12
$\begingroup$

What is the difference between LU Decomposition and Cholesky Decomposition about using these methods to solving linear equation systems?

Could you explain the difference with a simple example?

Also could you explain the differences between these decomposition methods in:

  • inverse of a matrix
  • forward and backward substitution
  • pivoting
$\endgroup$
  • 5
    $\begingroup$ The Cholesky decomposition is simply a particular case of the LU decomposition for symmetric (hermitian in the complex world) positive definite matrices, and those only. I think wikipedia has a decent page about it. en.wikipedia.org/wiki/Cholesky_decomposition $\endgroup$ – bartgol Aug 3 '16 at 21:36
4
$\begingroup$

Big question with a lot of possible tangents one could go down. I've tried to provide a somewhat brief summary.

In both an $LU$ factorization and a Cholesky $LL^T$ factorization, a square nonsingular matrix $A$ is factored into a product of triangular matrices.${}^*$ In a Cholesky factorization, the lower and upper triangular factors are transposes of each other—that is, an $LU$ factorization is a Cholesky factorization if $L^T=U$.

An example $LU$ factorization:

$$ \begin{pmatrix} 2 & 3 \\ 2 & 7 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 1 & 1\end{pmatrix}\begin{pmatrix} 2 & 3 \\ 0 & 4\end{pmatrix} $$

and a Cholesky factorization (which is also an $LU$ factorization)

$$ \begin{pmatrix} 1 & 1 \\ 1 & 2 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 1 & 1\end{pmatrix}\begin{pmatrix} 1 & 0 \\ 1 & 1\end{pmatrix}^T. $$

A matrix has a Cholesky factorization if, and only if, it is symmetric positive definite (SPD). If you try and compute a Cholesky factorization for matrix which is not SPD, it will always fail. However, for an SPD matrix, the Cholesky factorization is convenient in that one only has to store one triangular matrix $L$ rather a pair $L$ and $U$.

Once a matrix $A$ is factored as either an $LU$ or Cholesky factorization, then the linear system of equations $Ax=b$ can be solved by first solving $Ly=b$ by forward substitution and then $Ux=y$ (or $L^Tx=y$ for Cholesky) by backwards substitution. (The wikipedia page describes these quite nicely.) In either forward or backward substitution, the key insight is that, if you write out the equations for the unknowns $x_1,\ldots,x_n$ in the right order (forward for $L$ and backwards for $U$), each equation only possesses one new unknown than the previous equation, which can easily be computed.

If you need to compute $A^{-1}$, you can compute it from an $LU$ decomposition by the formula $A^{-1} = U^{-1}L^{-1}$. There exist methods to invert triangular matrices. In practical applications, it is widely accepted general wisdom to not compute a matrix inverse unless you have to, and there are usually ways around actually computing the inverse. (This, unfortunately, does not applying to linear algebra exams in school.)

When computing an $LU$ factorization, one may possibily encounter a zero pivot, which necessitates permuting the rows of the matrix to obtain a nonzero pivot. When pivoting is used, one actually computes a decomposition of the form $PA = LU$, where $P$ is a permutation matrix. When performing an $LU$ factorization on a computer using inexact arithmetic (such as floating point arithmetic), it is also important to permute when a pivot entry is simply small, not necessarily exactly zero. In fact, it is common to permute the matrix such that we always pick the largest pivot in the column, in a strategy known as partial pivoting. When performing Cholesky factorization on an SPD matrix, one will never encounter a zero pivot and one does not need to pivot to ensure the accuracy of the computation. (One may want to use permutations for other reasons, such as to maintain sparsity.)


${}^*$ One can do $LU$ factorization in a variety of ways for rectangular or rank deficient matrices as well, but the square case is the most common.

$\endgroup$
0
$\begingroup$

Both LU and Cholesky Decomposition is matrices factorization method we use for non-singular( matrices that have inverse) matrices. In general basic different between two method. the later one uses only for square matrices (A = A^T). however LU decomposition we can use any matrices that have inverses. for example see the following equation with 3 unknown 2x + y 3z = 4 2x - 2y -z = -1
-2 + 4y z = 1
the above equation since the coefficient of a variable can't form a square matrix so we use LU decomposition to factorize the matrices, basically two steps process first row reductions until make all zero below the main diagonal so that we find upper matrix (U) and save all the factors we use on this step to substitute to lower matrix(L) and put the values of 1's to the main diagonal and 0s to above the main diagonal. using either of two (L,U) we can easier solve the equation through back substitution values of x, y & z are 3/2, 1, 0. Cholesky factorization simple example Matrix A; 2x + y = 2 x + 2y = 4 this method use A = LL^T. Simply taking 2x2 lower triangular matrix multiply(components) with its transpose (with variables values).and matches with the coefficient of Matrix A and try to solve the unknown variable so that you can factor A with L, L^T.

$\endgroup$
  • 1
    $\begingroup$ Welcome to MSE! Your answer is rather messy; it would be helpful to future viewers if you edited it to make it look nicer. For some basic information about writing mathematics at this site see, e.g., here, here, here and here. $\endgroup$ – Sambo Nov 7 '18 at 3:04
  • $\begingroup$ You confuse square and symmetric matrices. $\endgroup$ – Yves Daoust May 14 at 13:05

Your Answer

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

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