I'm a linear algebra weakling and am looking for help anywhere I can find it.

I'm attempting to calculate the inverse of a 14x14 square matrix using Gauss-Jordan elimination. This seems to work for some matrix input, but not all. I end up with an almost zero value on the diagonal, which really blows up the scaling of the values in the matrix.

I'm comparing my results to those calculated by Matlab. For my initial set of matricies, the results match (within rounding differences). But for the following matrix, they do not. (sorry for the large data set: if there's a better way to show this please let me know).

Your matrix is very badly conditioned and will be extremely sensitive to numerical errors. If you want the same results as Matlab, then you're simply going to have to do the same calculations as Matlab.

As far as I remember, the Gauss-Jordan method is one of the worst general matrix inversion algorithms when it comes to numerical precision. You really need to feed it well-conditioned matrices for it to work.

@ Brother Bob; I'm sure that's good advice. The problem is I don't understand enough for it to help me. I wouldn't know how to properly condition the matrix. I also do not know how Matlab is coming up with its answer.

@ unbird: I'm looking at your linked article. My current solution claims to use row-pivoting; I'm going to attempt to better understand pivoting to see if: A) i'm already pivoting and B) if I'm not, if doing so will help.

Thanks.

I ran into this problem years ago and it looks like you may be pivoting on roundoff error.

Try partial pivoting.

http://en.wikipedia.org/wiki/Pivot_element

@ Brother Bob; I'm sure that's good advice. The problem is I don't understand enough for it to help me. I wouldn't know how to properly condition the matrix. I also do not know how Matlab is coming up with its answer.

The condition of the matrix is a property of the matrix itself and not something you can just adjust to improve it. Your specific matrix has a specific condition, and the specific matrix you posted is not well conditioned. In this context, you could see the condition of the matrix as how numerically sensitive it is to perform calculations on it.

If the exact matrix itself is not important but its general structure is, then you can improve the condition by a technique called regularization. It basically means adding a constant value to the diagonal of your matrix, effectively turning your matrix M into the new matrix M'=M+r*I where r is called a regularization factor and I is the identity matrix.

Matlab has a lot of algorithms it uses to solve problems where matrix inversion might be used. Which one it uses depends on its own analysis of the input matrix to determine what answer is most correct. In fact, Matlab tells you not to use the result of inv(A) to solve an Ax=b problem because of numerical error, probably because inv(A) is a quick naive Gauss elimination algorithm.

Looking at your matrix, it could be using a sparse matrix algorithm because a lot of the entries are zero, but the size of the matrix makes me think it's not. If you're not explicitly using a specific algorithm in your code, I think there is a command for it to tell you which algorithms it's using to solve your problem. By the way, which command are you using to get the inverse in Matlab?

For your information, naive Gaussian elimination is not numerically stable because of the mathematical roundoff errors that arise from the precision in the computer. Even though you might be calculating with double precision, tiny errors due to roundoff can compound to throw off the entire result if not checked. Partial pivoting checks these small errors so they don't become big. Complete pivoting does this even better, but takes more computations to do this. However, if you're trying to solve an Ax=B problem, don't use matrix inversion. There are better techniques for this, such as LU decomposition.