Understanding the implications of singular matrices

47 Views Asked by At

Context

An over-determined system of equations can not be generally solved by elimination operated on $Ax = b$.

Instead, it normally involves solving $$(A^t A) x = A^t b$$

which is obtained in the OLS method., this is, assuming we choose the sum of squared errors as the cost function.

Simplifying the $A$s product inverse to just $B$ then we could to LU decomposition $ B = LU $ and if $U$ has some zeroes in the diagonal then:

  • some columns are linearly dependent
  • it means $r<n$,
  • and $Det(B)=0$,
  • and $B^{-1}$ does not exist.

But does this means we can't find a solution ? Could you hint me why it is so conceptually, and how it relates to the obvious properties listed above ?