.\"
.\" $Id: mtwist.3,v 1.7 2010-06-24 01:53:59-07 geoff Exp $
.\"
.\" $Log: mtwist.3,v $
.\" Revision 1.7  2010-06-24 01:53:59-07  geoff
.\" Change all documented declarations to use types from stdint.h.  Fix
.\" some restriction descriptions.  Remove bugs that are no longer bugs.
.\"
.\" Revision 1.6  2007-10-26 00:21:06-07  geoff
.\" Document the new mt_u32bit_t type (barely).
.\"
.\" Revision 1.5  2002/10/30 07:39:53  geoff
.\" Document the new seeding routines.
.\"
.\" Revision 1.4  2001/06/20 08:15:51  geoff
.\" Correct the documentation of the generator's period.
.\"
.\" Revision 1.3  2001/06/19 00:43:01  geoff
.\" Document the lack of a newline in the << operator
.\"
.\" Revision 1.2  2001/06/18 10:09:24  geoff
.\" Fix the manual section.
.\"
.\" Revision 1.1  2001/06/16 21:20:31  geoff
.\" Initial revision
.\"
.\" 
.TH mtwist 3 "June 14, 2001" "" "Linux Programmer's Manual"
.SH NAME
mts_seed32new, mts_seed32, mts_seedfull, mts_seed, mts_goodseed, mts_bestseed,
mts_savestate, mts_loadstate, mt_seed32new, mt_seed32, mt_seedfull, mt_seed,
mt_goodseed, mt_bestseed, mt_getstate, mt_savestate, mt_loadstate,
mts_lrand, mts_llrand, mts_drand, mts_ldrand, mt_lrand, mt_llrand,
mt_drand, mt_ldrand,
mt_prng \- generate uniformly distributed pseudo-random numbers
.SH SYNOPSIS
.nf
.IR "#defines" " (see below)"
.br
.B
#include "mtwist.h"
.sp
C interface:
.sp
.BI "void mts_seed32(mt_state* " state ", uint32_t " seed ");"
.sp
.BI "void mts_seed32new(mt_state* " state ", uint32_t " seed ");"
.sp
.BI "void mts_seedfull(mt_state* " state ","
.BI "                  uint32_t " seeds "[MT_STATE_SIZE]);"
.sp
.BI "void mts_seed(mt_state* " state ");"
.sp
.BI "void mts_goodseed(mt_state* " state ");"
.sp
.BI "void mts_bestseed(mt_state* " state ");"
.sp
.BI "int mts_savestate(FILE* " statefile ", mt_state* " state ");"
.sp
.BI "int mts_loadstate(FILE* " statefile ", mt_state* " state ");"
.sp
.BI "void mt_seed32(uint32_t " seed ");"
.sp
.BI "void mt_seed32new(uint32_t " seed ");"
.sp
.BI "void mt_seedfull(uint32_t " seeds "[MT_STATE_SIZE]);"
.sp
.B void mt_seed(void);
.sp
.B void mt_goodseed(void);
.sp
.B void mt_bestseed(void);
.sp
.B mt_state* mt_getstate(void);
.sp
.BI "int mt_savestate(FILE* " statefile ");"
.sp
.BI "int mt_loadstate(FILE* " statefile ");"
.sp
.BI "uint32_t mts_lrand(mt_state* " state ");"
.sp
.BI "uint64_t mts_llrand(mt_state* " state ");"
.sp
.BI "double mts_drand(mt_state* " state ");"
.sp
.BI "double mts_ldrand(mt_state* " state ");"
.sp
.B uint32_t mt_lrand(void);
.sp
.B uint64_t mt_llrand(void);
.sp
.B double mt_drand(void);
.sp
.B double mt_ldrand(void);
.sp
.B "C++ interface:"
.sp
.BI "mt_prng " rng ;
.sp
.BI "mt_prng " rng "(bool " pickseed " = false);"
.sp
.BI "mt_prng " rng "(uint32_t " seed );
.sp
.BI "mt_prng " rng "(uint32_t " seeds [MT_STATE_SIZE]);
.sp
.BI "void " rng ".seed32(uint32_t " seed ");"
.sp
.BI "void " rng ".seedfull(uint32_t seeds[MT_STATE_SIZE]);"
.sp
.BI "void " rng ".seed();"
.sp
.BI "void " rng ".goodseed();"
.sp
.BI "void " rng ".bestseed();"
.sp
.BI "uint32_t " rng ".lrand();"
.sp
.BI "uint64_t " rng ".llrand();"
.sp
.BI "double " rng ".drand();"
.sp
.BI "double " rng ".ldrand();"
.sp
.BI "double " rng "();"
.sp
.IB "stream" " << " rng ";"
.sp
.IB "stream" " >> " rng ";"
.SH DESCRIPTION
These functions generate pseudo-random numbers using Matsumoto and
Nishimura's Mersenne Twist algorithm (see:
.nf
.sp
        http://www.math.keio.ac.jp/~matumoto/emt.html
.sp
.fi
for full information).
The period of this pseudo random-number generator (PRNG) is 2^19337-1
which is vastly longer than the life of the universe
even if the random numbers are being generated at an impossible rate.
The generator also has excellent statistical properties.
.PP
The
.B mtwist
package assumes a 32-bit machine with a modern C or C++ compiler that
supports inline functions and the
.B inttypes.h
header file.
If these features are not present, the package must be modified.
.PP
All of the PRNG functions work from a
.IR "state vector" ,
which is of type
.B mt_state
in C and type
.B mt_prng
in C++.
The state vector stores everything that the PRNG needs to generate new
numbers in the proper sequence.
