refactor(gaussian): switch to natural-parameter storage (pi, tau)

Mul and Div become two f64 adds/subs with no sqrt in the hot path.
mu() and sigma() are computed on demand from stored pi/tau.

Key implementation notes:
- exclude() returns N00 when var <= 0 to avoid inf/inf = NaN when
  two Gaussians have the same precision (ULP-level round-trip error
  from the pi→sigma accessor).
- Mul<f64> by 0.0 returns N00 (point mass at 0), matching old behavior.
- from_ms(0, 0) == N00 {pi:inf, tau:0}; from_ms(0, inf) == N_INF {pi:0, tau:0}.

Golden values in test_1vs1vs1_draw updated: nat-param arithmetic
rounds mu to 25.0 (was 24.999999) and shifts sigma by ~3e-7.
Both differences are bounded and validated against the original Python
reference values.

Part of T0 engine redesign.
This commit is contained in:
2026-04-24 06:59:43 +02:00
parent 06d3c886fe
commit a667deb7e1
6 changed files with 174 additions and 170 deletions

View File

@@ -2,143 +2,159 @@ use std::ops;
use crate::{MU, N_INF, SIGMA};
/// A Gaussian distribution stored in natural parameters.
///
/// `pi = 1 / sigma^2` (precision)
/// `tau = mu * pi` (precision-adjusted mean)
///
/// Multiplication and division in message passing become pure adds/subs of
/// the stored fields with no `sqrt` or reciprocal in the hot path. `mu()` and
/// `sigma()` are accessors computed on demand.
#[derive(Clone, Copy, PartialEq, Debug)]
pub struct Gaussian {
pub mu: f64,
pub sigma: f64,
pi: f64,
tau: f64,
}
impl Gaussian {
/// Construct from mean and standard deviation.
pub const fn from_ms(mu: f64, sigma: f64) -> Self {
Gaussian { mu, sigma }
if sigma == f64::INFINITY {
Self { pi: 0.0, tau: 0.0 }
} else if sigma == 0.0 {
// Point mass at mu. tau = mu * pi = mu * inf.
// For mu == 0 this is 0; for mu != 0 it is inf * mu = inf (IEEE).
// Only N00 (mu=0, sigma=0) is used in practice.
Self {
pi: f64::INFINITY,
tau: if mu == 0.0 { 0.0 } else { f64::INFINITY },
}
} else {
let pi = 1.0 / (sigma * sigma);
Self { pi, tau: mu * pi }
}
}
/// Construct directly from natural parameters.
#[inline]
pub(crate) const fn from_natural(pi: f64, tau: f64) -> Self {
Self { pi, tau }
}
#[inline]
pub fn pi(&self) -> f64 {
if self.sigma > 0.0 {
self.sigma.powi(-2)
} else {
f64::INFINITY
}
self.pi
}
#[inline]
pub fn tau(&self) -> f64 {
if self.sigma > 0.0 {
self.mu * self.pi()
self.tau
}
#[inline]
pub fn mu(&self) -> f64 {
if self.pi == 0.0 {
0.0
} else {
self.tau / self.pi
}
}
#[inline]
pub fn sigma(&self) -> f64 {
if self.pi == 0.0 {
f64::INFINITY
} else if self.pi.is_infinite() {
0.0
} else {
1.0 / self.pi.sqrt()
}
}
pub(crate) fn delta(&self, m: Gaussian) -> (f64, f64) {
((self.mu - m.mu).abs(), (self.sigma - m.sigma).abs())
pub(crate) fn delta(&self, other: Gaussian) -> (f64, f64) {
(
(self.mu() - other.mu()).abs(),
(self.sigma() - other.sigma()).abs(),
)
}
pub(crate) fn exclude(&self, m: Gaussian) -> Self {
Self {
mu: self.mu - m.mu,
sigma: (self.sigma.powi(2) - m.sigma.powi(2)).sqrt(),
pub(crate) fn exclude(&self, other: Gaussian) -> Self {
let var = self.sigma().powi(2) - other.sigma().powi(2);
if var <= 0.0 {
// When sigma_self ≈ sigma_other (including ULP-level rounding differences
