I am performing SVD on a matrix A where each column of A contains an image. The algorithms i am using have been thoroughly tested using matlab.

However when using the numerical recipes SVD some of the columns in the U and V matrix are sometimes the negative of the matlab columns. Therefore my images (basis images in the columns of the U matrix) are inverted and this is not acceptable. I have NEVER seen matlab invert the images.

So the question is... is there a way to correct the U/V matrices as a post processing step to make them the same as MATLAB

OR

Is numerical recipes not suitable for such an application and i need to use another library.

I have read the NR book and have not seen this effect mentioned anywhere .... can anyone point me to the page?

Thanks in advance.

After more testing matlab this too shows a sign ambiguity... sometimes choosing the alternate solution that recipes picks.
Reading some math texts this is perfectly valid.
My misunderstanding so i need to fix my algorithm.

By the way, the line " g = -SIGN(sqrt(s),f); "
fails with visual studio 6 compiler in release mode. Split the square root onto a separate line to get it to compile.