algebraic equation with trace

76 Views Asked by At

I have a problem which I don't know how to attack. Actually I am not even sure there is a way to do it.

Is it possible to solve an equation of this form $$A²-\frac{tr(A)²}{4}Id_{4\times 4}=B$$ where A, B are (4*4) matrices, A is the unknown. Is it possible to get A as function of B.

1

There are 1 best solutions below

0
On

I shall assume that $B$ is diagonalizable, because this gets messy if it's not (plus, the approach would be too numerically unstable).

Notice that $X$ and $X - \xi {\rm I}$ are similar, with $\lambda_i(X) = \lambda_i(X - \xi {\rm I}) + \xi$. Denoting a Jordan decomposition of $B$ as $B = S J S^{-1}$ and $\xi := (\operatorname{tr} A)^2 / 4$, we have

$$J = S^{-1} B S = S^{-1} (A^2 - \xi {\rm I}) S = S^{-1} A^2 S - \xi {\rm I},$$

i.e.,

$$S^{-1} A^2 S = J + \xi {\rm I},$$

which means that $A^2 = S (J + \xi {\rm I}) S^{-1}$, i.e., $A$ has the same similarity matrix in its Jordan normal form (diagonalization) as $B$. (This part wouldn't be true if $A$ could have been nilpotent)

Since we have assumed that $B$ is diagonalizable, $J$ is a diagonal matrix, and so is $J + \xi {\rm I}$. Also, if $A = S K S^{-1}$, then $A^2 = S K^2 S^{-1}$, $K$ is diagonal, and $\operatorname{tr} A = \operatorname{tr} K$. This means that we're solving a diagonal system

$$4J = 4K^2 - (\operatorname{tr} K)^2,$$

or, if we write it elementwise, with $J = \operatorname{diag}(j_1,\dots,j_n)$, $K = \operatorname{diag}(k_1,\dots,k_n)$,

$$4j_i = 4k_i^2 - \left( \sum_{i=1}^n k_i \right)^2.$$

That's a system of $n$ quadratic equations with $n$ unknowns (in your case, $n = 4$), solvable via numerical methods.

Once you solve that for $K$, you have $A = S K S^{-1}$.

Note: Jordan decomposition is not a numerically stable one, so be careful here. Given that your matrices are small, this might work, but do take some care here. If $B$ is real symmetric or complex Hermitian, you can use eigendecomposition, which is stable and should produce fine results.