7

This question is linked to this one: nilpotent elements of $M_2(\mathbb{R})$, $M_2(\mathbb{Z}/4\mathbb{Z})$

Let $R$ be a commutative ring with unity and let $A\in M_2(R)$. Show that $A$ is a nilpotent matrix iff $\det(A)$ and $\mathrm{trace}(A)$ are nilpotent elements of $R$.

  • I edited my answer. There is a simple solution for matrices of order 2. It might be true that $A$ is nilpotent iff the coefficients of the characteristic polynomial of $A$ are nilpotent (except the first coefficient 1). If the coefficients are nilpotent then A is nilpotent follows exactly like the proof for order 2. I don't know how to do the other direction. – Daniel Jan 17 '15 at 23:02
  • I added an observation. I believe this is true. – Daniel Jan 17 '15 at 23:32

3 Answers3

5

One direction can be found here: Trace of nilpotent matrix over a ring

For the other direction write $A^2=tr(A)A-\det(A)I$ and notice that $tr(A)A$ and $\det(A)I$ are nilpotent and commute.

Another Solution. Claim: Let $A\in M_2(R)$. If $tr(A^{2^k})$ is nilpotent then $tr(A^{2^{k-1}})$ is nilpotent.

By Cayley-Hamilton, $A^{2^k}=(A^{2^{k-1}})^2=tr(A^{2^{k-1}})A^{2^{k-1}}-\det(A^{2^{k-1}})I$.

Thus, $tr(A^{2^k})=tr(A^{2^{k-1}})^2-2\det(A^{2^{k-1}})$. Thus, $tr(A^{2^{k-1}})^2=tr(A^{2^k})+2\det(A^{2^{k-1}})$. Since $tr(A^{2^k})$ is nilpotent, by hypothesis of this claim, and $\det(A)$ is nilpotent then $tr(A^{2^{k-1}})^2$ is nilpotent and $tr(A^{2^{k-1}})$ is too.

Since $A$ is nilpotent exists $k$ such that $tr(A^{2^k})=0$. Now, use the claim $k$ times to obtain $tr(A)$ nilpotent.


Observation: In this answer , it is proved that $tr(A)$ is nilpotent if $A$ is nilpotent. Using the exactly same idea we can prove that the others coefficients of the characteristic polynomial of A (except the first) are also nilpotent.

Now if the coefficients of the characteristic polynomial of $A$ are nilpotent then $A^n$ is a sum of nilpotent matrices that commute: $A^n=\sum_{i=0}^{n-1}a_iA^i$, where $a_i$ are the coefficients of the characteristic polynomial of $A$ (Cayley-Hamilton). Thus, $A^n$ is nilpotent and $A$ is nilpotent.

Thus $A$ is nilpotent iff the coefficients of the characteristic polynomial of $A$ are nilpotent.

Daniel
  • 5,872
5

I write quickly an answer because this file will be put "on hold". Yes, the ayatollah act even in the mathematical field.

Daniel did a very good job that implies the following equivalence:

Assume that $n!$ is a unit in $R$ and $A\in M_n(R)$. Then $A$ is a nilpotent matrix iff for every $1\leq i\leq n$, $\mathrm{trace}(A^i)$ is a nilpotent element of $R$.

Proof: According to the Qiaochu Yuan'post in Formulas for the (top) coefficients of the characteristic polynomial of a matrix or my post (EDIT 2,3) in nth derivative of determinant wrt matrix the coefficient $a_{n-i}$ of Daniel is $a_{n-i}=\dfrac{1}{i!}(P_{n-i}(trace(A),\cdots, trace(A^{n-i-1}))\pm (n-i-1)trace(A^{n-i}))$ where $P_{n-i}$ is a polynomial with coefficient in $\mathbb{Z}$. According to Daniel's result, the $(a_{n-i})$ are nilpotent ; by a recurrence reasoning, we can conclude that the $(trace(A^k))$ are also nilpotent ; the converse is easy.

Remark: the condition "$n!$ is a unit is essential". Indeed, consider the case $R=\mathbb{Z}/4\mathbb{Z}$ and $n=2$. $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$ is nilpotent iff $2|a+d$ and $2|ad-bc$. Now the condition about the traces can be written $2|a+d$ and $2|a^2+d^2+2bc$, that is $2|a^2+d^2$. The entries $b,c$ are free and, consequently, the equivalence does not work.

