incorrect distribution of randomR for floating-point numbers #53

shaobo-he · 2019-04-01T23:02:36Z

Hello,

It appears that when the upper bound is very small, floating-point values generated are not correctly distributed.

For example, when the upper bound is 1.0e-45::Float which is the smallest subnormal number of single precision floating-point representation, randomR(0, (1.0e-45::Float)) does not produce 1.0e-45. Furthermore, if the upper bound is 4.0e-45::Float, 6.0e-45 is generated as follows,

filter (>4.0e-45) $ take 10 $ randomRs (0,(4.0e-45::Float)) $ mkStdGen 0
[6.0e-45,6.0e-45]

The text was updated successfully, but these errors were encountered:

cartazio · 2019-04-03T20:23:23Z

Yup. The current / usual naive approach to generating a unit interval doesn’t actually work well. 1) But: most / all unit interval algorithms are half open, they give the interval [0,1) and then things get rescaled , so you should never expect the 1 in your interval unless it’s from rounding... 2) the over shoot is definitely looking like a bug, I’m not sure if it’s a bug in our understanding of floating point near the cut off of normalized floats one. What’s the size of an ULP at that scale? I’d love some help understanding what’s the cause of this bug. Great job finding / reporting this stuff. Thank you for taking the time.

…

On Mon, Apr 1, 2019 at 7:02 PM Shaobo ***@***.***> wrote: Hello, It appears that when the upper bound is very small, floating-point values generated are not correctly distributed. For example, when the upper bound is 1.0e-45::Float which is the smallest subnormal number of single precision floating-point representation, randomR(0, (1.0e-45::Float)) does not produce 1.0e-45. Furthermore, if the upper bound is 4.0e-45::Float, 6.0e-45 is generated as follows, filter (>4.0e-45) $ take 10 $ randomRs (0,(4.0e-45::Float)) $ mkStdGen 0 [6.0e-45,6.0e-45] — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#53>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAAQwjf78XD4lTeVPJLdIdGldWAlV_Asks5vcpANgaJpZM4cWzgv> .

shaobo-he · 2019-04-04T03:17:06Z

I'm not sure how ULP is defined for subnormals but they are multiple of the minimum subnormal number. Let's call it MSN. 1.0e-45 is one MSN and 4.0e-45 is three MSNs. The number is generate using the following formula,

let (coef,g') = random g in
                  (2.0 * (0.5*l + coef * (0.5*h - 0.5*l)), g')

When the lower bound is 0, it becomes, 2.0*(coef*(0.5*h)). And when h is 3 MSNs, 0.5*h rounds to 2 MSNs. So when coef is greater than 0.75, coef*(0.5*h) rounds to 2 MSNs as well, making the entire expression evaluate to 4 MSNs, which is 4.0e-45.

There are also other values that could trigger values being greater than upper bound. I used sbv (https://hackage.haskell.org/package/sbv) to find counterexamples. It works pretty well.

