Metadata-Version: 2.1
Name: zarpy
Version: 1.0.1
Summary: Formally verified biased coin and n-sided die
Author: Alexander Bagnall
Author-email: Alexander Bagnall <abagnalla@gmail.com>
License: MIT License
        
        Copyright (c) [2022] [Alexander A. Bagnall]
        
        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.
Project-URL: Homepage, https://github.com/bagnalla/zar
Keywords: zar,zarpy,random,sampler,biased,coin,bernoulli,uniform,die,verified,coq
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python
Classifier: Programming Language :: Python :: 3
Requires-Python: >=3.6.3
Description-Content-Type: text/markdown
License-File: LICENSE

# Zarpy: formally verified biased coin and n-sided die.

See the [paper](https://arxiv.org/abs/2211.06747) (to appear in
PLDI'23) and [Github repository](https://github.com/bagnalla/zar).

## Why use Zarpy?

### 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:
```python
if random() < p:
  execute a1
else:
  execute a2
```

where `p` is a float in the range `[0,1]` and `random()` 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 guarantees of our overall system that
depend on the correctness of its probabilistic choices.

Zarpy provides an alternative that is guaranteed (by formal proof in
Coq) to execute `a1` with probability `p` (where `n` and `d` are
integers such that `p = n/d`):
```python
from zarpy import build_coin, flip
build_coin((n, d)) # Build and cache coin with bias p = n/d
if flip(): # Generate a Boolean value with Pr(True) = p 
  execute a1
else:
  execute a2
```

### 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`:
```python
x = rand() % n # Assign x random value from [0,n)
```
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](https://research.kudelskisecurity.com/2020/07/28/the-definitive-guide-to-modulo-bias-and-how-to-avoid-it/)
for more information on modulo bias). Zarpy provides a uniform sampler
that is guaranteed for any integer `0 < n` to generate samples from
`[0,n)` with probability `1/n` each:
```python
from zarpy import build_die, roll
build_die(n)
x = roll()
```

Although the Python function `random.randint` is ostensibly free from
modulo bias, our implementation guarantees so by a *formal proof of
correctness* in Coq.

## Trusted Computing Base

The samplers provided by Zarpy have been implemented and verified in
Coq and extracted to OCaml and bundled into Python package via
[pythonlib](https://github.com/janestreet/pythonlib). Validity of the
correctness proofs is thus dependent on the correctness of Coq's
extraction mechanism, the OCaml compiler and runtime, a small amount
of OCaml shim code (viewable
[here](https://github.com/bagnalla/zar/blob/main/python/zar/ocaml/zarpy.ml)),
and the pythonlib library.

## Proofs of correctness

The coin and die samplers are implemented as probabilistic programs in
the [Zar](https://github.com/bagnalla/zar) system and compiled to
[interaction trees](https://github.com/DeepSpec/InteractionTrees)
implementing them via reduction to sequences of fair coin flips. See
Section 3 of the [paper](https://arxiv.org/abs/2211.06747) for details
and the file
[zarpy.v](https://github.com/bagnalla/zar/blob/main/zarpy.v) for their
implementations and proofs of correctness.

Correctness is two-fold. For biased coin with bias `p`, we prove:

*
  [coin_itree_correct](https://github.com/bagnalla/zar/blob/main/zarpy.v#L57):
  the probability of producing `true` according to the formal probabilistic
  semantics of the constructed interaction tree is equal to `p`, and

*
  [coin_samples_equidistributed](https://github.com/bagnalla/zar/blob/main/zarpy.v#L75):
  when the source of random bits is uniformly distributed, for any
  sequence of coin flips the proportion of `true` samples converges to
  `p` 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`](https://v2.ocaml.org/api/Random.html) function. See
[the paper](https://arxiv.org/abs/2211.06747) for a more detailed
explanation.

Similarly, the theorem
[die_itree_correct](https://github.com/bagnalla/zar/blob/main/zarpy.v#L136)
proves semantic correctness of the n-sided die, and
[die_samples_equidistributed](https://github.com/bagnalla/zar/blob/main/zarpy.v#L161)
equidistribution of its samples.

## Usage

`seed()` initializes the PRNG via
[Random.self_init](https://v2.ocaml.org/api/Random.html).

### Biased coin

`build_coin((num, denom))` builds and caches a coin with `Pr(True) =
num/denom` for nonnegative integer `num` and positive integer `denom`.

`flip()` produces a single Boolean sample by flipping the cached coin.

`flip_n(n)` produces `n` Boolean samples by flipping the cached coin.

### N-sided die

`build_die(n)` builds and caches an n-sided die with `Pr(m) = 1/n` for
integer `m` where `0 <= m < n`.

`roll()` produces a single sample by rolling the cached die.

`roll_n(n)` produces `n` integer samples by rolling the cached die.
