Skip to main content
added the explanation how to calculate the solution
Source Link
mpiktas
  • 1.5k
  • 12
  • 28

Updated according to comments.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

Now since $R$ is triangular and $n>k$ we will have that the last $n-k$ rows of $R$ are zero. Since $Ax=b$ it follows that the last $n-k$ elements of $Q^Tb$ should be zero also. If they are not, then $x$ is not a solution.

Furthermore since we have an overdetermined matrix the solution exists only if $b$ lies in the linear space spanned by columns of $A$. So the real question is, how do we reliably check whether $b$ is in linear space spanned by columns of $A$.

Update 2

Since $n>k$, the $R$ matrix will look like:

\begin{align*} R=\begin{bmatrix} R_1\\\\ 0 \end{bmatrix} \end{align*} where $R_1$ is $k\times k$ upper triangular matrix. If the solution exists, then

\begin{align*} Q^Tb=\begin{bmatrix} b_1\\\\ 0 \end{bmatrix} \end{align*} where $b_1$ is $k\times 1$ vector. The solution for our system is then $$x=R_1^{-1}b_1$$

Updated according to comments.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

Now since $R$ is triangular and $n>k$ we will have that the last $n-k$ rows of $R$ are zero. Since $Ax=b$ it follows that the last $n-k$ elements of $Q^Tb$ should be zero also. If they are not, then $x$ is not a solution.

Furthermore since we have an overdetermined matrix the solution exists only if $b$ lies in the linear space spanned by columns of $A$. So the real question is, how do we reliably check whether $b$ is in linear space spanned by columns of $A$.

Updated according to comments.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

Now since $R$ is triangular and $n>k$ we will have that the last $n-k$ rows of $R$ are zero. Since $Ax=b$ it follows that the last $n-k$ elements of $Q^Tb$ should be zero also. If they are not, then $x$ is not a solution.

Furthermore since we have an overdetermined matrix the solution exists only if $b$ lies in the linear space spanned by columns of $A$. So the real question is, how do we reliably check whether $b$ is in linear space spanned by columns of $A$.

Update 2

Since $n>k$, the $R$ matrix will look like:

\begin{align*} R=\begin{bmatrix} R_1\\\\ 0 \end{bmatrix} \end{align*} where $R_1$ is $k\times k$ upper triangular matrix. If the solution exists, then

\begin{align*} Q^Tb=\begin{bmatrix} b_1\\\\ 0 \end{bmatrix} \end{align*} where $b_1$ is $k\times 1$ vector. The solution for our system is then $$x=R_1^{-1}b_1$$

added 460 characters in body
Source Link
mpiktas
  • 1.5k
  • 12
  • 28

Updated according to comments.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

whichNow since $R$ is triangular and $n>k$ we will have that the last $n-k$ rows of $R$ are zero. Since $Ax=b$ it follows that the last $n-k$ elements of $Q^Tb$ should be easier and numerically stablezero also. If they are not, then $x$ is not a solution.

Furthermore since we have an overdetermined matrix the solution exists only if $b$ lies in the linear space spanned by columns of $A$. So the real question is, how do we reliably check whether $b$ is in linear space spanned by columns of $A$.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

which should be easier and numerically stable.

Updated according to comments.

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

Now since $R$ is triangular and $n>k$ we will have that the last $n-k$ rows of $R$ are zero. Since $Ax=b$ it follows that the last $n-k$ elements of $Q^Tb$ should be zero also. If they are not, then $x$ is not a solution.

Furthermore since we have an overdetermined matrix the solution exists only if $b$ lies in the linear space spanned by columns of $A$. So the real question is, how do we reliably check whether $b$ is in linear space spanned by columns of $A$.

Source Link
mpiktas
  • 1.5k
  • 12
  • 28

If you are worried about the numerical stability do QR decomposition of matrix $A$. Then $A=QR$, where $Q$ is orthogonal and $R$ is triangular. Then you need to check whether $x$ satisfies the equation

$$Rx=Q^Tb$$

which should be easier and numerically stable.