…-randomM Use uniform instead of randomM in example

Use uniform instead of randomM in example

author Alexey Kuleshevich <[email protected]> 1581472095 +0300 committer Leonhard Markert <[email protected]> 1590493894 +0200 This patch is mostly backwards compatible. See "Breaking Changes" below for the full list of backwards incompatible changes. This patch fixes quality and performance issues, addresses additional miscellaneous issues, and introduces a monadic API. Issues addressed ================ Priority issues fixed in this patch: - Title: "The seeds generated by split are not independent" Link: haskell#25 Fixed: changed algorithm to SplitMix, which provides a robust 'split' operation - Title: "Very low throughput" Link: haskell#51 Fixed: see "Performance" below Additional issues addressed in this patch: - Title: "Add Random instances for tuples" Link: haskell#26 Addressed: added 'Uniform' instances for up to 6-tuples - Title: "Add Random instance for Natural" Link: haskell#44 Addressed: added 'UniformRange' instance for 'Natural' - Title: "incorrect distribution of randomR for floating-point numbers" Link: haskell#53 Addressed: see "Regarding floating-point numbers" below - Title: "System/Random.hs:43:1: warning: [-Wtabs]" Link: haskell#55 Fixed: no more tabs - Title: "Why does random for Float and Double produce exactly 24 or 53 bits?" Link: haskell#58 Fixed: see "Regarding floating-point numbers" below - Title: "read :: StdGen fails for strings longer than 6" Link: haskell#59 Addressed: 'StdGen' is no longer an instance of 'Read' Regarding floating-point numbers: with this patch, the relevant instances for 'Float' and 'Double' sample more bits than before but do not sample every possible representable value. The documentation now clearly spells out what this means for users. Quality (issue 25) ================== The algorithm [1] in version 1.1 of this library fails empirical PRNG tests when used to generate "split sequences" as proposed in [3]. SplitMix [2] passes the same tests. This patch changes 'StdGen' to use the SplitMix implementation provided by the splitmix package. Test batteries used: dieharder, TestU1, PractRand. [1]: P. L'Ecuyer, "Efficient and portable combined random number generators". https://doi.org/10.1145/62959.62969 [2]: G. L. Steele, D. Lea, C. H. Flood, "Fast splittable pseudorandom number generators". https://doi.org/10.1145/2714064.2660195 [3]: H. G. Schaathun, "Evaluation of splittable pseudo-random generators". https://doi.org/10.1017/S095679681500012X Performance (issue 51) ====================== The "improvement" column in the following table is a multiplier: the improvement for 'random' for type 'Float' is 1038, so this operation is 1038 times faster with this patch. | Name | Mean (1.1) | Mean (patch) | Improvement| | ----------------------- | ---------- | ------------ | ---------- | | pure/random/Float | 30 | 0.03 | 1038| | pure/random/Double | 52 | 0.03 | 1672| | pure/random/Integer | 43 | 0.33 | 131| | pure/uniform/Word8 | 14 | 0.03 | 422| | pure/uniform/Word16 | 13 | 0.03 | 375| | pure/uniform/Word32 | 21 | 0.03 | 594| | pure/uniform/Word64 | 42 | 0.03 | 1283| | pure/uniform/Word | 44 | 0.03 | 1491| | pure/uniform/Int8 | 15 | 0.03 | 511| | pure/uniform/Int16 | 15 | 0.03 | 507| | pure/uniform/Int32 | 22 | 0.03 | 749| | pure/uniform/Int64 | 44 | 0.03 | 1405| | pure/uniform/Int | 43 | 0.03 | 1512| | pure/uniform/Char | 17 | 