EDIT. About the orangeskid's answer. Consider the following instance $p(X)=X^3+X+4,R=\mathbb{Z}/15\mathbb{Z}$. Since $p$ admits no roots in $R$, we consider the ring $R_1=R[X]/(p(X))=\{ax^2+bx+c;a,b,c\in R\}$ with the condition $x^3=-x-4$ ; over $R_1$, $p$ admits $3^2=9$ roots ($3$ mod $3$ and $3$ mod $5$):

$x,6x+10,11x+5,6x^2+6x+4,6x^2+x+9,6x^2+11x+14,9x^2+3x+1,9x^2+13x+6,9x^2+8x+11.$

Yet, amongst these roots, we cannot find $\lambda,\mu,\nu$ s.t. $X^3+X+4=(X-\lambda)(X-\mu)(X-\nu)$. Then, as orangeskid wrote, we must consider an additional quotient in order to factorize p into a product of first degree factors.

1

The traces are more interesting. In dimension $d=2$ we have the implication

$$A^n = 0 \implies (\text{Tr} A)^{2n-1}=0$$ and $2n-1$ is the best exponent.

$\bf{Added}$ Here is a general proof that if $A$ is nilpotent then all the coefficients of the characteristic polynomial of $A$ are nilpotent.

We will use two basic facts from algebra.

  1. If $P(X)$ is a monic polynomial in $R[X]$ there exists an extension of rings $R\subset S$ in which the polynomial decomposes

$$P(X) = (X-\lambda_1)\cdot \ldots \cdot (X-\lambda_d)$$

  1. Let $A$ a matrix in $M_d(R)$ so that its characteristic polynomial $P_A(X)=\det(X\cdot I_d - A)$ decomposes in the extension $R\subset S$ as

    $$P_A(X) = (X-\lambda_1)\cdot \ldots \cdot (X-\lambda_d)$$

    Then we have for the characteristic polynomial of $A^n$

$$P_{A^n}(X) = (X-\lambda_1^n)\cdot \ldots \cdot (X-\lambda_d^n)$$

Assume now that $A^n=0$. Then $P_{A^n}(X)=X^d$, that is $$X^d= (X-\lambda_1^n)\cdot \ldots \cdot (X-\lambda_d^n)$$

Plug into this equality $X \mapsto \lambda_i^n$. We get $(\lambda_i^n)^d=0$, that is $\lambda_i^{nd}=0$. From here we conclude that all the fundamental symmetric polynomial in $\lambda_i$ are nilpontent, so the coefficients of the characteristic polynomial of $A$ (other than the leading $1$) are nilpotent, done.

Note that we only used the fact that the coefficients of the characteristic polynomial of $A^n$ are zero. This is weaker than $A^n=0$.

Obs: Using this method we can show that $A^2=0$ implies $(\text{Tr}A)^4$ for $A$ in $M_2(R)$. However, it's true that already implies $(\text{Tr}A)^3=0$. So this method will not give the optimal exponents. Nevertheless, we can have some explicit estimates.

$\bf{Added:}$ Here is an example: Let $A$ a $2\times 2$ matrix in $M_2(S)$ ($S$ commutative ring) and $\lambda_1$, $\lambda_2$ in $S$ so that $$\text{Tr} A = \lambda_1 + \lambda_2 \\ \det A = \lambda_1 \cdot \lambda_2$$ Then we have $$\text{Tr} A^2 = \lambda_1^2 + \lambda_2^2 \\ \det A = \lambda_1^2 \cdot \lambda_2^2 $$

We need to show that $\text{Tr} A^2 = (\text{Tr} A)^2 - 2 \det A$. This could be done directly, or we can use the known identity $A^2 - \text{Tr} A \cdot A + \det A \cdot I_2 = 0$.

