# Proof that an $n \times n$ real symmetric matrix has $n$ real eigenvalues [closed]

How to show that an $n \times n$ real symmetric matrix has $n$ real eigenvalues $$\lambda_n \geqslant \dots \geqslant \lambda_1$$ with corresponding eigenvectors $\mathbf v_n, \dots,\mathbf v_1$ such that $$\mathbf v_i^T \mathbf v_j = \delta_{ij}.$$ I am struggling a bit on how to start this proof. I tried looking online but could not find anything. Perhaps I am looking in the wrong direction, would it be the same to prove that such a matrix would have $n$ real orthogonal eigenvectors and thus $n$ real eigenvalues? Hints on where to start this proof would be appreciated.

## closed as too broad by Cameron Williams, user251257, graydad, Tom-Tom, user91500Oct 2 '15 at 9:44

Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. Avoid asking multiple distinct questions at once. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

• Look up the spectral theorem. The proof is much too long and easily found elsewhere to ask someone to do so on here. I have voted to close as too broad. – Cameron Williams Oct 2 '15 at 1:19
• If I recall correctly, the proof is by induction on $n$. You can find it in Linear Algebra, by Hoffman and Kunze (the book is available online, if you are so inclined). Although I think the book uses the name Spectral Theorem for something else. – Reveillark Oct 2 '15 at 1:21
• The gist of the proof: first prove that there is one real eigenvalue (for instance, apply the extreme value theorem to the Rayleigh quotient and check that the maximizer is an eigenvector). Then show that the action on the orthogonal complement of this eigenspace is given by a (lower dimensional) symmetric matrix. This latter step makes induction on $n$ possible. – Ian Oct 2 '15 at 1:28

## 2 Answers

This is actually fairly straightforward. Let the symmetric matrix be $A$. Let's start with the classic result:

Let $\lambda$ be an eigenvalue of $A$ with eigenvector $v$. Then $\lambda \in \mathbb{R}$.

Proof: $v^{\dagger} v$ is real for any complex vector. By definition, $$\lambda v^{\dagger}v = v^{\dagger} A v.$$ Taking conjugate transpose on both sides, $$\bar{\lambda} v^{\dagger} v = (v^{\dagger} Av)^{\dagger} = v^{\dagger} A^{\dagger} v = v^{\dagger} A v = \lambda v^{\dagger} v.$$ Since $v \neq 0$, $\lambda=\bar{\lambda}$ and so $\lambda$ is real. $\square$

(And further, therefore the eigenvectors are all real since $A$ is real, and we can go back to normal transposes.)

Now, having done that, we need to show that $A$ has an eigenvalue. There are two ways of doing this: firstly, we can just write down the characteristic polynomial $\det{(\lambda I-A)}$, and find its roots, which consist of the eigenvalues. Alternatively, consider the function $$R(x) = \frac{x^T A x }{x^Tx},$$ often called the Rayleigh quotient. One can prove that the minimiser of this quotient is an eigenvector. (It turns out to be equivalent to minimising $x^T A x$ subject to $x^T x=1$, and then you can use Lagrange multipliers and derive that a minimiser must be an eigenvector. Further, it's a continuous function on a compact set and hence it attains its minimum.) Indeed, the minimum produces the smallest eigenvector (this is easy to check: suppose not, &c.)

Now we need the other classic result:

Eigenvectors with different eigenvalues are orthogonal.

Proof: Let $\lambda,v$ and $\mu,u$ be eigenvalues and their eigenvectors. Then $$\lambda u^Tv = u^T A v = (v^T A^T u)^T = (v^T Au)^T = \mu (v^T u)^T = u^T v,$$ so if $\lambda \neq \mu$, $u^Tv=0$. $\square$

Now we can derive an algorithm to produce orthonormal eigenvectors. Go back to the minimisation problem, $x^T A x$ on the sphere $x^Tx=1$. We have already noticed that the minimum gives the smallest eigenvalue $\lambda_1$ and an eigenvector $v_1$ corresponding to it. Now add in an extra condition: $v_1^T x=0$. This is a closed set because $x \mapsto v_1^Tx$ is continuous. Hence the minimisation problem $$x^T A x, \quad \text{subject to } x^Tx=1, \quad v_1^Tx=0$$ has a minimiser, which you can check gives you another eigenvector $v_2$ and eigenvalue $\lambda_2 \geqslant \lambda_1$. It should be fairly obvious what's going to happen now: add in $v_2^T x=0$ to the conditions on $x$, solve the minimisation problem ...

This algorithm terminates when the set of admissable vectors becomes empty, which can only occur when we have found $n$ vectors for $x$ to be orthogonal to: the $n$ eigenvectors we have found form an orthonormal basis as required.

(Alternatively, you can go about diagonalising $A$ one eigenvalue at a time, but it's quite tedious: you have to write it as a direct sum of eigenspaces, and it looks more complicated than using the Rayleigh quotient, but it generalises to other fields more easily.)

Since $|\lambda I-A|$ is of degree $n$, it has $n$ roots and so $A$ has $n$ eigenvalues. Let $\lambda_i$ be any eigenvalue of $A$ and $x\ne 0$ be eigenvector for $\lambda_i$. Then $Ax=\lambda_i x$. So $\overline{x}^TAx=\lambda_i \overline{x}^Tx$. Take conjugate transpose on it, we have $$\overline{x}^T\overline{A}^Tx=\overline{x}^TAx=\overline{\lambda_i}\overline{x}^Tx=\lambda_i \overline{x}^Tx$$ So $\overline{\lambda_i}={\lambda_i}$ for $\overline{x}^Tx>0$. Thus $\lambda_i$ is real.

Since all eigenvalue of $A$ are real, so are eigenvectors of $A$. Let $x_1$ be a real normalized eigenvector for $\lambda_1$, i.e. $Ax_1=\lambda_1 x_1$. Then $x_1$ can be expanded to orthonormal base of $x_1, \cdots, x_n$. Let $S=(x_1, \cdots, x_n)$. Then $S^{-1}=S^T$. Note that $$AS=A(x_1,x_2,\cdots, x_n)=S\pmatrix{\lambda_1 & *\\ 0& B_1}=SB$$ So $S^{-1}AS=S^TAS=B$. Moreover, since $A$ is symmetric, $B^T=B$, i.e. $$B=\pmatrix{\lambda_1 & 0\\ 0& B_1}$$ where $B_1$ is $(n-1)\times (n-1)$ and symmetric. By induction hypothesis, there is an orthogonal matrix $S_1$ that $B_1S_1=S_1\Lambda_1$, where $\Lambda_1$ is diagonal. Hence $$\pmatrix{\lambda_1 & 0\\ 0& B_1}\pmatrix{1 & 0\\ 0& S_1}=\pmatrix{1 & 0\\ 0& S_1}\pmatrix{\lambda_1 & 0\\ 0& \Lambda_1}$$ Let $$B=\pmatrix{\lambda_1 & 0\\ 0& B_1}\quad\text{}\quad Q=\pmatrix{1 & 0\\ 0& S_1}\quad\text{}\quad \Lambda=\pmatrix{1 & 0\\ 0& \Lambda_1}$$ Then $BQ=Q\Lambda$. Let $P=SQ$. Then $P$ is orthogonal. Thus we have $$P^TAP=\Lambda$$ Clearly $\Lambda=Diag{(\lambda_1, \cdots, \lambda_n)}$ and each column of $P$ is eigenvector of $A$.