# Partial derivative with matrices

I have reforumulated my problem of computing some quantities $\mathbf{a}\in R^{m}$ from $\mathbf{b}\in R^{n}$ in a matricial form:

$$\mathbf{b} = (C\odot(\mathbf{1}_{n}\cdot \mathbf{a}^{T}))\cdot \mathbf{1}_{m}$$

where $\mathbf{C}\in R^{n\times m}$ and $\odot$ is the Hadamard (element-wise) product.

Now I'd like to compute the derivatives for my quantities in $\mathbf{a}$ (image to use them in an update rule for a gradient descent optimization step) and I'd like to derive them from my matrix formulation, so what I am trying to do is to compute $\frac{\partial \mathbf{b}}{\partial \mathbf{a}}$ (even though abusing notation).

Following the Matrix Cookbook I am doing this:

$$\frac{\partial \mathbf{b}}{\partial \mathbf{a}} = \partial(C\odot(\mathbf{1}_{n}\cdot \mathbf{a}^{T}))\cdot \mathbf{1}_{m} = \\ =\mathbf{C}\odot \partial(\mathbf{1}_{n}\cdot \mathbf{a}^{T})\cdot \mathbf{1}_{m} = \\ = \mathbf{C}\odot (\mathbf{1}_{n}\cdot \mathbf{1}_{m}^{T})\cdot \mathbf{1}_{m} = \\ = \mathbf{C}\cdot \mathbf{1}_{m}$$

But this does not feel right at all since is in $R^{n}$ while it shall be in $R^{m}$ like $\mathbf{a}$ is.

In the end I would expect it to be $\mathbf{C}$ since $\mathbf{b}=\mathbf{C}\cdot\mathbf{a}$ but still... I guess that not only I am making some mistake but that I am missing something theoretically.

• There is a mistake in your derivation, $\frac {\partial a^T} {\partial a} \neq 1^T_m$ – lynn Jan 8 '15 at 3:01
• I came to that conclusion too, but what is the correct derivative instead? – rano Jan 8 '15 at 7:52

Your intuition is indeed correct    $b = C\cdot a$.
To prove it, I'll need the 3rd order tensor $\beta_{ijk}$ whose components are unity whenever $i=j=k$ and zero otherwise.
Using $\beta$ you can express Hadamard products as: \eqalign { a\circ b &= a\cdot\beta\cdot b \cr C\circ(a\cdot b') &= a\cdot\beta\cdot C\cdot\beta\cdot b \cr } One more useful property is that $I = \beta\cdot 1$
Now we're ready to attack your problem \eqalign { (C\circ(1\cdot a'))\cdot 1 &= (1\cdot\beta\cdot C\cdot\beta\cdot a)\cdot 1 \cr &= (I\cdot C\cdot\beta\cdot a)\cdot 1 \cr &= (C\cdot\beta\cdot a)\cdot 1 \cr &= C\cdot\beta: (a\cdot 1') \cr &= C\cdot\beta: (1\cdot a') \cr &= (C\cdot\beta\cdot 1)\cdot a \cr &= (C\cdot I)\cdot a \cr &= C\cdot a \cr } Those middle steps are allowed because $\beta$ has a valence of 3, and is symmetric in all of its indices.
Another way to think of $\beta$ is in terms of diagonal operations, i.e. converting a vector into a diagonal matrix or converting the diagonal of a matrix into a vector: \eqalign { \beta\cdot a &= {\rm Diag}(a) = A \cr \beta:B &= {\rm diag}(B) = b \cr } As for the derivative
\eqalign { b &= C\cdot a \cr \frac {\partial b} {\partial a^T} &= C \cr }
• I see your point and I was pretty sure that it was equal to Ca btw thanks for your explanation. However I'd like to see the correct derivative for the long form instead of the shorter one, could you provide it? – rano Jan 8 '15 at 7:54