0.49 | 35| | pure/uniform/Bool | 18 | 0.03 | 618| | pure/uniform/CChar | 14 | 0.03 | 485| | pure/uniform/CSChar | 14 | 0.03 | 455| | pure/uniform/CUChar | 13 | 0.03 | 448| | pure/uniform/CShort | 14 | 0.03 | 473| | pure/uniform/CUShort | 13 | 0.03 | 457| | pure/uniform/CInt | 21 | 0.03 | 737| | pure/uniform/CUInt | 21 | 0.03 | 742| | pure/uniform/CLong | 43 | 0.03 | 1544| | pure/uniform/CULong | 42 | 0.03 | 1460| | pure/uniform/CPtrdiff | 43 | 0.03 | 1494| | pure/uniform/CSize | 43 | 0.03 | 1475| | pure/uniform/CWchar | 22 | 0.03 | 785| | pure/uniform/CSigAtomic | 21 | 0.03 | 749| | pure/uniform/CLLong | 43 | 0.03 | 1554| | pure/uniform/CULLong | 42 | 0.03 | 1505| | pure/uniform/CIntPtr | 43 | 0.03 | 1476| | pure/uniform/CUIntPtr | 42 | 0.03 | 1463| | pure/uniform/CIntMax | 43 | 0.03 | 1535| | pure/uniform/CUIntMax | 42 | 0.03 | 1493| API changes =========== MonadRandom ----------- This patch adds a class 'MonadRandom': -- | 'MonadRandom' is an interface to monadic pseudo-random number generators. class Monad m => MonadRandom g s m | g m -> s where {-# MINIMAL freezeGen,thawGen,(uniformWord32|uniformWord64) #-} type Frozen g = (f :: Type) | f -> g freezeGen :: g s -> m (Frozen g) thawGen :: Frozen g -> m (g s) uniformWord32 :: g s -> m Word32 -- default implementation in terms of uniformWord64 uniformWord64 :: g s -> m Word64 -- default implementation in terms of uniformWord32 -- plus methods for other word sizes and for byte strings -- all have default implementations so the MINIMAL pragma holds Conceptually, in 'MonadRandom g s m', 'g s' is the type of the generator, 's' is the state type, and 'm' the underlying monad. Via the functional dependency 'g m -> s', the state type is determined by the generator and monad. 'Frozen' is the type of the generator's state "at rest". It is defined as an injective type family via 'f -> g', so there is no ambiguity as to which 'g' any 'Frozen g' belongs to. This definition is generic enough to accommodate, for example, the 'Gen' type from the 'mwc-random' package, which itself abstracts over the underlying primitive monad and state token. The documentation shows the full 'MonadRandom Gen' instance. Four 'MonadRandom' instances ("monadic adapters") are provided for pure generators to enable their use in monadic code. The documentation describes them in detail. 'Uniform' and 'UniformRange' ---------------------------- The 'Random' typeclass has conceptually been split into 'Uniform' and 'UniformRange'. The 'Random' typeclass is still included for backwards compatibility. 'Uniform' is for types where it is possible to sample from the type's entire domain; 'UniformRange' is for types where one can sample from a specified range. Breaking Changes ================ This patch introduces these breaking changes: * requires 'base >= 4.10' (GHC-8.2) * 'StdGen' is no longer an instance of 'Read' * 'randomIO' and 'randomRIO' where extracted from the 'Random' class into separate functions In addition, there may be import clashes with new functions, e.g. 'uniform' and 'uniformR'. Deprecations ============ This patch introduces 'genWord64', 'genWord32' and similar methods to the 'RandomGen' class. The significantly slower method 'next' and its companion 'genRange' are now deprecated. Co-authored-by: Alexey Kuleshevich <[email protected]> Co-authored-by: idontgetoutmuch <[email protected]> Co-authored-by: Leonhard Markert <[email protected]>

