9
$\begingroup$

I need to prove that the determinant $\det: M(n, \mathbb{R}) \rightarrow \mathbb{R}$ is a continuous function given the euclidean metric on the vector space of all $n \times n$ matrices over $\mathbb{R}$, i.e. $\Vert M \Vert = \sqrt{\sum_{i,j=1}^n M_{i,j}^2}$.

So what I need to prove, I think, is that there exists a $\delta > 0$ such that if $\Vert M - M' \Vert < \delta$, for any $M' \in M(n,\mathbb{R})$ then it follows that for all $\epsilon > 0: |\det(M) - \det(M')|< \epsilon$.

Unfortunately I have no idea how to derive the correct inequalities between the given euclidean metric and a determinant, since I'm kind of struggling with the permutation part in the definition of a determinant. Can anybody help, please?

$\endgroup$
  • 10
    $\begingroup$ Do you have to use $\epsilon-\delta$? For example, can you also use that polynomials are continuous? Multiplications of continuous functions are continuous? $\endgroup$ – user99914 Jun 6 '15 at 11:40
  • $\begingroup$ I'm not entirely sure but I think the proof is supposed to use $\epsilon - \delta$, since the question explicitly says that the metric on the vector space of matrices is the euclidean norm. Would a specific metric on the domain of det somehow be relevant to a proof involving polynomials? $\endgroup$ – eager2learn Jun 6 '15 at 11:46
  • $\begingroup$ You may think it this way: Identified $M(n, \mathbb R)$ as $\mathbb R^{n^2}$. Note that the metric on $M(n, \mathbb R)$ is exactly the usual Euclidean metric on $\mathbb R^{n^2}$. If you think of $\det$ as a function from $\mathbb R^{n^2}$, this function is really a polynomial of degree $n$ with $n^2$ variables. $\endgroup$ – user99914 Jun 6 '15 at 11:49
  • 3
    $\begingroup$ Very interesting and apparently horrible, horrible question if it really has to be done with the epsilon delta definition and not using the fact that polynomials are continuous functions. And I wrote "apparently" above because there might be some slick trick to overcome the seemingly tought difficulties. $\endgroup$ – Timbuc Jun 6 '15 at 12:02
  • $\begingroup$ @user99914 how would the continuity of polynomials be helpful? The determinant can be computed easily using the characteristic polynomial, but a perturbation in the variable $\lambda$ of the characteristic polynomial is only tantamount to a perturbation in the diagonals of the matrix $A$. $\endgroup$ – Scott Jul 3 '19 at 15:20
15
$\begingroup$

Let $M=(a_{ij})_{n\times n}$ and $M'=(a'_{ij})_{n\times n}$. Recall $$ \det M=\sum_{\sigma\in S_n}\text{sgn}(\sigma)\prod_{i=1}^na_{i,\sigma_i},\det M'=\sum_{\sigma\in S_n}\text{sgn}(\sigma)\prod_{i=1}^na'_{i,\sigma_i} $$ where $S_n$ is the set of all permutations of $\{1,2,\cdots,n\}$ and $\sigma_i=\sigma(i)$. Define $m=\max_{1\le i, j\le n}\{|a_{ij}|,|a'_{ij}|\}$. Then \begin{eqnarray*} \vert\det M-\det M'\vert&=&\bigg|\sum_{\sigma\in S_n}\text{sgn}(\sigma)(\prod_{i=1}^na_{i,\sigma_i}-\prod_{i=1}^na'_{i,\sigma_i})\bigg|\\ &\le&\sum_{\sigma\in S_n}\bigg|\prod_{i=1}^na_{i,\sigma_i}-\prod_{i=1}^na'_{i,\sigma_i}\bigg|\\ &=&\sum_{\sigma\in S_n}\bigg|a_{1,\sigma_1}a_{2,\sigma_2}\cdots a_{n,\sigma_n}-a'_{1,\sigma_1}a'_{2,\sigma_2}\cdots a'_{n,\sigma_n}\bigg|\\ &=&\sum_{\sigma\in S_n}\bigg|(a_{1,\sigma_1}-a'_{1,\sigma_1})a_{2,\sigma_2}a_{3,\sigma_3}\cdots a_{n,\sigma_n}+a'_{1,\sigma_1}(a_2-a'_{2,\sigma_2})a_{3,\sigma_3}\cdots a_{n,\sigma_n}\\ &&+a'_{1,\sigma_1}a'_{2,\sigma_2}(a_{3,\sigma_2}-a'_{3,\sigma_3})\cdots a_{n,\sigma_n}+\dots+a'_{1,\sigma_1}a'_{2,\sigma_2}a'_{3,\sigma_3}\cdots a'_{n-1,\sigma_{n-1}}(a_{n,\sigma_n}-a'_{n,\sigma_n})\bigg|\\ &\le&\sum_{\sigma\in S_n}\sum_{i=1}^nm^{n-1}|a_{i,\sigma_i}-a'_{i,\sigma_i}|. \end{eqnarray*} For $\forall \varepsilon>0$, for $\delta=\frac{\varepsilon}{2nn!m^{n-1}}$, let $$ \|M-M'\|<\delta. $$ Then we have $|a_{i,j}-a'_{ij}|<\delta$ for all $1\le i,j\le n$ and hence $$ |\det M-\det M'|\le\sum_{\sigma\in S_n}\sum_{i=1}^nm^{n-1}\delta=nn!m^{n-1}\delta=\frac{\varepsilon}{2}<\varepsilon. $$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ do you know of a text or a paper where one might cite this factoid? $\endgroup$ – Scott Jul 3 '19 at 15:22
5
$\begingroup$

