A = LU decomposition when A is symmetric
The LU decomposition of a symmetric matrix is special in that it can be further decomposed as follows:
when \( A \) is a symmetric matrix.
Intuition
Consider the first column of a symmetric matrix \( A \), and remember that it is equal to the first row. During elimitation, \( L \) records in it's first column the multipliers used to eliminate the first column by the first pivot. But then, when the first row of \( A \) is scaled by the same multiplier to achieve a unity pivot, the elements in the first row get scaled by the same factor. These values are recorded in the first row of \( U \) then scaled by decomposing into \( DU' \), but we find that \( U' = L^T \) by the above argument.
Illustrated with an example
The matrix:
Goes through the steps of elimination:
Corresponding to the elimination matrices:
Knowing the elimination matricies gives us \( L \) by filling in the multiplies as the negative of the values in the elimination matrices:
And so we have:
So far, this is standard practice. Now we seek to normalize the diagonals of \( U \) to 1. We do this with a diagonal matrix \( D \):
Notice that \( U' = L^T \). How is this not a surprise? The entries \( \bigl( \begin{smallmatrix} 1 \\ 0.5 \\ 0 \end{smallmatrix} \bigr) \) are the \( \bigl( \begin{smallmatrix}4 \\ 2 \\ 0 \end{smallmatrix} \bigr) \) entries of \( A \) after scaling by the first pivot. But the act of separating out \( D \) from \( U \) causes the exact same scaling to occur in the first row of \( U \) which is also the first row of \( A \). This argument can be extended sequentially to the whole matrix as it undergoes elimination and then decomposition with \( D \).