This patch is mostly backwards compatible. See "Breaking Changes" below for the full list of backwards incompatible changes. This patch fixes quality and performance issues, addresses additional miscellaneous issues, and introduces a monadic API. Issues addressed ================ Priority issues fixed in this patch: - Title: "The seeds generated by split are not independent" Link: haskell#25 Fixed: changed algorithm to SplitMix, which provides a robust 'split' operation - Title: "Very low throughput" Link: haskell#51 Fixed: see "Performance" below Additional issues addressed in this patch: - Title: "Add Random instances for tuples" Link: haskell#26 Addressed: added 'Uniform' instances for up to 6-tuples - Title: "Add Random instance for Natural" Link: haskell#44 Addressed: added 'UniformRange' instance for 'Natural' - Title: "incorrect distribution of randomR for floating-point numbers" Link: haskell#53 Addressed: see "Regarding floating-point numbers" below - Title: "System/Random.hs:43:1: warning: [-Wtabs]" Link: haskell#55 Fixed: no more tabs - Title: "Why does random for Float and Double produce exactly 24 or 53 bits?" Link: haskell#58 Fixed: see "Regarding floating-point numbers" below - Title: "read :: StdGen fails for strings longer than 6" Link: haskell#59 Addressed: 'StdGen' is no longer an instance of 'Read' Regarding floating-point numbers: with this patch, the relevant instances for 'Float' and 'Double' sample more bits than before but do not sample every possible representable value. The documentation now clearly spells out what this means for users. Quality (issue 25) ================== The algorithm [1] in version 1.1 of this library fails empirical PRNG tests when used to generate "split sequences" as proposed in [3]. SplitMix [2] passes the same tests. This patch changes 'StdGen' to use the SplitMix implementation provided by the splitmix package. Test batteries used: dieharder, TestU1, PractRand. [1]: P. L'Ecuyer, "Efficient and portable combined random number generators". https://doi.org/10.1145/62959.62969 [2]: G. L. Steele, D. Lea, C. H. Flood, "Fast splittable pseudorandom number generators". https://doi.org/10.1145/2714064.2660195 [3]: H. G. Schaathun, "Evaluation of splittable pseudo-random generators". https://doi.org/10.1017/S095679681500012X Performance (issue 51) ====================== The "improvement" column in the following table is a multiplier: the improvement for 'random' for type 'Float' is 1038, so this operation is 1038 times faster with this patch. | Name | Mean (1.1) | Mean (patch) | Improvement| | ----------------------- | ---------- | ------------ | ---------- | | pure/random/Float | 30 | 0.03 | 1038| | pure/random/Double | 52 | 0.03 | 1672| | pure/random/Integer | 43 | 0.33 | 131| | pure/uniform/Word8 | 14 | 0.03 | 422| | pure/uniform/Word16 | 13 | 0.03 | 375| | pure/uniform/Word32 | 21 | 0.03 | 594| | pure/uniform/Word64 | 42 | 0.03 | 1283| | pure/uniform/Word | 44 | 0.03 | 1491| | pure/uniform/Int8 | 15 | 0.03 | 511| | pure/uniform/Int16 | 15 | 0.03 | 507| | pure/uniform/Int32 | 22 | 0.03 | 749| | pure/uniform/Int64 | 44 | 0.03 | 1405| | pure/uniform/Int | 43 | 0.03 | 1512| | pure/uniform/Char | 17 | 0.49 | 35| | pure/uniform/Bool | 18 | 0.03 | 618| | pure/uniform/CChar | 14 | 0.03 | 485| | pure/uniform/CSChar | 14 | 0.03 | 455| | pure/uniform/CUChar | 13 | 0.03 | 448| | pure/uniform/CShort | 14 | 0.03 | 473| | pure/uniform/CUShort | 13 | 0.03 | 457| | pure/uniform/CInt | 21 | 0.03 | 737| | pure/uniform/CUInt | 21 | 0.03 | 742| | pure/uniform/CLong | 43 | 0.03 | 1544| | pure/uniform/CULong | 42 | 0.03 | 1460| | pure/uniform/CPtrdiff | 43 | 0.03 | 1494| | pure/uniform/CSize | 43 | 0.03 | 1475| | pure/uniform/CWchar | 22 | 0.03 | 785| | pure/uniform/CSigAtomic | 21 | 0.03 | 749| | pure/uniform/CLLong | 43 | 0.03 | 1554| | pure/uniform/CULLong | 42 | 0.03 | 1505| | pure/uniform/CIntPtr | 43 | 0.03 | 1476| | pure/uniform/CUIntPtr | 42 | 0.03 | 1463| | pure/uniform/CIntMax | 43 | 0.03 | 1535| | pure/uniform/CUIntMax | 42 | 0.03 | 1493| API changes =========== MonadRandom ----------- This patch adds a class 'MonadRandom': -- | 'MonadRandom' is an interface to monadic pseudo-random number generators. class Monad m => MonadRandom g s m | g m -> s where {-# MINIMAL freezeGen,thawGen,(uniformWord32|uniformWord64) #-} type Frozen g = (f :: Type) | f -> g freezeGen :: g s -> m (Frozen g) thawGen :: Frozen g -> m (g s) uniformWord32 :: g s -> m Word32 -- default implementation in terms of uniformWord64 uniformWord64 :: g s -> m Word64 -- default implementation in terms of uniformWord32 -- plus methods for other word sizes and for byte strings -- all have default implementations so the MINIMAL pragma holds Conceptually, in 'MonadRandom g s m', 'g s' is the type of the generator, 's' is the state type, and 'm' the underlying monad. Via the functional dependency 'g m -> s', the state type is determined by the generator and monad. 'Frozen' is the type of the generator's state "at rest". It is defined as an injective type family via 'f -> g', so there is no ambiguity as to which 'g' any 'Frozen g' belongs to. This definition is generic enough to accommodate, for example, the 'Gen' type from the 'mwc-random' package, which itself abstracts over the underlying primitive monad and state token. The documentation shows the full 'MonadRandom Gen' instance. Four 'MonadRandom' instances ("monadic adapters") are provided for pure generators to enable their use in monadic code. The documentation describes them in detail. 'Uniform' and 'UniformRange' ---------------------------- The 'Random' typeclass has conceptually been split into 'Uniform' and 'UniformRange'. The 'Random' typeclass is still included for backwards compatibility. 'Uniform' is for types where it is possible to sample from the type's entire domain; 'UniformRange' is for types where one can sample from a specified range. Breaking Changes ================ This patch introduces these breaking changes: * requires 'base >= 4.10' (GHC-8.2) * 'StdGen' is no longer an instance of 'Read' * 'randomIO' and 'randomRIO' where extracted from the 'Random' class into separate functions In addition, there may be import clashes with new functions, e.g. 'uniform' and 'uniformR'. Deprecations ============ This patch introduces 'genWord64', 'genWord32' and similar methods to the 'RandomGen' class. The significantly slower method 'next' and its companion 'genRange' are now deprecated. Co-authored-by: Alexey Kuleshevich <[email protected]> Co-authored-by: idontgetoutmuch <[email protected]> Co-authored-by: Leonhard Markert <[email protected]>

