Let $f:V_1\times\dots\times V_N \to W$ be multilinear. Then $f$ s differentiable and $$df(a_1,\dots,a_n)(h_1+\dots+h_n)=f(h_1,a_2,\dots,a_n)+f(a_1,h_2,a_3,\dots,a_n)+f(a_1,\dots,a_{n-1},h_n)\tag{1}$$

Problem: Calculate the derivative of the cross product $\mathbb R^3\times \mathbb R^3 \to \mathbb R^3$.

Solution: Let $f$ be the cross product. It's easy to show that $f$ is multilinear by a direct calculation thus we can use (1). We calculate the derivative at $(u,v)=((u_1,u_2,u_3),(v_1,v_2,v_3))$. We then have $$df(u,v)(h_1+h_2)=h_1\times v + u\times h_2\tag{2}$$ Let $e_i, i=1,2,3$ be the unit vectors in $\mathbb R^3$. We calculate $$e_1\times v=\begin{pmatrix}0\\ -v_3 \\ v_2\end{pmatrix}$$ $$e_2\times v=\begin{pmatrix}v_3\\ 0 \\ -v_1\end{pmatrix}$$ $$e_3\times v=\begin{pmatrix}-v_2 \\ v_1 \\ 0\end{pmatrix}$$ For $u$ we use the identity $u\times e_i=-e_i\times u$. So we get the following derivative matrix: $$df(u,v)=\begin{pmatrix}0&v_3&-v_2&0&-u_3&u_2 \\ -v_3 & 0 & v_1 & u_3 & 0 & -u_1 \\ v_2 & -v_1 & 0 & -u_2 & u_1 & 0\end{pmatrix}$$

Question: 1. Apprently, the $+$ means writing the result of $f$ as a column, so it's not the $+$ from e.g. $\mathbb R^3$. Should I see that from the definition above without the example? How should I know the dimension of the imge of $df$ without them specifying it somehow?

  1. What $+$ is used in $f(a_i)(h_1+h_2)$? How does $h_1$ and $h_2$ and $h_1+h_2$ look like?
  • $\begingroup$ If $f$ is $p$-linear then $f(x_1 + h_1, \ldots, x_p + h_p) = f(x_1, \ldots, x_p) + \sum_i f(x_1, \ldots, h_i, \ldots, x_p) + E(h_1, \ldots, h_p),$ where $E$ is the sum of all terms were $f$ is evaluated at least in two different $h_i.$ It follows $E(h_1, \ldots, h_p) = o(\|h\|)$ and then, $f'(x_1, \ldots, x_p) = \sum_i f(x_1, \ldots, h_i, \ldots, x_p).$ This proof does not require any sort of finite dimension. Q.E.D. $\endgroup$
    – William M.
    Jan 27, 2019 at 1:17

1 Answer 1


For a multilinear mapping, it suffices to consider its Frechet derivative. Let $W$ be an $n$-D vector space, and each $V_i$ be an $m_i$-D vector space with $i=1,2,...,N$. Let $f:V_1\times V_2\times\cdots\times V_N\to W$ be multilinear. Then $\forall\left(v_1,v_2,...,v_N\right)\in V_1\times V_2\times\cdots\times V_N$, the Frechet derivative of $f$ at this location, denoted by $({\rm d}f)(v_1,v_2,...,v_N)$, is also a multilinear mapping, i.e., $$ ({\rm d}f)(v_1,v_2,...,v_N):V_1\times V_2\times\cdots\times V_N\to W. $$ According to Frechet, it follows that \begin{align} &({\rm d}f)(v_1,v_2,...,v_N)(h_1,h_2,...,h_N)\\ &=f(h_1,a_2,...,a_N)\\ &+f(a_1,h_2,...,a_N)\\ &+\cdots\\ &+f(a_1,a_2,...,h_N). \end{align}

Recall that, if $g$ is linear, its entry-wise form reads $$ g_i(v)=\sum_ja_{ij}v_j, $$ and if $g$ is bilinear, its entry-wise form reads $$ g_i(v_1,v_2)=\sum_{j_1,j_2}a_{ij_1j_2}v_{1j_1}v_{2j_2}. $$ Inductively and formally, the above multilinear $f$ observes the following entry-wise form $$ f_i(v_1,v_2,...,v_N)=\sum_{j_1=1}^{m_1}\sum_{j_2=1}^{m_2}\cdots\sum_{j_N=1}^{m_N}a_{ij_1j_2...j_N}v_{1j_1}v_{2j_2}...v_{Nj_N} $$ for $i=1,2,...,m$, where each $v_{kj_k}$ denotes the $j_k$-th entry of $v_k\in V_k$, while $a_{ij_1j_2...j_N}$'s are the coefficients of $f$.

Thanks to this entry-wise form, we may then write down the entry-wise form of ${\rm d}f$ as well, which reads \begin{align} &({\rm d}f)_i(v_1,v_2,...,v_N)(h_1,h_2,...,h_N)\\ &=\sum_{j_1=1}^{m_1}\sum_{j_2=1}^{m_2}\cdots\sum_{j_N=1}^{m_N}a_{ij_1j_2...j_N}h_{1j_1}v_{2j_2}...v_{Nj_N}\\ &+\sum_{j_1=1}^{m_1}\sum_{j_2=1}^{m_2}\cdots\sum_{j_N=1}^{m_N}a_{ij_1j_2...j_N}v_{1j_1}h_{2j_2}...v_{Nj_N}\\ &+\cdots\\ &+\sum_{j_1=1}^{m_1}\sum_{j_2=1}^{m_2}\cdots\sum_{j_N=1}^{m_N}a_{ij_1j_2...j_N}v_{1j_1}v_{2j_2}...h_{Nj_N}. \end{align} In other words, as $a_{ij_1j_2...j_N}$'s are known, the entry-wise form of ${\rm d}f$ could be expressed straightforwardly as above.

Finally, the "$+$" in OP's original post, i.e., $(h_1+h_2+\cdots+h_N)$, is a convention in some context, which is exactly $(h_1,h_2,...,h_N)$ here. When there is free of ambiguity, both expressions can be used as per ones preference.


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.