prob = NULL) dqsample.int(n, size = n, replace = FALSE, prob = NULL)
dqsample(x, size, replace = FALSE, prob = NULL) dqsample.int(n, size = n, replace = FALSE, prob = NULL)
x |
either a vector of one or more elements from which to choose, or a positive integer. |
size |
a non-negative integer giving the number of items to choose. |
replace |
should sampling be with replacement? |
prob |
a vector of probability weights for obtaining the elements of the vector being sampled. |
n |
a positive number, the number of items to choose from. |
vignette("sample", package = "dqrng")
, sample
and sample.int
Generate seed as a integer vector
generateSeedVectors(nseeds, nwords = 2L)
generateSeedVectors(nseeds, nwords = 2L)
nseeds |
Integer scalar, number of seeds to generate. |
nwords |
Integer scalar, number of words to generate per seed. |
Each seed is encoded as an integer vector with the most significant bits at the start of the vector. Each integer vector is converted into an unsigned integer (in C++ or otherwise) by the following procedure:
Start with a sum of zero.
Add the first value of the vector.
Left-shift the sum by 32.
Add the next value of the vector, and repeat.
The aim is to facilitate R-level generation of seeds with sufficient
randomness to cover the entire state space of pseudo-random number
generators that require more than the ~32 bits available in an
int
. It also preserves the integer nature of the seed, thus
avoiding problems with casting double-precision numbers to integers.
It is possible for the seed vector to contain NA_integer_
values. This should not be cause for alarm, as R uses -INT_MAX
to encode missing values in integer vectors.
A list of length n
, where each element is an integer vector that
contains nwords
words (i.e., 32*nwords
bits) of randomness.
Aaron Lun
generateSeedVectors(10, 2) generateSeedVectors(5, 4)
generateSeedVectors(10, 2) generateSeedVectors(5, 4)
The random-number generators (RNG) from this package can be
registered as user-supplied RNG. This way all r<dist>
functions make
use of the provided fast RNGs.
register_methods(kind = c("both", "rng")) restore_methods()
register_methods(kind = c("both", "rng")) restore_methods()
kind |
Which methods should be registered? Either |
Caveats:
While runif
and dqrunif
as well as rnorm
and
dqrnorm
will produce the same results, this is not the case for
rexp
and dqrexp
.
The dqr<dist>
functions are still faster than r<dist>
when many random numbers are generated.
You can use only the RNG from this package using
register_method("rng")
or both the RNG and the Ziggurat method
for normal draws with register_method("both")
. The latter
approach is used by default. Using only the Ziggurat method will give
undefined behavior and is not supported!
Calling dqset.seed(NULL)
re-initializes the RNG from R's RNG.
This no longer makes sense when the RNG has been registered as user-supplied
RNG. In that case set.seed{NULL}
needs to be used.
With R's in-build RNGs one can get access to the internal state using
.Random.seed
. This is not possible here, since the internal state
is a private member of the used C++ classes.
You can automatically register these methods when loading this package by
setting the option dqrng.register_methods
to TRUE
, e.g.
with options(dqrng.register_methods=TRUE)
.
Notes on seeding:
When a user-supplied RNG is registered, it is also seeded from the
previously used RNG. You will therefore get reproducible (but different)
whether you call set.seed()
before or after register_methods()
.
When called with a single integer as argument, both set.seed()
and dqset.seed()
have the same effect. However, dqset.seed()
allows you to call it with two integers thereby supplying 64 bits of
initial state instead of just 32 bits.
Invisibly returns a three-element character vector of the RNG, normal and sample kinds before the call.
RNGkind
and Random.user
register_methods() # set.seed and dqset.seed influence both (dq)runif and (dq)rnorm set.seed(4711); runif(5) set.seed(4711); dqrunif(5) dqset.seed(4711); rnorm(5) dqset.seed(4711); dqrnorm(5) # similarly for other r<dist> functions set.seed(4711); rt(5, 10) dqset.seed(4711); rt(5, 10) # but (dq)rexp give different results set.seed(4711); rexp(5, 10) set.seed(4711); dqrexp(5, 10) restore_methods()
register_methods() # set.seed and dqset.seed influence both (dq)runif and (dq)rnorm set.seed(4711); runif(5) set.seed(4711); dqrunif(5) dqset.seed(4711); rnorm(5) dqset.seed(4711); dqrnorm(5) # similarly for other r<dist> functions set.seed(4711); rt(5, 10) dqset.seed(4711); rt(5, 10) # but (dq)rexp give different results set.seed(4711); rexp(5, 10) set.seed(4711); dqrexp(5, 10) restore_methods()