This patch is mostly backwards compatible. See "Breaking Changes" below for the full list of backwards incompatible changes. This patch fixes quality and performance issues, addresses additional miscellaneous issues, and introduces a monadic API. Issues addressed ================ Priority issues fixed in this patch: - Title: "The seeds generated by split are not independent" Link: haskell#25 Fixed: changed algorithm to SplitMix, which provides a robust 'split' operation - Title: "Very low throughput" Link: haskell#51 Fixed: see "Performance" below Additional issues addressed in this patch: - Title: "Add Random instances for tuples" Link: haskell#26 Addressed: added 'Uniform' instances for up to 6-tuples - Title: "Add Random instance for Natural" Link: haskell#44 Addressed: added 'UniformRange' instance for 'Natural' - Title: "incorrect distribution of randomR for floating-point numbers" Link: haskell#53 Addressed: see "Regarding floating-point numbers" below - Title: "System/Random.hs:43:1: warning: [-Wtabs]" Link: haskell#55 Fixed: no more tabs - Title: "Why does random for Float and Double produce exactly 24 or 53 bits?" Link: haskell#58 Fixed: see "Regarding floating-point numbers" below - Title: "read :: StdGen fails for strings longer than 6" Link: haskell#59 Addressed: 'StdGen' is no longer an instance of 'Read' Regarding floating-point numbers: with this patch, the relevant instances for 'Float' and 'Double' sample more bits than before but do not sample every possible representable value. The documentation now clearly spells out what this means for users. Quality (issue 25) ================== The algorithm [1] in version 1.1 of this library fails empirical PRNG tests when used to generate "split sequences" as proposed in [3]. SplitMix [2] passes the same tests. This patch changes 'StdGen' to use the SplitMix implementation provided by the splitmix package. Test batteries used: dieharder, TestU1, PractRand. [1]: P. L'Ecuyer, "Efficient and portable combined random number generators". https://doi.org/10.1145/62959.62969 [2]: G. L. Steele, D. Lea, C. H. Flood, "Fast splittable pseudorandom number generators". https://doi.org/10.1145/2714064.2660195 [3]: H. G. Schaathun, "Evaluation of splittable pseudo-random generators". https://doi.org/10.1017/S095679681500012X Performance (issue 51) ====================== The "improvement" column in the following table is a multiplier: the improvement for 'random' for type 'Float' is 1038, so this operation is 1038 times faster with this patch. | Name | Mean (1.1) | Mean (patch) | Improvement| | ----------------------- | ---------- | ------------ | ---------- | | pure/random/Float | 30 | 0.03 | 1038| | pure/random/Double | 52 | 0.03 | 1672| | pure/random/Integer | 43 | 0.33 | 131| | pure/uniform/Word8 | 14 | 0.03 | 422| | pure/uniform/Word16 | 13 | 0.03 | 375| | pure/uniform/Word32 | 21 | 0.03 | 594| | pure/uniform/Word64 | 42 | 0.03 | 1283| | pure/uniform/Word | 44 | 0.03 | 1491| | pure/uniform/Int8 | 15 | 0.03 | 511| | pure/uniform/Int16 | 15 | 0.03 | 507| | pure/uniform/Int32 | 22 | 0.03 | 749| | pure/uniform/Int64 | 44 | 0.03 | 1405| | pure/uniform/Int | 43 | 0.03 | 1512| | pure/uniform/Char | 17 | 0.49 | 35| | pure/uniform/Bool | 18 | 0.03 | 618| | pure/uniform/CChar | 14 | 0.03 | 485| | pure/uniform/CSChar | 14 | 0.03 | 455| | pure/uniform/CUChar | 13 | 0.03 | 448| | pure/uniform/CShort | 14 | 0.03 | 473| | pure/uniform/CUShort | 13 | 0.03 | 457| | pure/uniform/CInt | 21 | 0.03 | 737| | pure/uniform/CUInt | 21 | 0.03 | 742| | pure/uniform/CLong | 43 | 0.03 | 1544| | pure/uniform/CULong | 42 | 0.03 | 1460| | pure/uniform/CPtrdiff | 43 | 0.03 | 1494| | pure/uniform/CSize | 43 | 0.03 | 1475| | pure/uniform/CWchar | 22 | 0.03 | 785| | pure/uniform/CSigAtomic | 21 | 0.03 | 749| | pure/uniform/CLLong | 43 | 0.03 | 1554| | pure/uniform/CULLong | 42 | 0.03 | 1505| | pure/uniform/CIntPtr | 43 | 0.03 | 1476| | pure/uniform/CUIntPtr | 42 | 0.03 | 1463| | pure/uniform/CIntMax | 43 | 0.03 | 1535| | pure/uniform/CUIntMax | 42 | 0.03 | 1493| API changes =========== StatefulGen ----------- This patch adds a class 'StatefulGen': -- | 'StatefulGen' is an interface to monadic pseudo-random number generators. class Monad m => StatefulGen g m where uniformWord32 :: g -> m Word32 -- default implementation in terms of uniformWord64 uniformWord64 :: g -> m Word64 -- default implementation in terms of uniformWord32 -- plus methods for other word sizes and for byte strings -- all have default implementations so the MINIMAL pragma holds In 'StatefulGen g m', 'g' is the type of the generator and 'm' the underlying monad. Four 'StatefulGen' instances ("monadic adapters") are provided for pure generators to enable their use in monadic code. The documentation describes them in detail. FrozenGen --------- This patch also introduces a class 'FrozenGen': -- | 'FrozenGen' is designed for stateful pseudo-random number generators -- that can be saved as and restored from an immutable data type. class StatefulGen (MutableGen f m) m => FrozenGen f m where type MutableGen f m = (g :: Type) | g -> f freezeGen :: MutableGen f m -> m f thawGen :: f -> m (MutableGen f m) 'f' is the type of the generator's state "at rest" and 'm' the underlying monad. 'MutableGen' is defined as an injective type family via 'g -> f' so for any generator 'g', the type 'f' of its at-rest state is well-defined. Both 'StatefulGen' and 'FrozenGen' are generic enough to accommodate, for example, the 'Gen' type from the 'mwc-random' package, which itself abstracts over the underlying primitive monad and state token. The documentation shows the full instances. 'Uniform' and 'UniformRange' ---------------------------- The 'Random' typeclass has conceptually been split into 'Uniform' and 'UniformRange'. The 'Random' typeclass is still included for backwards compatibility. 'Uniform' is for types where it is possible to sample from the type's entire domain; 'UniformRange' is for types where one can sample from a specified range. Breaking Changes ================ This patch introduces these breaking changes: * requires 'base >= 4.10' (GHC-8.2) * 'StdGen' is no longer an instance of 'Read' * 'randomIO' and 'randomRIO' where extracted from the 'Random' class into separate functions In addition, there may be import clashes with new functions, e.g. 'uniform' and 'uniformR'. Deprecations ============ This patch introduces 'genWord64', 'genWord32' and similar methods to the 'RandomGen' class. The significantly slower method 'next' and its companion 'genRange' are now deprecated. Co-authored-by: Alexey Kuleshevich <[email protected]> Co-authored-by: idontgetoutmuch <[email protected]> Co-authored-by: Leonhard Markert <[email protected]>

