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

The question is similar to this one Implicit derivative?

Let $x_1, x_2, x_3$ be three points in $\mathbb{R}^3$, $A=(a_{ij})$ is a $3\times 3$ matrix with $a_{ii}=0$ and $a_{ij}=\frac{1}{|x_i-x_j|}$ for $1\le i,j\le 3$. My question: what is $\frac{\partial A}{\partial x_1}$, $\frac{\partial A}{\partial x_2}$ and $\frac{\partial A}{\partial x_3}$? I came across this in a paper <'Note on an inequality' by Y.Xu 2006. Annales de l'Institut Henri Poincare>, but I was puzzled with this expressions.

share|cite|improve this question
You don't think they're just the respective matrices of partial derivatives of each entry? – Cam McLeman Apr 26 '10 at 4:09
You probably should elaborate what you found puzzling... Cam's reading of the notation surely does not deserve that adjective, so you probably understood something else and/or the paper meant something else, but it is impossible to tell from what you wrote! – Mariano Suárez-Alvarez Apr 26 '10 at 4:21
I am puzzled by these partial derivatives, too. How do I take the partial derivative with respect to x_i of a_ij = 1/|x_i-x_j|, if x_i is a three-dimensional vector? I can't even do 1/|x_i|^2 . – TonyK Apr 26 '10 at 9:06
Some people use the partial derivative notation when they mean a partial differential. My guess would be that the author of the paper in question belongs to that category. – fedja Apr 26 '10 at 14:57
@miwalin: Please edit your question to include the citation to the paper. That will vastly improve it. – Pete L. Clark Apr 26 '10 at 17:12

I didn't look at the paper in too much detail, but my best guess is that $\partial_{x_i}A$ is a rank-3 tensor, in the sense that we can write $$v\cdot\partial_{x_i}A = \lim_{h\to 0} \frac{1}{h}(A[x_1, \ldots,x_i + hv, \ldots,x_p] - A[x_1,\ldots, x_p])$$

Or, in index notation, writing $a_{ij} = \frac{1}{x_i - x_j}$, we have that $$(\partial_{x_l}A)_{ijk} = \partial_{(x_l)^k} (\sum_{m} [(x_i)^m - (x_j)^m]^2)^{-1/2}$$ where $(x_l)^k$ denotes the $k$-th component of the point $x_l$. So in conjecture 1 of the linked paper, the term inside the absolute value sign in the first supremum is in fact, a vector, with the $i$ and $j$-th component in the above expression contracted against a vector $u$.

More intuitively, you can think of it as the gradient on the total space projected to the $k$-th copy of $\mathbb{R}^3$.

It is, admittedly, somewhat questionable and sloppy notation. I may have written it as the variation $\delta A / \delta x_i$ instead to get over the cognitive disconnect of taking a partial relative to a vector, but still it implicitly depends on the Euclidean structure of $\mathbb{R}^p$.

share|cite|improve this answer
I should probably add that rank-3 is not the best word for it. Let $V = \mathbb{R}^n, W = \mathbb{R}^m$, then $A: \oplus^m V \to W\otimes W$ is some matrix valued function. Then the $\partial_{x_i}A : \oplus^m V \to W\otimes W\otimes V^*$ is the tensor. – Willie Wong Apr 26 '10 at 16:26

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.