Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

More general Exp #969

Closed
saona-raimundo opened this issue Apr 26, 2020 · 13 comments
Closed

More general Exp #969

saona-raimundo opened this issue Apr 26, 2020 · 13 comments

Comments

@saona-raimundo
Copy link
Contributor

Background

What is your motivation?
I am implementing stochastic processes, continuous markov chains in particular, and I need an Exponential clock. I wanted to use rand_distr::Exp as an internal field, but I have the following two problems:

  1. To include absorbent states, I need a zero rate exponential, i.e. that sample always INFINITY.
  2. It would be nice to have a method set_lambda to change it, instead of having to construct a new one.

What type of application is this?
Numerical simulation

Feature request

Generalize the construction of Exp to include the case of lambda = 0.0.
Also, implement a "setter" method to change the lambda in an already created Exp.

I know that allowing `lambda = 0.0` could imply a comparison before each sample (checking if lambda is finite or not) and I do not know of any solution for this. If there is no way around to maintain performance, then I would prefer this feature not to be implemented.
@vks
Copy link
Collaborator

vks commented Apr 26, 2020

I'm not sure I understand what the problem is with constructing a new one. Is this about performance or convenience?

If you need this edge case behavior, couldn't you implement your own distribution on top of Exp?

@dhardy
Copy link
Member

dhardy commented Apr 26, 2020

  1. I don't understand whether this infinite case is merely a possible random sample (in which case it should happen rarely enough it probably doesn't matter) or is explicitly selected. But I understand your point: from this point of view, λ = 0 is a valid parametrisation. Perhaps we should add a second constructor fn new_allowing_zero(lambda: N) -> Self?
  2. Using distr = Exp::new(...) should have the same effect (probably exactly the same code when optimised). This is one reason we always mark constructors with #[inline].

@saona-raimundo
Copy link
Contributor Author

Thank you for the quick answer!!

Sorry for not being clear. I will try again.

General parametrization

λ = 0 is a valid parametrisation.

Yes, from my point of view it is (Wikipedia does not think the same).

Implications

Perhaps we should add a second constructor fn new_allowing_zero(lambda: N) -> Self?

The problem is that the implementation changes. An exponential with λ = 0 is always (samples always) infinity. But, an exponential with λ > 0 samples always finite and has the same distribution as Exp(1) / λ.
Therefore, I thought that allowing λ = 0 in the construction and using the naive implementation (use of if for every sample) will be "less performant for the general use" since the variable Exp is usually used with λ > 0.
If that is the case and there is no way around it,

couldn't you implement your own distribution on top of Exp?

I am okay using my own wrapper, I just wanted to ask if you had better ideas.

Setter method

I'm not sure I understand what the problem is with constructing a new one.

There is no problem with constructing a new one, a method set_lambda would be purely for convenience. Feel free to say no, I just thought the code would look nicer.

@dhardy
Copy link
Member

dhardy commented Apr 26, 2020

(use of if for every sample)

I was thinking optimise for the common case, waste a sample and multiply by INF in the zero case. Is that not acceptable?

purely for convenience

When designing an API for general usage, it is always a good idea to minimise complexity: avoid methods you don't need. And I don't think a setter adds enough.

@saona-raimundo
Copy link
Contributor Author

waste a sample and multiply by INF in the zero case. Is that not acceptable?

Good!! I didn't think about it. In fact, there is no need to waste a sample! This is the current implementation.

#[derive(Clone, Copy, Debug)]
pub struct Exp<N> {
    /// `lambda` stored as `1/lambda`, since this is what we scale by.
    lambda_inverse: N
}

/// Error type returned from `Exp::new`.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Error {
    /// `lambda <= 0` or `nan`.
    LambdaTooSmall,
}

impl<N: Float> Exp<N>
where Exp1: Distribution<N>
{
    /// Construct a new `Exp` with the given shape parameter
    /// `lambda`.
    #[inline]
    pub fn new(lambda: N) -> Result<Exp<N>, Error> {
        if !(lambda > N::from(0.0)) {
            return Err(Error::LambdaTooSmall);
        }
        Ok(Exp { lambda_inverse: N::from(1.0) / lambda })
    }
}

impl<N: Float> Distribution<N> for Exp<N>
where Exp1: Distribution<N>
{
    fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> N {
        rng.sample(Exp1) * self.lambda_inverse
    }
}

One could change the constructor method (or create another constructor to not have breaking change) like this.

    pub fn new(lambda: N) -> Result<Exp<N>, Error> {
        if lambda > N::from(0.0) {
            return Ok(Exp { lambda_inverse: N::from(1.0) / lambda });
        } else if lambda < N::from(0.0) {
            return Err(Error::LambdaTooSmall);
        } else {
        Ok(Exp { lambda_inverse: N::infinity() })
        }
    }
}

Consequences

  1. No performance change for the common use case.
  2. Ask to implement infinity for the rand_distr::Float trait.
  3. Change documentation of LambdaTooSmall for strictly negative lambda.

@dhardy
Copy link
Member

dhardy commented Apr 28, 2020

@rasa200 that's roughly what I was thinking, yes. It does still sample Exp1 when calling sample(), but I think that's acceptable enough.

@vks any thoughts on whether we should allow zero-rate in the new constructor? The only issue is whether some users expect an error in this case. It may be okay just to relax the bound to lambda >= 0?

@vks
Copy link
Collaborator

vks commented Apr 28, 2020

I think either way is fine as long as it is documented.

@saona-raimundo
Copy link
Contributor Author

It does still sample Exp1 when calling sample(), but I think that's acceptable enough.

Sure! It is good!

@dhardy
Copy link
Member

dhardy commented Apr 29, 2020

Okay then, I guess the only code change we need is to Exp::new. Documentation-wise, we need a clear note on that constructor, tweaks to the error type, and a note in the changelog.

@rasa200 would you like to make a PR?

@saona-raimundo
Copy link
Contributor Author

@dhardy Happy to make a PR :D It will be my first ever.

One question: To use N::infinity(), I would need to add the infinity method to the rand_distr::Float trait. Is that okay?

Another possibility is using
N::from(f64::INFINITY) instead if N::infinity(), which:

  • Does not require to change the Float trait
  • Ask for possibly one unnecessary casting in the "non-common" case.

Historical note: In rand, version 0.7.3, rand::distributions::Exp can be created only with f64

@dhardy
Copy link
Member

dhardy commented Apr 29, 2020

I would need to add the infinity method to the rand_distr::Float trait. Is that okay?

You shouldn't need to. I just tested:

    /// Construct a new `Exp` with the given shape parameter
    /// `lambda`.
    #[inline]
    pub fn new(lambda: N) -> Result<Exp<N>, Error> {
        if !(lambda >= N::from(0.0)) {
            return Err(Error::LambdaTooSmall);
        }
        Ok(Exp {
            lambda_inverse: N::from(1.0) / lambda,
        })
    }
}

...

#[cfg(test)]
mod test {
    use super::*;

    #[test]
    fn test_zero() {
        let d = Exp::new(0.0).unwrap();
        assert_eq!(d.sample(&mut crate::test::rng(21)), std::f64::INFINITY);
    }

Yes, supporting both float types is a recent change.

@saona-raimundo
Copy link
Contributor Author

True!! My bad, I should have tried first.

@saona-raimundo
Copy link
Contributor Author

There is a PR :D

I had a bit of travel with the documentation (see PR), because of the possibility of a custom implementation of 1 / 0 for types implementing the rand_distr::Float trait. Other than that, I should have taken all the discussion into account.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants