Given the matrix: $$A=\begin{pmatrix}i&1\\1&-i\end{pmatrix}$$ Find a matrix $P$ such that $P^T A P$ is diagonal, how should I go about this? we know from Sylvester's theorem that $A$ is congruent to the matrix $D=\begin{pmatrix}1&0\\0&0\end{pmatrix}$ since it has rank $1$ and we're considering it over the complex field $\mathbb{C}$, however i'm not sure how can I go about this, if $A$ was real, since it is symmetric, I could consider the identity matrix and do row/column operations until I got A into a diagonal form, mimic such operations on the identity matrix and that should do the job, but in this case it doesn't work (unless I made some calculation mistake...)

Is there a general way to find the matrix which gives the congruence between a symmetric matrix and a diagonal matrix? thanks in advance...


3 Answers 3


0) outline of general approach (may be skipped)
a very flexible way of dealing with problems involving congruence transforms is to consider your matrix as representing a symmetric bilinear form

i.e. abstractly having some vector space $V$ where $\mathbf v, \mathbf v'\in V$ may be related by

$\langle \mathbf v,\mathbf v'\rangle = \langle \mathbf v',\mathbf v\rangle = c \in \mathbb C$

where $\langle ,\rangle$ denotes some particular symmetric bilinear form, not an inner product per se

After introducing some basis we have $P\mathbf x = \mathbf v$ and $P\mathbf y = \mathbf v'$ and

The coordinate interpretation is
$\langle \mathbf v,\mathbf v'\rangle = \mathbf x^T P^TA P\mathbf y$

OP's problem amounts to selecting(/changing) the basis wisely so that $ P^TA P = D$ for some rank $k$ (k=1 in this problem) diagonal matrix $D$, preferably with all non-zero components on the unit circle.

In general the process consists of a modified Gram Schmidt-- here we can first figure out the dimension of the space of null vectors = r, then the subspace of non-null vectors $W$ has dimension $k=n-r$; in this particular problem $r=1=k$. A null-vector is a vector with that is orthogonal to every vector; orthogonal is defined as $\langle \mathbf v, \mathbf v'\rangle = 0$ (again this is not an inner product).

Working over $\mathbb C$ (in fact any field of characteristic $\neq 2$) we can easily find vectors that are not self-orthogonal and use this to run Gram Schmidt -- and in particular the normalization stage won't fail since we have not self-orthogonal vectors and in $\mathbb C$ we can always find square roots to normalize the 'length' with respect to the bilinear form to be 1. The computational test at this stage amounts to matrix vector multiplication with $\mathbf w\in W$ having coordinate vector $\mathbf z_1$ and collecting $n-r$ linearly independent vectors from $W$ in coordinate form in a matrix $Z$ and then computing $Z^T A\mathbf z$.

Artin's Algebra's chapter on Bilinear Forms has the details for the general approach for symmetric (and Hermitian) and skew bilinear forms.

1) easy answer for OP's specific problem
whenever $\text{rank}\big(A\big)=1$ for some symmetric $A$, focus building a basis for the kernel of $A$. For this particular problem:

$\mathbf p_2: = \begin{bmatrix}-1 \\ i \end{bmatrix}$
$A\mathbf p_2 =\mathbf 0$

select $\mathbf p_1$ to be linearly independent of $\mathbf p_2$ (e.g. a standard basis vector will do) and

$P := \bigg[\begin{array}{c|c} \mathbf p_1 & \mathbf p_2 \end{array}\bigg]$

$P^TAP = P^T(AP)= \bigg[\begin{array}{c|c} P^T(A\mathbf p_1) & P^T\mathbf 0 \end{array}\bigg]=\begin{bmatrix}\eta &0 \\ *&0 \end{bmatrix} =\begin{bmatrix}\eta &0 \\ 0&0 \end{bmatrix}$
for some $\eta \neq 0$

1.) we know $*=0$ by symmetry, that is $\big(P^TAP\big)^T = \big(P^TA^TP\big) = \big(P^TAP\big)$
2.) we also know $\eta \neq 0$ because $\text{rank}\big(P^TAP\big)=\text{rank}\big(A\big)=1$

from here we may effect one more congruence transform, this time using an elementary type 3 matrix so as to map $\eta \mapsto 1$.


Consider this: $\mathrm{det}(P^TAP)=0$, since $\mathrm{det}A=0$. Thus one of diagonal elements of $P^TAP$ must be zero.

Let $P=\begin{bmatrix}p_1& p_2\\ p_3 &p_4\end{bmatrix} $ then $P^TAP=\begin{bmatrix}2p_1p_3+i(p_1^2-p_3^2)& p_1p_4+p_2p_3+i(p_1p_2+p_3p_4)\\ p_1p_4+p_2p_3+i(p_1p_2+p_3p_4) &2p_2p_4+i(p_2^2-p_4^2)\end{bmatrix}$.

Then to control your only diagonal element (lets say $d$) you can set off-diagonal elements to zero and $2p_2p_4+i(p_2^2-p_4^2)=0$ by choosing $p_2=p_4=0$ and $p_1=p_3=\sqrt{\frac{d}{2}}$.


yes, take one elementary matrix at a time on the right; if there are $r$ steps needed to finish the job, the final is $P = P_1 P_2 \ldots P_r.$ This time, just one is needed, $$ \left( \begin{array}{rr} 1 & 0 \\ i & 1 \\ \end{array} \right) \left( \begin{array}{rr} i & 1 \\ 1 & -i \\ \end{array} \right) \left( \begin{array}{rr} 1 & i \\ 0 & 1 \\ \end{array} \right) = \left( \begin{array}{rr} i & 0 \\ 0 & 0 \\ \end{array} \right) $$ There was no need to use anything other than the Gaussian integers here. For larger size matrices, it would still suffice to use a string of elementary matrices with elements in $\mathbb Q [i].$ The algorithmic description of this is at reference for linear algebra books that teach reverse Hermite method for symmetric matrices

If you really want $1$ as the $1,1$ element, you can multiply on both sides by a diagonal matrix, with 1,1 element being a chosen $\sqrt{1/i}.$ So, $\frac{1-i}{\sqrt 2}$ works.

If we ask about eigenvalues first, we find that the characteristic polynomial and the minimal polynomial are $\lambda^2,$ so that the Jordan form is not diagonal: $$ \left( \begin{array}{rr} 1 & 0 \\ i & 1 \\ \end{array} \right) \left( \begin{array}{rr} i & 1 \\ 1 & -i \\ \end{array} \right) \left( \begin{array}{rr} 1 & 0 \\ -i & 1 \\ \end{array} \right) = \left( \begin{array}{rr} 0 & 1 \\ 0 & 0 \\ \end{array} \right) $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.