I'm studying stochastic gradient descent algorithm for optimization. It looks like this:

\begin{aligned} L(w) &= \frac{1}{N} \sum_{n=1}^{N} L_n(w) \\ w^{(t+1)} &= w^{(t)} - \gamma \nabla L_n(w^{(t)}) \end{aligned}

I assume that $$n$$ is chosen randomly each time the algorithm iterates. The problem comes when my notes state that $$\Bbb E[\nabla L_n(w)] = \nabla L(w)$$. Where does this come from?

• For starters, do you understand that $E[L_n(w)] = L(w)$? Commented Oct 10, 2016 at 22:48
• If I'm understanding correctly, it sounds like you want to minimize the average of several objective functions, and you do so by following the gradient direction of one of them chosen uniformly at random from all of them. In this case simple linearity of expectation and of differentiation gives you that the expected value of a given gradient is equal to the true gradient of the sum.
– Ian
Commented Oct 10, 2016 at 22:49
• @arkeet I do have the concept of expectation of a random variable. So here I assume that the random variable is something called W which takes values w and that we are computing the expectation of a function of a random variable Commented Oct 10, 2016 at 22:51
• @Rodrigo No; as you said, $n$ is chosen randomly, so the expectation is taken over the possible values of $n$. Commented Oct 10, 2016 at 22:53

Let's assume we are talking about stochastic gradient descent where we update the weights based on a single example (not minibatch), out of a total data set of size $N$.
The total error over the whole set is given by: $$L(w) = \frac{1}{N}\sum\limits_{n=1}^{N} L_n(w)$$ Then, at every step, a random sample point $n\sim U$ is chosen, and we update the weights via: $$w \leftarrow w - \gamma\nabla L_n(w)$$ where $U$ means uniform over the data set. Now we want to know whether $\mathbb{E}_{n\sim U}[\nabla L_n(w)]=\nabla L(w)$.
The first step is probably the nastiest (although not in the discrete case I guess), but we can interchange the gradient and expectation assuming $L$ is sufficiently smooth and bounded (which it probably is). See here and here.
• Suggestion to make clear why gradient and expectation can be interchanged: $$\mathbb{E}[\nabla L_n] = \sum_i P(n=i)\nabla L_i = \nabla \sum_i P(n=i)L_i = \nabla L$$ Ist this ok? Commented Apr 21, 2023 at 6:38