hyperiax.prebuilt.bffg.discrete_bf_sweep

hyperiax.prebuilt.bffg.discrete_bf_sweep(prxy_scale_fn, prxy_shift_fn, prxy_covar_fn)[source]

Build the discrete-edge backward-filtering up-sweep (Theorem 14 §6.1).

At each non-leaf parent, pulls each child’s canonical message back through the linear-Gaussian auxiliary \(\tilde P(x, dy) = \mathcal N(y; \Phi x + \beta, Q)\,dy\) and fuses (sums) the contributions. The auxiliary is evaluated per-child at the child’s anchor.

Parameters:

  • prxy_scale_fn – Callable (anchor, params) -> (d, d) array returning the auxiliary scale matrix \(\Phi\).

  • prxy_shift_fn – Callable (anchor, params) -> (d,) array returning the auxiliary shift vector \(\beta\).

  • prxy_covar_fn – Callable (anchor, params) -> (d, d) SPD array returning the auxiliary covariance \(Q\).

Return type:

SweepFn

Returns:

A hyperiax.SweepFn that, when applied to a tree, writes (prec, ptnl, log_norm) at every non-leaf node.

Notes

Linear-Gaussian models (auxiliary equal to truth) ignore anchor; for nonlinear models, iterate this sweep with discrete_refine_anchor() (Algorithm 3 §7.1).