Let's say I have two sets of paired points $\{x_i\}_{i=1..n}$, $\{y_i\}_{i=1..n}$ defined in an $N$-dimensional space and I want to use the Kabsch algorithm to estimate the best rotation matrix / translation vector that transform each point in $\{y_i\}_{i=1..n}$ into its corresponding version in $\{x_i\}_{i=1..n}$ by solving the Orthogonal Procrustes problem based on the RMSD (root mean squared deviation) criterion:

\begin{equation} R = \arg \min_{R} \|R Y - X {\|}_F \end{equation} where:

  • $R$ is a rotation matrix (${R}^T R = I_N$),
  • $\| . {\|}_F$ is the Frobenius norm and $X$,
  • $Y$ is the matrix representation of $\{x_i\}_{i=1..n}$ and $\{y_i\}_{i=1..n}$.

Question: Knowing that the distribution of the random variable $x$ which generates each $x_i$ is multivariate Gaussian, is there any theoretical garentees about the distribution of the estimated vectors $\hat{y}_i$ using this algorithm ?

Attempt: I have doubts about this since the procedure is not Bayesian and there is no use of prior knowledge about the distribution of the target random variable $p(x)$. But I don't know whether this information is "implicitly" accounted for in the model.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.