use ndarray::{Array1, Array2};
use solow_core::error::{Error, Result};
#[derive(Debug, Clone)]
pub struct InnovationsResult {
pub theta: Array2<f64>,
pub sigma2: Array1<f64>,
}
pub fn innovations_algo(acovf: &Array1<f64>, nobs: usize) -> Result<InnovationsResult> {
if nobs == 0 {
return Err(Error::Value("nobs must be positive".into()));
}
if acovf.len() < nobs {
return Err(Error::Value(
"acovf must contain at least nobs autocovariances".into(),
));
}
let cols = nobs.saturating_sub(1);
let mut theta = Array2::<f64>::zeros((nobs, cols));
let mut v = Array1::<f64>::zeros(nobs);
v[0] = acovf[0];
for n in 1..nobs {
for k in 0..n {
let mut sub = 0.0;
for j in 0..k {
sub += theta[[k, k - j - 1]] * theta[[n, n - j - 1]] * v[j];
}
if v[k] == 0.0 {
return Err(Error::Value(
"innovations algorithm encountered a zero prediction variance".into(),
));
}
theta[[n, n - k - 1]] = (acovf[n - k] - sub) / v[k];
}
let mut s = 0.0;
for j in 0..n {
let t = theta[[n, n - j - 1]];
s += t * t * v[j];
}
v[n] = acovf[0] - s;
}
Ok(InnovationsResult { theta, sigma2: v })
}
#[cfg(test)]
mod tests {
use super::*;
use ndarray::Array1;
#[test]
fn ma1_recovers_coefficient_and_variance() {
let acov = Array1::from_vec(vec![1.25, 0.5, 0.0, 0.0, 0.0, 0.0]);
let r = innovations_algo(&acov, 6).unwrap();
assert!((r.theta[[5, 0]] - 0.5).abs() < 1e-3);
assert!(r.theta[[5, 0]] < 0.5);
assert!((r.sigma2[5] - 1.0).abs() < 1e-3);
assert_eq!(r.sigma2[0], 1.25);
for n in 1..6 {
for j in n..5 {
assert_eq!(r.theta[[n, j]], 0.0);
}
}
}
#[test]
fn white_noise_has_zero_coefficients() {
let acov = Array1::from_vec(vec![2.0, 0.0, 0.0, 0.0]);
let r = innovations_algo(&acov, 4).unwrap();
for v in r.theta.iter() {
assert_eq!(*v, 0.0);
}
for v in r.sigma2.iter() {
assert_eq!(*v, 2.0);
}
}
#[test]
fn rejects_short_acovf() {
let acov = Array1::from_vec(vec![1.0, 0.5]);
assert!(innovations_algo(&acov, 5).is_err());
}
}