compose

Contents

compose#

hybrid_aif.compose(kernel, n)[source]#

The kernel of n repeats of a per-tick linear-Gaussian kernel, per mode.

With x' = F x + G u + b + eps and eps ~ N(0, Q) per tick (a control input held over the step), n ticks give F_n = F^n, b_n = sum_{i<n} F^i b, Q_n = sum_{i<n} F^i Q F^iᵀ and G_n = sum_{i<n} F^i G. A per-mode F or G (leading axes innovations.parent_shape) is composed mode by mode; a shared one stays shared, and an innovation noise shared across the modes (ConditionalGaussian.shared_chol) under a shared F stays shared (under a per-mode F the composed noise is per mode). n is a static positive int. The returned innovation is a ConditionalGaussian carrying the Cholesky factor of Q_n (a kernel with a singular Q per tick needs n ticks to be non-singular in the directions it reaches, as the Cholesky factor requires). Design: docs/design/step-durations.md T3.