DiagonalNoise#
- class hybrid_aif.inference.DiagonalNoise[source]#
Bases:
NamedTupleAn observation noise of independent readings,
N(0, diag(scale²)), in the form the state-space update reads.scale(..., d_obs)holds the readings’ standard deviations, which must be strictly positive (a zero or negative value is not checked and gives an infinite or wrong log determinant).low_rank(..., d_obs, k), when given, adds a correlated part of rank at mostk: the noise covariance is thendiag(scale²) + low_rank low_rankᵀ(the residual covariance of an unscented reading,linear_model()). The update treats that part askfurther standard-normal latent coordinates read throughlow_rank, so it stays in state space at dimensiond + k. The update checks the shapes (statically, at no cost underjit):scale’s last axis is the reading’s length, its batch axes are no more than the head’s (a Cholesky factor passed asscaleis refused), andlow_rank’sd_obsaxis matchesscale’s.- static __new__(_cls, scale, low_rank=None)#
Create new instance of DiagonalNoise(scale, low_rank)