# Difference between lsq(A,b) and A\b (on Scilab)

Can you explain me the difference between lsq(A,b) and A\b? Why do I get a positif solution when I use lsq(A,b,1)? Where can I get the source code of lsq function?

Thank you.

-

the difference between lsq(A,b) and A\b?

There should be no difference when $Ax=b$ has a unique solution. The differences arise when the system is underdetermined or inconsistent.

If $Ax=b$ has multiple solutions, A\b will find one, and call it good. lsq(A,b) will find the solution of smallest norm. For example, consider $x_1+x_2=1$:

• [1 1]\[1] returns $x_1=1$, $x_2=0$
• lsq([1,1],[1]) returns $x_1=0.5$, $x_2=0.5$

If there are no exact solutions, A\b proceeds to look for the least squares solution, and may return the same output as lsq(A,b). For example, try the inconsistent system $x=1$, $x=2$:

• both [1;1]\[1;2] and lsq(1;1],[1;2]) return $x=1.5$.

However, the algorithm of A\b appears to be less optimized for the task of least squares solution. Consider the inconsistent system $$\begin{cases} x_1+x_2&=1 \\ x_1+x_2&=2 \end{cases}$$

• [1 1;1 1]\[1;2] displays a warning message and returns $x_1=-\infty$, $x_2=\infty$.
• lsq([1 1;1 1],[1;2]) returns $x_1=x_2=0.75$ with no drama.

Why do I get a positif solution when I use lsq(A,b,1)?

Depends on your $A$ and $b$. For example, lsq([1],[-1],1) returns $-1$, as it should.

Where can I get the source code of lsq function?

Scilab source code is available. Also, Scilab documentation says that the lsq function is based on the LApack functions dgelsy (for real matrices) and zgelsy (for complex matrices). LApack source is also available.

-

I though A\b is using the LU approach. This is my opinion, may be there are slight different in the result since A\b use direct method, while lsq is using approximation via least square

-