MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let $(e_j)_{j\in\mathbb N}$ be an orthonormal basis of a Hilbert space $V$. Let $T:V\to V$ be continuous, selfadjoint linear operator. Assume that for all $i,j\in\mathbb N$ the number $\langle Te_i,e_j\rangle$ is rational. Let $\lambda\in\mathbb C$. Suppose there exists an eigenvector $v\in V$ with $Tv=\lambda v$. The question is this: Does there exist an eigenvector $w$ for the eigenvalue $\lambda$ such that $\langle w,e_j\rangle$ lies in the field ${\mathbb Q}(\lambda)$ for every $j\in\mathbb N$?

This is true for finite dimensional spaces, even without the self-adjointness, since $\lambda$ is a zero of the characteristic polynomial.

If it turns out to be false as stated, you may add further conditions on $T$, like the one that it preserves the linear span of the basis vectors.

share|cite|improve this question
For finite dimensional spaces this is true regardless the fact that $\lambda$ is a zero of the characteristic polynomial: indeed the hypothesis on $T$ implies that $T-\lambda Id:\mathbb{C}^n\to \mathbb{C}^n$ is induced by extension of scalars by an operator $T-\lambda Id:\mathbb{Q}(\lambda)^n\to \mathbb{Q}(\lambda)^n$, and so its kernel will have a basis consisting of vectors in $\mathbb{Q}(\lambda)^n$. The fact that $\lambda$ is a zero of the characteristic polynomial of $T$ gives the further information that $\lambda$ is algebraic over $\mathbb{Q}$. – domenico fiorenza Dec 20 '10 at 10:45
up vote 15 down vote accepted

The answer is no.

Update. Here is an explicit example (which also preserves the $\mathbb Q$-span of the basis). Let $V=\ell_2$ over $\mathbb R$ and $(e_j)$ the standard basis. Define $T:V\to V$ by $$ \begin{aligned} (Tx)_1 &= x_1+3x_2+x_3 \\ (Tx)_2 &= 3x_1+10x_2+6x_3+x_4 \\ (Tx)_n &= x_{n-2}+6x_{n-1}+11x_n+6x_{n+1}+x_{n+2}, \qquad n>2 . \end{aligned} $$ Let $\alpha$ be the root of $p(x)=x^2+3x+1$ such that $|\alpha|<1$. Then the vector $v\in\ell_2$ defined by $v_n=\alpha^n$ satisfies $Tv=0$, so it is an eigenvector for $\lambda=0$. On the other hand, no nonzero rational vector $w$ satisfies $Tw=0$ because the relation $Tw=0$ implies the recurrence relation $$ w_{n+4}=-6w_{n+3}-11w_{n+2}-6w_{n+1}-w_n , $$ which does not have rational solutions tending to 0.

Below is the original answer.

Let $V=\ell_2$ over $\mathbb R$ and $(e_j)$ the standard basis, so that $\langle x,e_j\rangle=x_j$ for every $x\in\ell_2$. Let $v\in V$ be a unit vector with irrational ratios of coordinates (for definiteness, take a geometric progression like $(c\pi^{-n})_{n\in\mathbb N}$). Then there exist a continuous self-adjoint $T:V\to V$ with rational coordinates such that $\ker T=\langle v\rangle$. So we have $\lambda=0$ but no $\lambda$-eigenvector is rational.

I represent $T$ by its (infinite) matrix $(t_{ij})_{i,j\in\mathbb N}$, $t_{ij}=\langle Te_i,e_j\rangle$. This matrix should be symmetric, and continuity of the resulting operator should be taken care of.

Begin with a self-adjoint $A$ such that $\ker A=\langle v\rangle$ with irrational components $a_{ij}$. For example, let $A$ be the projection to the orthogonal complement of $v$, so $a_{ij}=\delta_{ij}-v_iv_j$. I am going to approximate the matrix $(a_{ij})$ by a rational matrix $(t_{ij})$ such that the kernel stays the same.

Since $Av=0$, each row of our matrix is orthogonal to $v$. Approximate the first row $a_i=(a_{1i})$ by a rational vector $t_1=(t_{1i})$ such that $\langle t_1,v\rangle=0$ and $|t_{1i}-a_{1i}|$ is bounded by a rapidly decaying geometric progression (see below for details). Replace the first row and the first column of the matrix by this approximation. Then adjust the diagonal elements $a_{ii}$ so that the rows remain orthogonal to $v$, namely change $a_{ii}$ to $a_{ii}-(a_{1i}-t_{1i})v_1v_i^{-1}$. Recall that $v_i$ is a not-so-fast decaying geometric progression, so this adjustment is a small change.

Now remove the first row and the first column from the matrix, and the first component from $v$. Apply the same procedure to the truncated data: namely approximate the first remaining row and the first remaining column by a rational vector whose scalar product with $v$ is the same as before, then adjust the diagonal elements so that the scalar products of all rows with $v$ are preserved. Repeat ad infinitum. Note that every element of the original matrix is changed only finitely many times, and $v$ belongs to the kernel of the matrix after each step.

The approximations are controlled and the adjustments are bounded in terms of approximations, so the sum $\sum (a_{ij}-t_{ij})^2$ can be made arbitrarily small. This implies that $T$ is continuous and $\|T-A\|$ is small (say, less than 1/2), so we have $Tv=0$ and $\|Tw\|\ge\frac12\|w\|$ for all $w$ orthogonal to $w$. Therefore $\ker T=\langle v\rangle$. However $v$ cannot be rescaled to a rational vector.

It remains to show that every vector $w\in\ell_2$ can be approximated by a rational vector $w'$ such that $\langle w',v\rangle=\langle w,v\rangle$ and $w'_i-w_i$ is bounded by a small, fast decaying progression. Let $w'_1=w_1+\varepsilon_1$ be a rational approximation of $w_1$, then let $w'_2=w_2-\varepsilon_1v_1v_2^{-1}+\varepsilon_2$ be a rational approximation of $w_2-\varepsilon_1v_1v_2^{-1}$, then let $w'_3=w_3-\varepsilon_2v_2v_3^{-1}+\varepsilon_3$ be a rational approximation of $w_3-\varepsilon_2v_2v_3^{-1}$, and so on. The $\varepsilon_i$ at each step can be chosen arbitrarily small, and the resulting vector $w'$ satisfies $\langle w'-w,v\rangle=0$.

share|cite|improve this answer
Great. That takes care of the general case. But what if we assume that $T$ preserves the $\mathbb Q$-span of the basis $(e_j)$? – Anton Dec 21 '10 at 11:23
@Anton: see update. – Sergei Ivanov Dec 21 '10 at 16:17

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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