This patch is mostly backwards compatible. See "Breaking Changes" below for the full list of backwards incompatible changes. This patch fixes quality and performance issues, addresses additional miscellaneous issues, and introduces a monadic API. Issues addressed ================ Priority issues fixed in this patch: - Title: "The seeds generated by split are not independent" Link: haskell#25 Fixed: changed algorithm to SplitMix, which provides a robust 'split' operation - Title: "Very low throughput" Link: haskell#51 Fixed: see "Performance" below Additional issues addressed in this patch: - Title: "Add Random instances for tuples" Link: haskell#26 Addressed: added 'Uniform' instances for up to 6-tuples - Title: "Add Random instance for Natural" Link: haskell#44 Addressed: added 'UniformRange' instance for 'Natural' - Title: "incorrect distribution of randomR for floating-point numbers" Link: haskell#53 Addressed: see "Regarding floating-point numbers" below - Title: "System/Random.hs:43:1: warning: [-Wtabs]" Link: haskell#55 Fixed: no more tabs - Title: "Why does random for Float and Double produce exactly 24 or 53 bits?" Link: haskell#58 Fixed: see "Regarding floating-point numbers" below - Title: "read :: StdGen fails for strings longer than 6" Link: haskell#59 Addressed: 'StdGen' is no longer an instance of 'Read' Regarding floating-point numbers: with this patch, the relevant instances for 'Float' and 'Double' sample more bits than before but do not sample every possible representable value. The documentation now clearly spells out what this means for users. Quality (issue 25) ================== The algorithm [1] in version 1.1 of this library fails empirical PRNG tests when used to generate "split sequences" as proposed in [3]. SplitMix [2] passes the same tests. This patch changes 'StdGen' to use the SplitMix implementation provided by the splitmix package. Test batteries used: dieharder, TestU1, PractRand. [1]: P. L'Ecuyer, "Efficient and portable combined random number generators". https://doi.org/10.1145/62959.62969 [2]: G. L. Steele, D. Lea, C. H. Flood, "Fast splittable pseudorandom number generators". https://doi.org/10.1145/2714064.2660195 [3]: H. G. Schaathun, "Evaluation of splittable pseudo-random generators". https://doi.org/10.1017/S095679681500012X Performance (issue 51) ====================== The "improvement" column in the following table is a multiplier: the improvement for 'random' for type 'Float' is 1038, so this operation is 1038 times faster with this patch. | Name | Mean (1.1) | Mean (patch) | Improvement| | ----------------------- | ---------- | ------------ | ---------- | | pure/random/Float | 30 | 0.03 | 1038| | pure/random/Double | 52 | 0.03 | 1672| | pure/random/Integer | 43 | 0.33 | 131| | pure/uniform/Word | 44 | 0.03 | 1491| | pure/uniform/Int | 43 | 0.03 | 1512| | pure/uniform/Char | 17 | 0.49 | 35| | pure/uniform/Bool | 18 | 0.03 | 618| API changes =========== StatefulGen ----------- This patch adds a class 'StatefulGen': -- | 'StatefulGen' is an interface to monadic pseudo-random number generators. class Monad m => StatefulGen g m where uniformWord32 :: g -> m Word32 -- default implementation in terms of uniformWord64 uniformWord64 :: g -> m Word64 -- default implementation in terms of uniformWord32 -- plus methods for other word sizes and for byte strings -- all have default implementations so the MINIMAL pragma holds In 'StatefulGen g m', 'g' is the type of the generator and 'm' the underlying monad. Four 'StatefulGen' instances ("monadic adapters") are provided for pure generators to enable their use in monadic code. The documentation describes them in detail. FrozenGen --------- This patch also introduces a class 'FrozenGen': -- | 'FrozenGen' is designed for stateful pseudo-random number generators -- that can be saved as and restored from an immutable data type. class StatefulGen (MutableGen f m) m => FrozenGen f m where type MutableGen f m = (g :: Type) | g -> f freezeGen :: MutableGen f m -> m f thawGen :: f -> m (MutableGen f m) 'f' is the type of the generator's state "at rest" and 'm' the underlying monad. 'MutableGen' is defined as an injective type family via 'g -> f' so for any generator 'g', the type 'f' of its at-rest state is well-defined. Both 'StatefulGen' and 'FrozenGen' are generic enough to accommodate, for example, the 'Gen' type from the 'mwc-random' package, which itself abstracts over the underlying primitive monad and state token. The documentation shows the full instances. 'Uniform' and 'UniformRange' ---------------------------- The 'Random' typeclass has conceptually been split into 'Uniform' and 'UniformRange'. The 'Random' typeclass is still included for backwards compatibility. 'Uniform' is for types where it is possible to sample from the type's entire domain; 'UniformRange' is for types where one can sample from a specified range. Breaking Changes ================ This patch introduces these breaking changes: * requires 'base >= 4.8' (GHC-7.10) * 'StdGen' is no longer an instance of 'Read' * 'randomIO' and 'randomRIO' where extracted from the 'Random' class into separate functions In addition, there may be import clashes with new functions, e.g. 'uniform' and 'uniformR'. Deprecations ============ This patch introduces 'genWord64', 'genWord32' and similar methods to the 'RandomGen' class. The significantly slower method 'next' and its companion 'genRange' are now deprecated. Co-authored-by: Alexey Kuleshevich <[email protected]> Co-authored-by: idontgetoutmuch <[email protected]> Co-authored-by: Leonhard Markert <[email protected]>