By using multiple state vectors, programs can draw random numbers from
independent sequences, which is important in applications such as
simulation (where each independent random variable should be drawn
from its own sequence to avoid unintentional correlations).
.PP
For convenience, the C interface also provides a built-in default
state vector that can be used in simple applications.
The
.BI mt_ xxx
functions use the default state vector to control their behavior,
while the
.BI mts_xxx
functions accept a user-provided state vector.
.PP
In C, a user-provided state vector has the following structure:
.PP
.nf
#define MT_STATE_SIZE 624

typedef struct {
.in +8
uint32_t statevec[MT_STATE_SIZE];
.in +16
/* Vector holding current state */
.in -16
int stateptr;   /* Next state entry to be used */
int initialized;
.in +16
/* NZ if state has been initialized */
.in -24
} mt_state;
.fi
.PP
An uninitialized PRNG is indicated by zeros in
.I both
.B stateptr
and
.BR initialized .
It is the programmer's responsibility to ensure that these fields are
zero before calling any of the
.BI mts_xxx
functions.
.PP
It is occasionally useful to directly access the default state vector, so
.B mt_getstate
will return a pointer to the default state.
.PP
In both C and C++, the functionality is divided into two categories:
seeding and pseudorandom-number generation.
If one of the generation functions is called on an unseeded generator,
a default seed (specified by Matsumoto and Nishimura) will be used.
Usually, the programmer will wish to override the default seed and
choose a more appropriate one.
The simplest way to seed a PRNG is by calling one of the
.B *seed32new
functions.
This will invoke Matsumoto and Nishimura's revised Knuth-style seed
generator.
.PP
The
.B *seed32
functions
will invoke Matsumoto and Nishimura's original Knuth-style seed
generator, which is now deprecated.
In C++, the same effect can be achieved by passing a 32-bit
.RB ( "uint32_t" )
seed to the constructor.
The original 32-bit seeder did not work correctly if the seed was zero,
so in that
case the default seed of 4357 will be substituted.
The original seeder is still supported so that older software will
continue to work in the same fashion without changes.
.PP
The
.B *seed32new
and
.B *seed32
functions are simple to use, but they have the drawback that only 4
billion distinct pseudorandom sequences can be generated using them.
To allow access to sequences beginning anywhere in the entire space of
possibilities, the
.B *seedfull
functions can be passed an initial state vector of 624 32-bit numbers,
or a C++ PRNG can be constructed with a 624-element array as an
argument.
The initialization vector must contain at least one nonzero value;
if this rule is violated, the program will be aborted (unfortunately
without a diagnostic message due to C/C++ portability issues).
.PP
The
.BR *seed32new ,
.BR *seed32 ,
and
.B *seedfull
functions allow fixed, reproducible seeds, which is useful for
simulation and experimentation.
For game-like applications, non-reproducible seeds are usually more
appropriate.
The
.BR mts_seed ,
.BR mt_seed ,
and
.B seed
functions use the system time to generate an argument to the
.B *seed32new
functions to satisfy this need.
The microseconds portion of the time is included in the seed to
enhance the probability that two programs will generate different
random sequences.
.PP
Since the various "plain"
.B seed
functions are also somewhat limited in the variety they can produce,
two other functions are available on systems that have support for the
.B /dev/random
device.
The
.B *goodseed
functions attempt to use
.B /dev/urandom
to get truly random values for use with
.BR *seedfull .
If
.B /dev/urandom
isn't available, these functions fall back to calling the equivalent "plain"
.B seed
function.
C++ programmers can also invoke
.B goodseed
at construction time by passing an argument of
.B true
to the constructor.
.PP
For the most random seed possible, the
.B *bestseed
functions attempt to use
.B /dev/random
to acquire values for
.BR *seedfull ,
falling back to
.B *seed
if
.B /dev/random
is unavailable.
The disadvantage of these functions is that it usually takes a
significant amount of (wall-clock) time before
.B /dev/random
can produce enough entropy to provide a seed.
Therefore, it is nearly always better to stick with the
.B *goodseed
functions.
.PP
Finally, it is often useful to be able to save and restore the PRNG
state for later use.
In C, the functions
.B *savestate
.B *loadstate
will save the current state into an open
.B stdio
.B FILE
as a single long line (in ASCII)
and later restore it such that the restored PRNG will pick up where
the saved one left off.
In C++, the same effect can be achieved by writing to or reading from
a C++
.B stream
using the usual
.B "<<"
and
.B ">>"
operators.
As with all well-behaved C++ types, the
.B "<<"
operator does not add a newline after the saved state.
.PP
Once a generator has been seeded,
uniformly distributed pseudorandom numbers can be produced in several
formats.
(The functions in the
.IR randistrs (3)
library can be used to produce other statistical distributions.)
The
.B *lrand
and
.B *llrand
generate 32-bit and 64-bit random integers uniformly distributed
between 0 and the maximum unsigned value.
(The
.B *llrand
functions are only available on machines that support a 64-bit
data type.
The
.B *drand
functions generate a double-precision number in the range [0,1)
(i.e., 0 is a possible value but 1 is not).
The number generated by
.B *drand
has 32 bits of precision.
For convenience, the C++ interface also defines a function operator
that returns the same result as
.BR drand ,
so that a PRNG can be called as if it were a function.
For applications that demand increased precision, the
.B *ldrand
functions generate a double-precision number in [0,1) with up to 64
bits of precision (usually 52 bits).
.SH "SEE ALSO"
.BR randistrs "(3), " drand48 "(3), " rand "(3), " random (3)