package zar
Install
Dune Dependency
Authors
Maintainers
Sources
sha256=2a7a509909c8f6c477a17729098d7ae4d3a6c3dea2eb711079abe751d3cd95a0
sha512=501aae0bae51d4a48fbc96118641d1eb5ca414cb187cf007cf9f5308b9a34226ba571a3c6019aa008d1afb85735fe389263d54289d75416377745a6f09e4f2b8
Description
See the paper (to appear in PLDI'23) and Github repository.
README
Zar OCaml: formally verified sampling from discrete probability distributions.
See the related paper (to appear in PLDI'23), the main Zar Github repository, and similar Haskell package.
Why use Zar?
Probabilistic Choice
A basic operation in randomized algorithms is probabilistic choice: for some p ∈ [0,1]
, execute action a1
with probability p
or action a2
with probability 1-p
(i.e., flip a biased coin to determine the path of execution). A common method for performing probabilistic choice is as follows:
if Random.float 1.0 < p then a1 else a2
where p
is a float in the range [0,1]
and Random.float 1.0
produces a random float in the range [0,1)
. While good enough for many applications, this approach is not always correct due to float roundoff error. We can only expect a1
to be executed with probability p + ϵ
for some small error term ϵ
, which technically invalidates any correctness guarantee of our overall system that depends on the correctness of its probabilistic choices.
Zar provides an alternative that is formally proved (in Coq) to execute a1
with probability p
(where num
and denom
are integers such that p = num / denom
):
let coin_stream = Zar.coin num denom in (* Build coin stream *)
if Zar.first coin_stream (* Look at first flip of the coin *)
then a1 else a2
The expression Zar.coin num denom
builds a stream of bool
s resulting from flipping the coin with bias p = num / denom
. See lib/stream.mli for the stream type interface. We may integrate with an existing library in the future (e.g., streaming) similar to the Zar Haskell package's integration with the pipes library.
Internally, the coin is constructed as a stream transformer of type bool stream -> bool stream
that transforms an input source of fair coin flips into an output stream of biased coin flips. The coin transformer is applied to a default source of fair coin flips based on the OCaml Random module. The following code is equivalent:
let coin_transformer = Zar.coin_transformer num denom
let coin_stream = coin_transformer (Zar.bits ())
if Zar.first coin_stream (* Look at first flip of the coin *)
then a1 else a2
The user has the option to apply the coin transformer to their own source of fair coin flips instead (perhaps one connected to a "true" source of randomness such as the NIST randomness beacon).
Uniform Sampling
Another common operation is to randomly draw from a finite collection of values with equal (uniform) probability of each. An old trick for drawing an integer uniformly from the range [0, n)
is to generate a random integer from [0, RAND_MAX]
and take the modulus wrt. n
:
k = rand() // n # Assign k random value from [0,n)
// do something with k
but this method suffers from modulo bias when n
is not a power of 2, causing some values to occur with higher probability than others (see, e.g., this article for more information on modulo bias). Zar provides a uniform sampler that is guaranteed for any integer 0 < n
to generate samples from the range [0,n)
with probability 1/n
each:
let die = Zar.die n in
let k = Zar.first die in (* k drawn uniformly from [0,n) *)
(* do something with k *)
Although the OCaml function Random.int
is ostensibly free from modulo bias, our implementation guarantees it by a formal proof of correctness in Coq.
Finite Distributions
The coin and die samplers are special cases of a more general construction for finite probability distributions that we provide here. Given a list of nonnegative integer weights weights
such that 0 < weightsᵢ
for some i
(at least one of the weights is nonzero), we can draw an integer k
from the range [0, |weights|)
with probability weightsₖ / ∑ⱼweightsⱼ
(the corresponding weight of k
normalized by the sum of all weights):
let findist = Zar.findist weights in
let k = Findist.sample () in
(* do something with k *)
...
For example, Zar.findist [1; 3; 2]
builds a sampler that draws integers from the set {0, 1, 2}
with Pr(0) = 1/6
, Pr(1) = 1/2
, and Pr(2) = 1/3
.
Trusted Computing Base
The samplers provided by Zar have been implemented and verified in Coq and extracted to OCaml for execution. Validity of the correctness proofs is thus dependent on the correctness of Coq's extraction mechanism, the OCaml compiler and runtime, and a small amount of OCaml shim code (viewable here and thoroughly tested with QCheck here),
Proofs of Correctness
The samplers are implemented as choice-fix (CF) trees (an intermediate representation used in the Zar compiler) and compiled to interaction trees that implement them via reduction to sequences of fair coin flips. See Section 3 of the paper for details and the file ocamlzar.v for their implementations and proofs of correctness.
Correctness is two-fold. For biased coin with bias p
, we prove:
coin_itree_correct: the probability of producing
true
according to the formal probabilistic semantics of the constructed interaction tree is equal top
, andcoin_true_converges: when the source of random bits is uniformly distributed, for any sequence of coin flips the proportion of
true
samples converges top
as the number of samples goes to +∞.
The equidistribution result is dependent on uniform distribution of the Boolean values generated by OCaml's Random.bool
function. See the paper for a more detailed explanation.
Similarly, Theorem die_itree_correct proves semantic correctness of the n-sided die, and Corollary die_eq_n_converges that for any m < n
the proportion of samples equal to m
converges to 1 / n
.
Theorem findist_itree_correct proves semantic correctness of findist samplers, and Corollary findist_eq_n_converges that for any weight vector weights
and integer 0 <= i < |weights|
, the proportion of samples equal to i
converges to weightsᵢ / ∑ⱼweightsⱼ
.
Usage
See zar.mli for the top-level interface.
Zar.bits ()
produces a stream of uniformly distributed random bits.
Zar.seed ()
initializes the PRNG for Zar.bits
(currently just calls Random.self_init).
Biased Coin
Zar.coin_transformer num denom
builds a stream transformer that when applied to a stream of uniformly distributed random bits generates bool
samples with Pr(True) = num/denom
. Requires 0 <= num < denom
and 0 < denom
.
Zar.coin num denom
composes Zar.coin_transformer num denom
with the default source of uniformly distributed random bits.
N-sided Die
Zar.die_transformer n
builds a stream transformer that when applied to a stream of uniformly distributed random bits generates int
samples with Pr(m) = 1/n
for integer m where 0 <= m < n
.
Zar.die n
composes Zar.die_transformer n
with the default source of uniformly distributed random bits.
Finite Distribution
Zar.findist_transformer weights
builds a stream transformer from list of nonnegative integer weights weights
(where 0 < weightsᵢ
for some i
) that when applied to a stream of uniformly distributed random bits generates int
samples with Pr(i) = weightsᵢ / ∑ⱼweightsⱼ
for integer 0 <= i < |weights|
.
Zar.findist weights
composes Zar.findist_transformer weights
with the default source of uniformly distributed random bits.
Performance and Limitations
The samplers here are optimized for sampling performance at the expense of build time. Thus, this library not be ideal if your use case involves frequent rebuilding due to changes in the samplers' parameters (e.g., the coin bias or the number of sides of the die). For example, in our experiments it takes ~0.22s to build a 100000-sided die and 1.83s to build a 500000-sided die, but only ~1.85s and ~2.19s respectively to generate one million samples from them.
The size of the in-memory representation of a coin with bias p = num / denom
is proportional to denom
(after bringing the fraction to reduced form). The size of an n
-sided die is proportional to n
, and the size of a finite distribution to the sum of its weights. The formal results we provide are partial in the sense that they only apply to samplers that execute without running out of memory. I.e., we do not provide any guarantees against stack overflow or out-of-memory errors when, e.g., n
is too large.