curiousleo · 2020-06-24T14:25:13Z

In version 1.2.0 (released yesterday), we decided not to attempt to give stronger guarantees for floating point numbers. This decision was taken after a long discussion, see the summary here: idontgetoutmuch#113 (comment).

What is new in 1.2.0 is that these issues are now documented: https://hackage.haskell.org/package/random-1.2.0/docs/System-Random-Stateful.html#g:14 -- I also used sbv to find these counterexamples btw :)

The decision not to tackle floating point guarantees was a pragmatic one: a lot of improvements had been made, and we didn't want to delay the release. Generating uniformly random floating point numbers is surprisingly tricky, and we may pick this up again in the future. In the meantime, I am exploring this topic in a little separate experiment.

This was referenced Apr 29, 2020

Check issue remains fixed idontgetoutmuch/random#117

Merged

Interface to performance idontgetoutmuch/random#1

Open

curiousleo pushed a commit to curiousleo/random that referenced this issue May 5, 2020

Merge pull request haskell#53 from idontgetoutmuch/uniform-instead-of…

937d958

…-randomM Use uniform instead of randomM in example

cartazio pushed a commit that referenced this issue May 19, 2020

Merge pull request #53 from idontgetoutmuch/uniform-instead-of-randomM

85a0a4d

Use uniform instead of randomM in example

idontgetoutmuch mentioned this issue May 20, 2020

V1.2 proposal #61

Merged

lehins mentioned this issue Jun 4, 2020

V1.2 proposal #62

Merged

Bodigrim mentioned this issue Jun 24, 2020

Why does random for Float and Double produce exactly 24 or 53 bits? #58

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

incorrect distribution of randomR for floating-point numbers #53

incorrect distribution of randomR for floating-point numbers #53

shaobo-he commented Apr 1, 2019

cartazio commented Apr 3, 2019 via email

shaobo-he commented Apr 4, 2019

curiousleo commented Jun 24, 2020

incorrect distribution of randomR for floating-point numbers #53

incorrect distribution of randomR for floating-point numbers #53

Comments

shaobo-he commented Apr 1, 2019

cartazio commented Apr 3, 2019 via email

shaobo-he commented Apr 4, 2019

curiousleo commented Jun 24, 2020