pub fn kullback_leibler_divergence_impl(p: &[f64], q: &[f64]) -> Result<f64, String> {
if p.len() != q.len() {
return Err(format!(
"true and pred must have the same length, got {} and {}",
p.len(),
q.len()
));
}
let p_total: f64 = p.iter().sum();
let q_total: f64 = q.iter().sum();
let mut divergence = 0.0;
for (&p_value, &q_value) in p.iter().zip(q) {
let p_normalized = p_value / p_total;
if p_normalized != 0.0 {
let q_normalized = q_value / q_total;
divergence += p_normalized * (p_normalized / q_normalized).ln();
}
}
Ok(divergence)
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn normalizes_raw_counts() {
let probabilities = kullback_leibler_divergence_impl(&[0.4, 0.6], &[0.5, 0.5]).unwrap();
let counts = kullback_leibler_divergence_impl(&[4.0, 6.0], &[5.0, 5.0]).unwrap();
assert!((probabilities - counts).abs() < 1e-15);
}
#[test]
fn zero_reference_mass_contributes_nothing() {
assert_eq!(
kullback_leibler_divergence_impl(&[0.0, 2.0], &[0.0, 2.0]).unwrap(),
0.0
);
}
#[test]
fn missing_predicted_support_is_infinite() {
assert!(
kullback_leibler_divergence_impl(&[1.0, 1.0], &[0.0, 1.0])
.unwrap()
.is_infinite()
);
}
#[test]
fn mismatched_lengths_are_rejected() {
assert!(kullback_leibler_divergence_impl(&[1.0], &[1.0, 2.0]).is_err());
}
}