atgjack / prob Goto Github PK

Returning math.NaN() is preferable to throwing an error. Also, when a validation occurs, a NaN should be used instead of 0.0.

Currently there are none. Need to find a valid resource for generating values against.

I added the Negative Binomial distribution in #32 . Needs a better sampling test.

Similar to #1. Add some discrete distributions.

Discrete Distributions:

• Binomial
• Negative Binomial
• Poisson
• Geometric
• Hypergeometric
• Multinomial

Will probably update this list as I come up with more.
I'd like at least the following added:

Continuous

• Cauchy
• Chi-Square
• Exponential
• Gamma
• Normal
• Pareto
• Student's T
• Uniform
• Weibull
• Beta
• Logistic
• Log-Normal
• F
• Laplace

I added the Chi-Squared distribution in #12. The sampling tests are lame. I need to find a better way to verify them.

I added the Student's t-distribution in #14. The sampling tests are lame. I need to find a better way to verify them.

I added the Cauchy distribution in #4 but it's sampling tests are not really valid.
Looking for a better way to test this.

I've been using math.Pow for simple exponents, and that's probably got a small performance penalty.

Here is a list of where it's called.
grep -rno math.Pow

The chi square CDF does not account for underflow or overflow and therefore can emit (very) wrong answers.

package demo

import (
	"math"
	"testing"

	"github.com/atgjack/prob"
)

func TestProb(t *testing.T) {
	N := 200.0

	v := prob.ChiSquared{1}
	t.Log(1 - v.Cdf(N))
}

This will yield ~ 0.4933. The correct answer is approximately 0. Either extremes chi square values should be handled, or an error (or a panic, as is done by the probab/dst functions) should be returned. (I'm using github.com/tokenme/probab/dst instead so I don't plan on submitting a pull request -- my apologies -- but I do want to at least mention the issue.)

Update function names and change throughout source.

• Choose should be BinomialCoefficient
• A better name for Lowerincgamma

It looks like each distribution's Sample method is returning []float64{}, nil when it should be []float64, err. Need to go back through and fix this.

I added the Log-Normal distribution in #29. The sample average is fine, but the variance is off by quite a bit. This probably has to do with the natural log, but I need to fix this to test more accurately.

I added the Pareto distribution in #8. The sampling test seems alright, but I found this when looking for better tests.

Basically, it suggests estimating the parameters and comparing them to the base distribution. Seem's pretty simple to implement.