Struct LogKernelSumJet 

pub struct LogKernelSumJet {
    pub value: f64,
    pub d1: f64,
    pub d2: f64,
    pub d3: f64,
    pub d4: f64,
    pub mode: IntegratedExpectationMode,
}

Derivatives of log(Σ_j a_j · K_{k_j, m_j}(μ, σ)) with respect to μ.

This is the workhorse for row-level log-likelihood derivatives in all latent families. The numerator and denominator of a row likelihood are each a small signed sum of kernel terms.

The value path is assembled from log-space kernel bundles and ratio jets, so individual kernel terms are never exponentiated before the final signed sum. That avoids the old overflow/underflow problems from value-space kernels. When the signed sum is zero or negative, this returns an invalid row (value = -∞) instead of trying to continue with a floored surrogate. Signed two-term differences (e.g. interval censoring K_{0,M_L} − K_{0,M_R}) are still combined through the shared sign-aware log-sum path.

Fields

value: f64

log(Σ a_j K_j)

d1: f64

d/dμ log(Σ a_j K_j)

d2: f64

d²/dμ² log(Σ a_j K_j)

d3: f64

d³/dμ³ log(Σ a_j K_j)

d4: f64

d⁴/dμ⁴ log(Σ a_j K_j)

mode: IntegratedExpectationMode

Implementations

impl LogKernelSumJet

pub fn single_term( quadctx: &QuadratureContext, k: usize, m: f64, mu: f64, sigma: f64, ) -> Result<Self, EstimationError>

Evaluate for a single positive kernel term (fast path).

Computes log(K_{k,m}) and its μ-derivatives from exact recurrences, entirely in log-space.

pub fn evaluate( quadctx: &QuadratureContext, terms: &[KernelSumTerm], mu: f64, sigma: f64, ) -> Result<Self, EstimationError>

Evaluate log(Σ a_j K_j) and its μ-derivatives for a small signed sum.

All terms share the same (μ, σ). Both the value and derivative ratios are computed entirely in log-space. The runtime latent-survival rows in this repo are almost always one-term or two-term sums, so those cases stay on dedicated stack paths; the heap-backed logic below is only for genuinely longer symbolic sums:

1. Per-term log-magnitudes log|a_j| + log K_{k_j,m_j} and signs.
2. Sign-aware log-sum-exp to get log|S| and sign(S).
3. Importance weights w_j = a_j K_j / S formed in log-space.
4. Weighted ratio sums R_n = Σ w_j · (∂ⁿK_j / K_j) for the final log-derivatives.

Trait Implementations

impl Clone for LogKernelSumJet

fn clone(&self) -> LogKernelSumJet

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for LogKernelSumJet

impl Debug for LogKernelSumJet

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.