Features in rust are not supposed to change the signatures of the API, because it means that turning on the feature is a breaking change. This is the problem because cargo's choice of features as additive across the dependency tree.

For example suppose I have library A that doesn't use f64, and another library B that does. If I add both as a dependency, the KDE crate will be compiled with the f64 enabled. This will cause library A to not compile, because it was using f32.

From the cargo book:

features should be additive. That is, enabling a feature should not disable functionality, and it should usually be safe to enable any combination of features. A feature should not introduce a SemVer-incompatible change

Currently only univariate distributions are supported. A complete implementation would include seamless support for multivariate distributions. The only type that should be changed is the KernelDensityEstimator struct. Currently, the data structure is as follows:

pub struct KernelDensityEstimator<B, K> {
    observations: Vec<Float>,
    bandwidth: B,
    kernel: K,
}

The observations field will need to be converted to a nalgebra::DMatrix to support a multivariate distribution. The type should be hidden behind an alias so that end-users do not need to add nalgebra as a dependency in their own projects.

pub type Matrix2D = nalgebra::DMatrix<Float>;

pub struct KernelDensityEstimator<B, K> {
    observations: Matrix2D,
    bandwidth: B,
    kernel: K,
}

To prevent needless conversions for users working with univariate data, the data structure could instead add a generic parameter T representing the type of observations. For univariate data T could be concretely represented as Vec<Float> and for multivariate data T could be concretely represented as Matrix2D. However, this would require the introduction of two new traits UnivariateKDE and MultivariateKDE to mimic overloading of the method names pdf, cdf, and sample.

pub struct KernelDensityEstimator<T, B, K> {
    observations: T,
    bandwidth: B,
    kernel: K,
}

pub trait UnivariateKDE {
    // Unimplemented.
}

pub trait MultivariateKDE {
    // Unimplemented.
}

// Univariate case.
impl<B, K> UnivariateKDE for KernelDensityEstimator<Vec<Float>, B, K>
where
    B: Bandwidth,
    K: Kernel,
{
    // Unimplemented.
}

// Multivariate case.
impl<B, K> MultivariateKDE for KernelDensityEstimator<Matrix2D, B, K>
where
    B: Bandwidth,
    K: Kernel,
{
    // Unimplemented.
}

Lastly, the traits UnivariateKDE and MultivariateKDE should be sealed to prevent end-user implementations.

All the kernel functions described here should be implemented.

• [x] Uniform
• [ ] Triangular
• [x] Epanechnikov
• [ ] Quartic
• [ ] Triweight
• [ ] Tricube
• [x] Normal (Gaussian)
• [ ] Cosine
• [ ] Logistic
• [ ] Sigmoid
• [ ] Silverman