# Gradient vector function using sum and scalar

Could someone take a look on my attempt to compute the gradient for:

$$f(x) = \lambda \sum_{x = 1}^n g(x_i)$$

Where $x \in \mathbb{R^d}$, $\lambda \in \mathbb{R}$ and

$$g(x_i) = \begin{cases} x_i - \varepsilon/2 & \textbf{if } |x_i| \geq \varepsilon\\ x_i^2 / (2\varepsilon) & \textbf{if } |x_i| < \varepsilon\\ \end{cases}$$

This is what I have done so far:

The function $g(x_i)$ is not differentiable if $x = -\varepsilon$, for the rest:

$$\frac{\partial}{\partial\beta_i}\sum_{i=1}^n g(x_i)= \begin{cases} 1&|x_i|\ge\epsilon\;,\\ x_i/\epsilon&|x_i|\lt\epsilon\;. \end{cases}$$

For $f(x)$ I would apply the product rule:

\begin{align*} \frac{\partial}{\partial x} f(x) &= (\frac{\partial}{\partial x} \lambda) \cdot \sum_{x = 1}^n g(x_i) + \lambda \cdot (\frac{\partial}{\partial x} \sum_{x = 1}^n g(x_i))\\ &= 0 \cdot \sum_{x = 1}^n g(x_i) + \lambda \cdot (\frac{\partial}{\partial x} \sum_{x = 1}^n g(x_i))\\ &= \lambda \cdot \frac{\partial}{\partial x_i} \sum_{x = 1}^n g(x_i) \end{align*}

If this is correct, then my question is of which domain is then $\frac{\partial}{\partial x} f(x)$?

Either it is $\mathbb{R}^n$ or $\mathbb{R}$. I am not sure, for the fact, that this is a gradient I would say $\mathbb{R}^n$, but how are then the components of the resulting vector computed?

$$\begin{pmatrix} ???\\ ???\\ \vdots\\ ??? \end{pmatrix}$$

-

First, $\frac{\partial f(x)}{\partial x_i} = \lambda g'(x_i)$, so the gradient is given by $\nabla f(x) = \lambda (g'(x_1),...,g'(x_n))^T$

Second, $g'(t) = 1$, when $|t|>\epsilon$, and $g'(t) = \frac{t}{\epsilon}$when $|t| < \epsilon$.

Hence the gradient is either $(\lambda,...,\lambda)^T$ if $||x||_{\infty} < \epsilon$, or $\frac{\lambda}{\epsilon} x$ if $x_i > \epsilon$, $\forall i$. For other $x$, you need to use the appropriate value depending on $x_i$.

No product rule is needed.

And $\nabla f(x) \in \mathbb{R}^n$.

-
Finally you brought some light into the darkness. – Mahoni May 7 '12 at 20:22
Note that I fixed a minor error in the solution; the gradient is $\frac{\lambda}{\epsilon}$ only when all $x_i>\epsilon$. – copper.hat May 7 '12 at 23:18