Householder method : Which α do we take?

  • Context: MHB 
  • Thread starter Thread starter mathmari
  • Start date Start date
  • Tags Tags
    Method
Click For Summary
SUMMARY

The discussion focuses on the application of the Householder method for QR factorization of the matrix $A=\begin{pmatrix}1 & 4 & -1 \\ 2 & 2 & 7 \\ 2 & -4 & 7\end{pmatrix}$. The first column vector $a_0=\begin{pmatrix}1 \\ 2 \\ 2\end{pmatrix}$ leads to the calculation of $\alpha=-3$ to maximize the norm difference. The resulting Householder matrix $H_1$ is computed as $H_1=\begin{pmatrix}-1/3 & -2/3 & -2/3 \\ -2/3 & 2/3 & -1/3 \\ -2/3 & -1/3 & 2/3\end{pmatrix}$, transforming $A$ into $A^{(1)}=\begin{pmatrix}-3 & 0 & -9 \\ 0 & 0 & 3 \\ 0 & -6 & 3\end{pmatrix}$. The discussion concludes that when the norms are equal, either choice of $\alpha$ will yield a valid QR decomposition, minimizing calculation errors equally.

PREREQUISITES
  • Understanding of QR factorization and its significance in numerical linear algebra.
  • Familiarity with the Householder transformation method for matrix decomposition.
  • Knowledge of vector norms, specifically $\ell_2$ norms.
  • Basic linear algebra concepts, including matrix operations and eigenvalues.
NEXT STEPS
  • Study the derivation and properties of the Householder transformation in detail.
  • Explore the implications of different choices of $\alpha$ in QR factorization.
  • Learn about numerical stability and error analysis in matrix decompositions.
  • Investigate alternative methods for QR factorization, such as Gram-Schmidt process.
USEFUL FOR

Mathematicians, data scientists, and engineers involved in numerical analysis, particularly those working with matrix decompositions and optimization techniques.

mathmari
Gold Member
MHB
Messages
4,984
Reaction score
7
Hey! 😊

We have the matrix $A=\begin{pmatrix}1 & 4 & -1 \\ 2 & 2 & 7 \\ 2 & -4 & 7\end{pmatrix}$ and we want to calculate the $QR$ factorisation using the Householder method.

First we take the first column of the matrix $a_0=\begin{pmatrix}1 \\ 2 \\ 2\end{pmatrix}$.
We have that $\alpha=\pm \|a_0\|=\pm \sqrt{9}\pm 3$.
Then we take the sign of that so that $\|a_0-\alpha e_1\|_2$ is big.
We have $$\|a_0-\alpha e_1\|_2=\begin{cases}\|\begin{pmatrix}1 \\ 2 \\ 2\end{pmatrix}-\begin{pmatrix}3 \\ 0 \\ 0\end{pmatrix}\|_2 =\|\begin{pmatrix}-2 \\ 2 \\ 2\end{pmatrix} =\sqrt{12}\\ \|\begin{pmatrix}1 \\ 2 \\ 2\end{pmatrix}-\begin{pmatrix}-3 \\ 0 \\ 0\end{pmatrix}\|_2 =\|\begin{pmatrix}4 \\ 2 \\ 2\end{pmatrix}\end{cases} =\sqrt{20}$$
So we take $\alpha=-3$.
Then $u=\frac{1}{\|a_0-\alpha e_1\|_2}\begin{pmatrix}4 \\ 2 \\ 2\end{pmatrix}=\begin{pmatrix}\frac{4}{\sqrt{20}} \\ \frac{2}{\sqrt{20}} \\ \frac{2}{\sqrt{20}}\end{pmatrix}$.
Then $$H_1=I-2uu^T=\cdots =\begin{pmatrix}-1/3 & -2/3 & -2/3 \\ -2/3 & 2/3 & -1/3 \\ -2/3 & -1/3 & 2/3\end{pmatrix}$$
Then $A^{(1)}=H_1\cdot A=\begin{pmatrix}-3 & 0 & -9 \\ 0 & 0 & 3 \\ 0 & -6 & 3\end{pmatrix}$.
Then we consider the first column of the $2\times 2$-submatrix, $a_1=\begin{pmatrix}0 \\ -6\end{pmatrix}$.
Then $\alpha=\pm \|a_1\|=\pm 6$.
We have $$\|a_1-\alpha e_1\|_2=\begin{cases}\|\begin{pmatrix}0 \\ -6\end{pmatrix}-\begin{pmatrix}6\\ 0\end{pmatrix}\|_2 =\|\begin{pmatrix}-6 \\ -6\end{pmatrix} =6\sqrt{2}\\ \|\begin{pmatrix}0\\ -6\end{pmatrix}-\begin{pmatrix}-6 \\ 0\end{pmatrix}\|_2 =\|\begin{pmatrix}6 \\ 0\end{pmatrix}\end{cases} =6\sqrt{2}$$
In this case the norms are equal, which $\alpha$ do we take then?

:unsure:
 
Physics news on Phys.org
mathmari said:
In this case the norms are equal, which $\alpha$ do we take then?
Since the norms are equal, the impact on the calculation error is the same.
So it doesn't matter which one we pick. You can choose. 🤔
 
Klaas van Aarsen said:
Since the norms are equal, the impact on the calculation error is the same.
So it doesn't matter which one we pick. You can choose. 🤔

But the result that we get will be different, or not? :unsure:
 
mathmari said:
But the result that we get will be different, or not?
We may find a different QR decomposition, but the impact of calculation errors will have been minimized to the same extent. 🤔
 
Klaas van Aarsen said:
We may find a different QR decomposition, but the impact of calculation errors will have been minimized to the same extent. 🤔

Ok! Thank you! 🤩
 

Similar threads

  • · Replies 2 ·
Replies
2
Views
1K
  • · Replies 9 ·
Replies
9
Views
3K
  • · Replies 2 ·
Replies
2
Views
2K
  • · Replies 21 ·
Replies
21
Views
4K
  • · Replies 10 ·
Replies
10
Views
3K
  • · Replies 5 ·
Replies
5
Views
2K
  • · Replies 13 ·
Replies
13
Views
4K
  • · Replies 8 ·
Replies
8
Views
1K
  • · Replies 9 ·
Replies
9
Views
5K
  • · Replies 7 ·
Replies
7
Views
2K