PrecisionRegressionResult
Full multivariate-Gaussian posterior in square-root information form: the joint uncertainty is carried as the Cholesky factor of the posterior precision H, and the covariance is H⁻¹.
Every consumer reaches the covariance through a triangular solve rather than an inverse. A Thompson draw w ~ N(mean, S) is mean + L⁻ᵀ u for u ~ N(0, I), because L⁻ᵀ (L⁻ᵀ)ᵀ = H⁻¹; a predictive standard deviation sqrt(xᵀ S x) is the 2-norm of L⁻¹ x. Both are O(n²), the same cost the covariance form paid, and neither needs the n×n inverse materialised.
Constructors
PrecisionRegressionResult
Properties
bias
biasPrecision
Posterior precision (inverse variance) on the bias term.
link
precisionL
Cholesky factor of the posterior precision over weights: H = L·Lᵀ, covariance S = H⁻¹. Only the lower triangle is meaningful, matching koblas's factor convention.
sse
step
Number of com.eignex.kumulant.core.RegressionStat.update calls absorbed; useful as a bookkeeping counter for learning-rate decay or retraining cadence.
totalWeights
Cumulative observation weight folded in.
weights
Number of features in weights.
Mean squared error: sse / totalWeights. Zero on an empty stream.
Unbiased sample standard deviation: sqrt([sampleVariance]).
Unbiased sample variance: [sst] / ([totalWeights] - 1). Zero when totalWeights <= 1.
Functions
covariance
Posterior covariance S = (L·Lᵀ)⁻¹ materialised from precisionL. O(n³) and a fresh n×n matrix per call, so it belongs in reporting and prior fitting; scoring paths should stay on the factor.
covarianceInto
Linear predictor eta = bias + x . weights, before the inverse link.
Mean response: link.invMean(linearPredictor(x)). For Link.Identity this is the linear predictor itself, matching plain linear regression.