Proof that real quadratic forms are always diagonalisable

73 Views Asked by At

I'm having some trouble understanding Theorem 1 in Chapter 8 of Lax's book on Linear Algebra. This chapter is "spectral theory of self-adjoint mappings of a Euclidean space". The theorem is that

Given a real quadratic form $$q(y) = \sum_{i,j} h_{ij} y_i y_j$$ it is possible to change variables as in $Ly=z$ s.t. in terms of the new variabls, $z, q$ is diagonal, i.e. of the form $$q \left( L^{-1} z \right) = \sum_{1}^{n}d_iz_i^2 \tag{11}\label{11}$$


The proof is as follows.

The proof is entirely elementary and constructive. Suppose that one of the diagonal elements of $q$ is non-zero, say $h_{11}\ne0$. We then group together all terms containing $y_1$ $$q(y)=h_{11}y_1^2+\sum_{2}^{n}h_{1j}y_1y_j+\sum_{2}^{n}h_{ij}y_iy_j$$ Since $H$ is symmetric, $h_{j1}=h_{1j}$, and so we can write $$h_{11}\left(y_1+h_{11}^{-1}\sum_{2}^{n}h_{1j}y_j\right)^2-h_{11}^{-1}\left(\sum_{2}^{n}h_{1j}y_j\right)^2$$ Set $$y_1+h_{11}^{-1}\sum_{2}^{n}h_{1j}y_j=z_1\tag{12}$$ We can then write $$q(y)=h_{11}z_1^2+q_2(y)\tag{13}\label{13}$$ where $q_2$ depends only on $y_2,\dots,y_n$.

If all diagonal terms of $q$ are zero but there is some non-zero off-diagonal term, say $h_{12}=h_{21}\ne0$, then we introduce $y_1+y_2$ and $y_1-y_2$ as new variables, which produces a non-zero diagonal term. If all diagonal and off-diagonal terms are zero, then $q(y)\equiv0$ and there is nothing to prove.

We now apply induction on the number of variables $n$. Using (13) shows that if the quadratic function $q_2$ in $(n-1)$ variables can be written in form (11), then so can $q$ itself. Since $y_2,\dots,y_n$ are related by an invertible matrix to $z_2,\dots,z_n$, it follows from (12) that the full set $y$ is related to $z$ by an invertible matrix.

This proof seems to skip the step of showing that the matrix is actually invertible. As far as I can tell, this method gives an upper triangular matrix $L$. However it doesn't seem to show or give a reason why $L$ is immediately invertible. I think I'm able to follow the reasoning up to (and including) the induction, but it's the last few sentences about invertibility that I'm confused about.

1

There are 1 best solutions below

0
On

The answer is quite simple. Since $y_i$ and $z_i$ are "coordinates", they can be treated as constant. Since $q_2$ does not include any terms with $y_1$, this mapping is injective and so the matrix representing this mapping is invertible.