For a more geometric approach, you can use the following inequality: \begin{equation}\tag{1} |\det M|\le \prod_{j=1}^n \lVert M_j\rVert. \end{equation} Here $\lVert M_j\rVert$ denotes the Euclidean norm of the $j$-th column of $M$. This inequality is best understood if the determinant of $M$ is interpreted as the signed volume of the parallelogram spanned by the vectors $M_1\ldots M_n$. It expresses the intuitive fact that a skew parallelogram has smaller volume than the right parallelogram having the same sides. (Look here for more information and a proof).

Observing that $\det(M)$ is a linear map in each column $M_j$, one obtains the following formula, in which the right hand side is a telescoping sum: $$ \det M-\det M'=\sum_{j=1}^n\det\begin{bmatrix} M'_1&\ldots& M'_{j-1}& M_j-M'_j& M_{j+1}&\ldots& M_n\end{bmatrix}.$$ Taking absolute values and using inequality $(1)$ one has $$\tag{2} \lvert \det M-\det M'\rvert \le \sum_{j=1}^n\lVert M'_1\rVert\ldots \lVert M_j-M'_j\rVert\ldots\lVert M_n\rVert.$$

The inequality $(2)$ shows that $\det$ is locally Lipschitz continuous.

Indeed, if $C>0$ is such that $$\lVert M_k\rVert\le C,\ \lVert M'_k\rVert\le C, \qquad \forall k=1\ldots n, $$ then using the elementary inequality $$ \sum_{j=1}^n a_j\le \sqrt{n}\left(\sum_{j=1}^n a_j^2\right)^{\frac{1}{2}}\qquad \forall a_1\ldots a_n\ge 0, $$ one sees from $(2)$ that $$ \begin{split} \lvert \det M-\det M'\rvert&\le C^{n-1}\sum_{j=1}^n\lVert M_j-M'_j\rVert \\ &\le \sqrt{n}C^{n-1}\sqrt{\sum_{j=1}^n \lVert M_j-M'_j\rVert^2} \\ &=\sqrt{n}C^{n-1}\lVert M-M'\rVert_{\mathrm{matrix}} \end{split} $$ where $\lVert M\rVert_{\mathrm{matrix}}^2=\sum_{k=1}^n\lvert M_{i\,j}\rvert^2.$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Side note: The "telescoping sum trick" is exactly the same xpaul used. This proof and xpaul's one are not so different, actually. This proof is "column-oriented" while xpaul's is "element-oriented" but that's all. The algebraic machinery is the same and it relies on the multilinear character of the determinant function. $\endgroup$ – Giuseppe Negro Jun 9 '15 at 21:01
  • $\begingroup$ Do you know of a text or paper where one might cite this factoid? $\endgroup$ – Scott Jul 3 '19 at 15:22
  • $\begingroup$ @Scott: Which "factoid"? I am sorry, I don't understand what you mean. $\endgroup$ – Giuseppe Negro Jul 3 '19 at 15:48
  • $\begingroup$ That the determinant is a continuous function. Is this proof you've provided available in a textbook somewhere? $\endgroup$ – Scott Jul 4 '19 at 9:20
  • $\begingroup$ Well, surely it is, but I don't know exactly. You can safely dismiss this as "well-known", though, no need for an exact reference. $\endgroup$ – Giuseppe Negro Jul 4 '19 at 9:22

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.