Let $M,N$ be $n \times n$ square matrices over an algebraically closed field with the properties that the trace of both matrices coincides along with all powers of the matrix. More specifically, suppose that $\mathrm{Tr}(M^k) = \mathrm{Tr}(N^k)$ for all $1\leq k \leq n$. The following questions about eigenvalues is then natural and I was thinking it would be an application of Cayley-Hamilton but I am having trouble writing out a proof.

How do we show that $M$ and $N$ have the same eigenvalues?

Added (because this question is now target of many duplicates, it should state its hypotheses properly). Assume that all the mentioned values of $k$ are nonzero in the field considered; in other words either the field is of characteristic $0$, or else its prime characteristic $p$ satisfies $p>n$.

  • $\begingroup$ I am suspicious about this result... because I realize that it means it is also true for diagonal matrices with integer entries and thus gives a quite astonishing result for number theory. $\endgroup$ Dec 5 '11 at 2:03
  • 8
    $\begingroup$ The result is false over arbitrary fields: if the field has characteristic $p\gt 0$, then take $M$ to be the $p\times p$ zero matrix, and $N$ to be the $p\times p$ identity matrix. So presumably, you are working over $\mathbb{R}$ or $\mathbb{C}$ or some other restriction? $\endgroup$ Dec 5 '11 at 2:03
  • $\begingroup$ Yes sorry I forgot to give conditions on the base field: we can assume the matrices are over an algebraically close field. $\endgroup$
    – user7980
    Dec 5 '11 at 2:32
  • 3
    $\begingroup$ @user7980: The examples I give still hold over an algebraically closed field of positive characteristic. You need to assume more than that: as Splice note, you need either $n$ to be smaller than the characteristic, or for the characteristic to be zero. $\endgroup$ Dec 5 '11 at 2:58
  • 4
    $\begingroup$ @PatrickDaSilva If you work over integers/reals and all eigenvalues are real, there is a very simple argument to prove this result: If $|a_1| \leq |a_2| \leq ... \leq |a_n|$ then $\lim_k \frac{\sum_{i=1}^n a_i^{2k}}{a_n^{2k}}=\lim_k \frac{\sum_{i=1}^n b_i^{2k}}{a_n^{2k}}$. Left side counts the multiplicity of $a_n^2$ as an eigenvalue for $A^2$ while right side is finite and nonzero if and only $b_n^2=a_n^2$. Then you get the multiplicity of $b_2^2$ is the same... You eliminate these and repeat... Then it is easy to take care of signs.... $\endgroup$
    – N. S.
    Dec 5 '11 at 7:16

$M$ and $N$ have the same generalized eigenvalues (with multiplicity) if and only if their characteristic polynomials are the same. Thus it suffices to show that the power sums $T_k = \sum _{i=1}^{n} \alpha^k_i$ for $k = 1$ to $n$ generate the ring of symmetric polynomials (the coefficients of the characteristic polynomial are symmetric polynomials in the eigenvalues). This is a result due to Newton. For example, if $S_k$ is the usual $k$-th elementary symmetric polynomial (the sum of all products of $k$ distinct $\alpha_i$), then one has:

$$S_1 = \sum \alpha_i = T_1,$$

$$S_2 = \sum_{i > j} \alpha_i \alpha_j = \frac{1}{2} \left( \left(\sum \alpha_i\right)^2 - \sum \alpha^2_i \right) = \frac{1}{2}(T^2_1 - T_2)$$

More generally, one has:

$$\log \sum_{k=0}^{n} S_k X^k = \log \prod_{i=1}^{n} (1 + \alpha_i X) = \sum_{k=1}^{n} \log(1 + \alpha_i x)$$

which, expanding the logarithm, becomes:

$$ \sum_{k=1}^{n} \sum_{j=1}^{\infty} \frac{\alpha^j_i (-1)^{j-1} X^j}{j} = \sum_{j=1}^{\infty} \frac{T_j (-1)^{j-1} X^j}{j}$$

In particular, from the $S_k$ one can determine all the $T_k$, and from the $T_k$ (for $k = 1$ to $n$) one can determine the $S_k$ (and hence all the $T_k$ as well).

This even shows that the ring generated by $T_k$ over any ring $R$ for $k = 1$ to $n$ is the same as the ring generated by $S_k$ for $k = 1$ to $n$, as long as $n!$ is invertible in $R$. So the result also holds for any field of characteristic $p > n$. It is false if $p \le n$ - for example, the identity $p \times p$ matrix and all its powers has trace $0$, which is the same as the zero matrix.


What about this one?

We know that if $\psi_A(x) = |xI_n - A| = x^n+c_{n-1}x^{n-1}+c_{n-2}x^{n-2}+\cdots+c_1x+c_0I_n$ be characteristic polynomial of $A$ (thus $c_0 = (-1)^n|A|$), then the coefficients are given by $$c_{n-m}=\frac{(-1)^m}{m}\left| \begin{array}{ccccc} t_1 & m-1 & 0 & \cdots & 0 \\ t_2 & t_1 & m-2 & \cdots & 0 \\ \vdots & \vdots & & & \vdots \\ t_{m-1} & t_{m-2} & t_2 & t_1 & 1 \\ t_m & t_{m-1} & t_3 & t_2 & t_1 \\ \end{array} \right|$$ where $t_r:= \operatorname{tr}(A^r)$.

Assume that $\psi_{M}(x)=x^n+a_{n-1}x^{n-1}+a_{n-2}x^{n-2}+\cdots+a_1x+a_0I_n$ and $\psi_N(x)=x^n+b_{n-1}x^{n-1}+b_{n-2}x^{n-2}+\cdots+b_1x+b_0I_n$. Then $$a_r=b_r$$ (by using $\operatorname{tr}(M^k) = \operatorname{tr}(N^k)$ and the above determinant).

Hence $\psi_M(x)=\psi_N(x)$ which means $M$ and $N$ will have same eigenvalues.


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.