// from the pi→sigma accessor round-trip), the excluded contribution is N00.
// Computing from_ms(tiny_mu, 0.0) would give {pi:inf, tau:inf}, whose
// mu() = inf/inf = NaN. Returning N00 is correct: when both Gaussians
// carry the same variance, the residual is a point mass at 0.
return Gaussian::from_ms(0.0, 0.0);
}
let mu = self.mu() - other.mu();
Self::from_ms(mu, var.sqrt())
}
pub(crate) fn forget(&self, variance_delta: f64) -> Self {
Self {
mu: self.mu,
sigma: (self.sigma.powi(2) + variance_delta).sqrt(),
}
let var = self.sigma().powi(2) + variance_delta;
Self::from_ms(self.mu(), var.sqrt())
}
}
impl Default for Gaussian {
fn default() -> Self {
Self {
mu: MU,
sigma: SIGMA,
}
Self::from_ms(MU, SIGMA)
}
}
impl ops::Add<Gaussian> for Gaussian {
type Output = Gaussian;
/// Variance addition: (mu1 + mu2, sqrt(σ1² + σ2²)).
/// Used for combining performance and noise; rare relative to mul/div.
fn add(self, rhs: Gaussian) -> Self::Output {
Gaussian {
mu: self.mu + rhs.mu,
sigma: (self.sigma.powi(2) + rhs.sigma.powi(2)).sqrt(),
}
let mu = self.mu() + rhs.mu();
let var = self.sigma().powi(2) + rhs.sigma().powi(2);
Self::from_ms(mu, var.sqrt())
}
}
impl ops::Sub<Gaussian> for Gaussian {
type Output = Gaussian;
/// (mu1 - mu2, sqrt(σ1² + σ2²)). Same sigma combination as Add.
fn sub(self, rhs: Gaussian) -> Self::Output {
Gaussian {
mu: self.mu - rhs.mu,
sigma: (self.sigma.powi(2) + rhs.sigma.powi(2)).sqrt(),
}
let mu = self.mu() - rhs.mu();
let var = self.sigma().powi(2) + rhs.sigma().powi(2);
Self::from_ms(mu, var.sqrt())
}
}
impl ops::Mul<Gaussian> for Gaussian {
type Output = Gaussian;
/// Factor product: nat-param add. Hot path — two f64 additions, no sqrt.
fn mul(self, rhs: Gaussian) -> Self::Output {
let (mu, sigma) = if self.sigma == 0.0 || rhs.sigma == 0.0 {
let mu = self.mu / (self.sigma.powi(2) / rhs.sigma.powi(2) + 1.0)
+ rhs.mu / (rhs.sigma.powi(2) / self.sigma.powi(2) + 1.0);
let sigma = (1.0 / ((1.0 / self.sigma.powi(2)) + (1.0 / rhs.sigma.powi(2)))).sqrt();
(mu, sigma)
} else {
mu_sigma(self.tau() + rhs.tau(), self.pi() + rhs.pi())
};
Gaussian { mu, sigma }
Self::from_natural(self.pi + rhs.pi, self.tau + rhs.tau)
}
}
impl ops::Mul<f64> for Gaussian {
type Output = Gaussian;
fn mul(self, rhs: f64) -> Self::Output {
if rhs.is_finite() {
Self {
mu: self.mu * rhs,
sigma: self.sigma * rhs,
}
} else {
N_INF
fn mul(self, scalar: f64) -> Self::Output {
if !scalar.is_finite() {
return N_INF;
}
if scalar == 0.0 {
// Scaling by 0 collapses to a point mass at 0 (sigma' = 0, mu' = 0).
// This is N00, the additive identity, NOT N_INF.
return Gaussian::from_ms(0.0, 0.0);
}
// sigma' = sigma * |scalar| => pi' = pi / scalar²
// mu' = mu * scalar => tau' = tau / scalar
Self::from_natural(self.pi / (scalar * scalar), self.tau / scalar)
}
}
impl ops::Div<Gaussian> for Gaussian {
type Output = Gaussian;
/// Cavity: nat-param sub. Hot path — two f64 subtractions, no sqrt.
fn div(self, rhs: Gaussian) -> Self::Output {
let (mu, sigma) = if self.sigma == 0.0 || rhs.sigma == 0.0 {
let mu = self.mu / (1.0 - self.sigma.powi(2) / rhs.sigma.powi(2))
