2
$\begingroup$

I am trying to prove mathematically that the update step in a Kalman filter can not result in a increase in uncertainty. I found the following proof which is based on the inversion lemma and the concept of positive semidefiniteness, but I can't grasp why the matrix product in the end is positive semi-definite.

Proof:

The uncertainty in the belief after the update step is expressed by the covariance matrix :

$\Sigma_t = (I - K_tC_t)\bar{\Sigma_{t}}$ (1)

where $\bar{\Sigma_t}$ is the predicted covariance before the update and $K_t$ is the Kalman Gain term, which is defined as:

$K_t = \bar{\Sigma_t}{C_t}^T(C_t\bar{\Sigma_t}{C_t}^T + {Q_t})^{-1}$ (2)

where $C_t$ and $Q_t$ are the parameters of the $z_t$ measurement's distribution $p(z_t | x_t) \sim \mathcal{N}(C_tx_t, Q_t) $

By expanding the $K_t$ term in equation (1) and then by applying the matrix inversion lemma/Woodburry matrix identity, we get:

$ \Sigma_t = \bar{\Sigma_{t}} - \bar{\Sigma_t}{C_t}^T(C_t\bar{\Sigma_t}{C_t}^T + {Q_t})^{-1}C_t\bar{\Sigma_t} \\ = (\bar{\Sigma}^{-1}_t + {C_t}^T{Q^{-1}_t}C_t)^{-1}$

The proof mentions that ${C_t}^T{Q^{-1}_t}C_t$ is positive semi-definite and hence the uncertainty cannot increase.

End of proof

Question: I know that $Q_t$ is positive semi-definite since it is a covariance matrix but I haven't been able to understand why ${C_t}^T{Q^{-1}_t}C_t$ is positive semi-definite. Can someone explain?

$\endgroup$

1 Answer 1

2
$\begingroup$

$Q_t$ is real-valued and positive definite, thus $Q_t^{-1}$ is real-valued and positive definite.

Now it's just making up a lemma of the Cholesky decomposition:

  • If $Q_t^{-1}$ is real-valued and positive definite, then there's some real-valued $q$ such that $q_t^T q_t = Q_t^{-1}$ (Cholesky decomposition)
  • If that holds, then $\left(q_t C_t \right)^T\left(q_t C_t \right)$ must be real-valued and at least positive semidefinite for any real-valued $C_t$ (Cholesky again -- that guy gets around)
  • $\left(q_t C_t \right)^T\left(q_t C_t \right) = C_t^T q_t^T q_t C_t = C_t^T Q_t^{-1} C_t$
  • Cogito, ergo sum. Or QED$^*$, or something brainy, and Latin.

${\tiny ^* {\rm quod\ erat\ demonstrandum.}}$

$\endgroup$
3
  • 2
    $\begingroup$ BTW: I knew this in my bones, but I had to whomp up the proof -- thanks for the exercise! $\endgroup$
    – TimWescott
    Nov 9, 2021 at 0:56
  • 1
    $\begingroup$ I don't see the second point "$\left(q_t C_t \right)^T\left(q_t C_t \right)$ must be real-valued and at least positive semidefinite for any real-valued $C_t$" straightforward. Could you please elaborate? Thanks. $\endgroup$
    – AlexTP
    Nov 9, 2021 at 9:44
  • 1
    $\begingroup$ If there is some matrix $L$ with all real-valued entries, of any rank, then $L^T L$ is positive semidefinite. If $L$ is of rank $n$ and $L^T L$ is $n \times n$ then $L^T L$ is positive definite. So let $q_t C_t = L$. $\endgroup$
    – TimWescott
    Nov 9, 2021 at 16:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.