orangeskid
  • 53,909
  • @ orangeskid , I'm stumped. Have you a reference for your property 1 ? Moreover a polynomial of degree $d$ over a ring may have more than $d$ roots. I don't know.... –  Jan 18 '15 at 15:33
  • 1
    @loupblanc: I've seen it somewhere as an exercise and the proof works as follows: show that for every ring $R$ and polynomial $P$ monic there exists an extension of $R$ in which $R$ has a root. You can take $R[X]/(P)$. The map $R \to R[X]/(P)$ is injective, for instance because the polynomial is monic ( in fact, it would work for polynomials for which the dominant coefficient is not a zero divisor..) Anyways, so we get an extension where $P$ has a root $\alpha$. Make a division $P(X) = (X-\alpha)Q(X)$ and continue. This is analogous to the field situation, ours do not need to be fields. – orangeskid Jan 18 '15 at 15:43
  • 1
    @loupblanc: Also, it is not important that the roots are distinct or not,or if there are any other roots; only that the Viete relation are valid. This is the key. The magic stuff is that if $\lambda_i$ satisfy the Viete for $P_A(X)$ then $\lambda_i^n$ do for $P_{A^n}(X)$. This is the meaning of: the eigenvalues of $A^n$ are (eigenvalues of $A$)$\ ^n$ – orangeskid Jan 18 '15 at 15:45
  • @loupblanc: Note that I need $R\to S$ to be injective, because the nilpotency will be proved in $S$. – orangeskid Jan 18 '15 at 15:48
  • @loupblanc: The fact : eigenvalues of $A^n$ are (eigenvalues of $A$)$\ ^n$ is equivalent to $d$ ($A\in M_d(\cdot)$ ) identities in the entries of $A$. It involves symmetric functions in some way. – orangeskid Jan 18 '15 at 15:55
  • @loupblanc: If you want, I can add some more detail to the answer. – orangeskid Jan 18 '15 at 16:06
  • @ orangeskid , I never saw such an argument and I am not sure that it is correct ; yet, it is very interesting and this is the reason why I upvote your post ; it would imply that we can work with pseudo-eigenvalues of any matrix. In this connection, you say that the $({\lambda_i}^n)$ are pseudo-eigenvalues of $A^n$ ; can you give some details about this assertion ? –  Jan 19 '15 at 15:23
  • @loupblanc: This about the eigenvalues of $A^n$. Remember any exercises of this sort: you are given a polynomial, with roots named $\lambda_i$ ( the polynomial is given, the roots no). Now they ask you to produce a polynomial with roots $\lambda_i^2$. What do you do: you form the polynomial with roots $\lambda_i^2$. For this the Viete relations have to be satisfied. So you form the fundamental symmetric functions in $\lambda_i^2$. These, being symmetric functions in $\lambda_i$, can be expressed in terms of the fun symmetric functions of $\lambda_i$, so in terms of the coeffn of the polynomia – orangeskid Jan 20 '15 at 05:30
  • @loupblanc: So we see that the coeffs of monic the pol with roots $\lambda_i^2$ are polynomials ( very precise) in the coeffn of the monic pol with roots $\lambda_i$. Now let's consider a matrix $A$. You have the characteristic pol $P_A$. For matrices over say $\mathbb{C}$ you know that the char pol of $A^n$ $P_{A^n}$ has the roots the roots of $P_A$ raised to power $n$. What this means is : that procedure that from a poly produces a poly with roots the $n$-th power of the roots of the given one, takes $P_A$ to $P_{A^n}$. So you have a bunch of identities in the entries of $A$. – orangeskid Jan 20 '15 at 05:36
  • @loupblanc: Since these identities work for any elements in $\mathbb{C}$ they are just that, identities. So they will work for any ring ( commutative). What they say imply is: if $P_A(t)$ factors as $\prod (t - \lambda_i)$ then $P_{A^n}$ factors as $(\prod t - \lambda_i^n)$. – orangeskid Jan 20 '15 at 05:40
  • @ orangeskid , OK for $A^n$ ; it's a formal proof. Moreover, I am convinced by your entire post. cf. my EDIT. –  Jan 21 '15 at 18:02
  • @loupblanc: Neat, +1. – orangeskid Jan 22 '15 at 00:26
  • @loupblanc: For a commutative ring $R$ and a monic polynomial $P(X)=X^n + a_1 X^{n-1} + \ldots + a_n$ the map from $R$ to $R[ u_1, u_2, \ldots, u_n] /( \sum u_i = -a_1, \sum u_i u_j = a_2, \ldots, u_1 \ldots u_n =(-1)^n a_n)$ is an injection. You have the universal roots of the polynomial $P$ , $\bar u_1$, $\ldots \bar u_n$. – orangeskid Jan 22 '15 at 00:33
  • @loupblanc: I also want to point out two examples of decompositions: over $\mathbb{Z}/32$ we have $x^2 = (x-8)(x-24)$ and over $\mathbb{Z}[t]/(t^2)$ we have $x^2 = (x-t)(x+t)$ – orangeskid Jan 22 '15 at 00:48