In most of the papers I have read, calculation of weights is done by the following formula:
$$W_m=\lambda/(L+\lambda)$$ $$W_c=\lambda/(L+\lambda)+(1−\alpha^2+\beta)$$
Where $L$ is the dimensionality of my state and $\lambda$ is calculated by:
$$\lambda=\alpha^2(L+\kappa) - L$$
And the values for $\alpha$, $\beta$ and $\kappa$ are given by $\alpha=0.001$, $\beta=2$ (for Gaussian distributions) and $\kappa=0$.
My problem is that this would give a value of $\lambda$ that is very close to $-L$, which drives the weights to huge negative values. This is affecting my covariance estimation, leading to wildly inaccurate predictions. This seems like a fundamental error in the way the weights are calculated, but I'm not sure if I'm missing some critical intuition here.
