Pseudo inverse very inaccurate

divB

Hi,

I've a very trivial numerical problem where I'm currently stuck. In MATLAB the matrix Hf:

Code:
>> Hf
Hf =

1.0e+003 *

1.6443    1.6516    1.6583
4.8373    4.8349    4.8334
4.6385    4.6418    4.6445
-9.6014   -9.6084   -9.6154
And the following vectors which are very close:

Code:
>> [yl1 , yl3 , yl1 - yl3 ]
ans =

1.0e+006 *

0.2966    0.2972   -0.0006
0.8705    0.8703    0.0002
0.8352    0.8355   -0.0003
-1.7288   -1.7295    0.0006
yl1 is my result as it should be:

Code:
>> Hf \ yl1
ans =

100.0000
75.0000
5.0000
yl3 is obtained in a different way but is very close to the original. But sill:

Code:
>> Hf \ yl3
ans =

56.0412
72.5578
51.4007
The result is not just a little bit away, it is terrible, unuseable!

I have much redundancy in the data, so I can ad much lines to the matrix Hf. However, it does not matter how much, the result is always the same ... unuseable.

Can anyone explain why the least squares is so terrible in this case? I'm a bit confused because least squares should be pretty robust ...

Thanks,
divB

divB

Hmm, sorry I think I know: The columns of Hf are too close, right?

Number Nine

Hmm, sorry I think I know: The columns of Hf are too close, right?
Likely, yes. The columns are very nearly linearly dependent, which can make calculations like the pseudoinverse numerically unstable.

divB

Hmm, thanks. Is there a good way to stabilize such a system? E.g. I have y=Hf*h; I want to find h and may modify y and Hf

chiro

It's been a while since I did this myself, but have you tried looking at the various numeric iterative schemes to solve systems with a high condition number?

divB

Hmm, I wonder why MATLAB would not implement these itself?
Anycase, I get similar results when the condition number is not so high, e.g. cond(Hf) = 183. Starting from which condition number could one say that the LS becomes unstable?

I am currently trying to reformulate the problem to avoid these similar columns. In my current setup I sum over a series where each column differs only with one sample of the series. Clearly, this yields very similar numbers in each column per row.

However, I am currently confused about the linear dependency of the rows/columns in such a linear system. Can you clearify that? I thought for the solvability, only the linear dependency of the rows (rather than the cols) plays a role?

E.g. when there are fewer lin. independent rows than unknowns, then the system cannot be solved.

So does this mean that the rows AND columns must be lin. independent?

I found no clear explanation.

chiro

That is an extremely high condition number: 183 is very big.

AlephZero

Homework Helper
Instead of Hy \ y1, try pinv(Hf)*y1. That may give you more "consistent" results.

Judging by http://www.mathworks.co.uk/help/matlab/ref/arithmeticoperators.html, \ uses a fairly simple minded (but quick) algorithm to solve badly conditioned problems. The bottom line is that the solution isn't unique, because there is a nonzero vector v such that Hf.v = 0, and you can add any multiple of v to the "solution" to get another solution. And the algorithm that \ uses doesn't care what that multiple is, but the algorithm for pinv() does care - it also tries to make the components of the "answer" as small as possible.

Physics Forums Values

We Value Quality
• Topics based on mainstream science
• Proper English grammar and spelling
We Value Civility
• Positive and compassionate attitudes
• Patience while debating
We Value Productivity
• Disciplined to remain on-topic
• Recognition of own weaknesses
• Solo and co-op problem solving