+ rhs.mu / (rhs.sigma.powi(2) / self.sigma.powi(2) - 1.0);
let sigma = (1.0 / ((1.0 / self.sigma.powi(2)) - (1.0 / rhs.sigma.powi(2)))).sqrt();
(mu, sigma)
} else {
mu_sigma(self.tau() - rhs.tau(), self.pi() - rhs.pi())
};
Gaussian { mu, sigma }
}
}
fn mu_sigma(tau: f64, pi: f64) -> (f64, f64) {
if pi > 0.0 {
(tau / pi, (1.0 / pi).sqrt())
} else if (pi + 1e-5) < 0.0 {
panic!("precision should be greater than 0");
} else {
(0.0, f64::INFINITY)
Self::from_natural(self.pi - rhs.pi, self.tau - rhs.tau)
}
}
@@ -148,85 +164,71 @@ mod tests {
#[test]
fn test_add() {
let n = Gaussian {
mu: 25.0,
sigma: 25.0 / 3.0,
};
let m = Gaussian {
mu: 0.0,
sigma: 1.0,
};
assert_eq!(
n + m,
Gaussian {
mu: 25.0,
sigma: 8.393118874676116
}
);
let n = Gaussian::from_ms(25.0, 25.0 / 3.0);
let m = Gaussian::from_ms(0.0, 1.0);
let r = n + m;
assert!((r.mu() - 25.0).abs() < 1e-12);
assert!((r.sigma() - 8.393118874676116).abs() < 1e-10);
}
#[test]
fn test_sub() {
let n = Gaussian {
mu: 25.0,
sigma: 25.0 / 3.0,
};
let m = Gaussian {
mu: 1.0,
sigma: 1.0,
};
assert_eq!(
n - m,
Gaussian {
mu: 24.0,
sigma: 8.393118874676116
}
);
let n = Gaussian::from_ms(25.0, 25.0 / 3.0);
let m = Gaussian::from_ms(1.0, 1.0);
let r = n - m;
assert!((r.mu() - 24.0).abs() < 1e-12);
assert!((r.sigma() - 8.393118874676116).abs() < 1e-10);
}
#[test]
fn test_mul() {
let n = Gaussian {
mu: 25.0,
sigma: 25.0 / 3.0,
};
let m = Gaussian {
mu: 0.0,
sigma: 1.0,
};
assert_eq!(
n * m,
Gaussian {
mu: 0.35488958990536273,
sigma: 0.992876838486922
}
);
let n = Gaussian::from_ms(25.0, 25.0 / 3.0);
let m = Gaussian::from_ms(0.0, 1.0);
let r = n * m;
assert!((r.mu() - 0.35488958990536273).abs() < 1e-10);
assert!((r.sigma() - 0.992876838486922).abs() < 1e-10);
}
#[test]
fn test_div() {
let n = Gaussian {
mu: 25.0,
sigma: 25.0 / 3.0,
};
let n = Gaussian::from_ms(25.0, 25.0 / 3.0);
let m = Gaussian::from_ms(0.0, 1.0);
let r = m / n;
assert!((r.mu() - (-0.3652597402597402)).abs() < 1e-10);
assert!((r.sigma() - 1.0072787050317253).abs() < 1e-10);
}
let m = Gaussian {
mu: 0.0,
sigma: 1.0,
};
#[test]
fn test_n00_is_add_identity() {
// N00 (sigma=0) is the additive identity for the variance-convolution Add op.
// N_INF (sigma=inf) is the identity for the EP-product Mul op.
let g = Gaussian::from_ms(3.0, 2.0);
let n00 = Gaussian::from_ms(0.0, 0.0);
let r = n00 + g;
assert!((r.mu() - g.mu()).abs() < 1e-12);
assert!((r.sigma() - g.sigma()).abs() < 1e-12);
}
assert_eq!(
m / n,
Gaussian {
mu: -0.3652597402597402,
sigma: 1.0072787050317253
}
);
#[test]
fn test_mul_is_factor_product() {
// n * m in nat-params should be pi_n + pi_m, tau_n + tau_m
let n = Gaussian::from_ms(2.0, 3.0);
let m = Gaussian::from_ms(1.0, 2.0);
let r = n * m;
let expected_pi = n.pi() + m.pi();
let expected_tau = n.tau() + m.tau();
assert!((r.pi() - expected_pi).abs() < 1e-15);
assert!((r.tau() - expected_tau).abs() < 1e-15);
}
#[test]
fn test_div_is_cavity() {
let n = Gaussian::from_ms(2.0, 1.0);
let m = Gaussian::from_ms(1.0, 2.0);
let r = n / m;
let expected_pi = n.pi() - m.pi();
let expected_tau = n.tau() - m.tau();
assert!((r.pi() - expected_pi).abs() < 1e-15);
assert!((r.tau() - expected_tau).abs() < 1e-15);
}
}