Efficient Sensor Fusion Through Covariance-Constrained Observation Decimation (CCOD)

What the math gives to ML

The paper provides a constructive covariance-growth mechanism for skipping intermediate observations: over a decimation interval of length d, the dynamics become A^d and accumulated process noise becomes Q_d = sum from j=0 to d-1 of A^j Q (A^j)^T. A single Riccati solve then predicts the steady-state uncertainty of the decimated estimator, enabling selection of the largest d satisfying a prescribed covariance bound. This transfers naturally to neural state-space models and recurrent world models by using a local Jacobian for A, learned residual covariance for Q, and an observation-head Jacobian for H. The resulting controller can adaptively skip encoder, sensor, or cross-attention updates while retaining a falsifiable uncertainty boundary.