refactor(notes): Remove unnecessary emphasis and improve clarity in S…

…GD analysis section

section/6/notes.md

@@ -305,4 +305,4 @@ An especially striking aspect of our analysis is that *none of our convergence g
 
 We initially framed the setting as having $n$ fixed data points $\{x_i\}$, each drawn from a uniform distribution over the indices $\{1,\dots,n\}$. However, if we interpret $\{x_i\}$ as i.i.d. samples from an underlying distribution $\mathcal{D}$, we can simply view $x_{i_k}$ as an independent draw from $\mathcal{D}$ at each SGD step. The mean of that distribution is $\mu$, and the variance is $\sigma^2$. The identical update analyses still apply as long as each step’s sample $x_{i_k}$ has expectation $\mu$ and variance $\sigma^2$. 
 
-Hence, even if the sample space is continuous (e.g., $x_i\sim\mathcal{N}(\mu, \sigma^2)$), and we draw a fresh $x_{i_k}$ at every iteration from that continuous distribution, **the same proofs** hold. We never used anything special about the discrete nature or the finite size $n$. Our SGD update formulas and convergence in mean argument relied only on the unbiasedness of $\nabla \ell_{i_k}(w_k)$ and the boundedness of its variance. As a result, the entire analysis generalizes seamlessly to large-scale or infinite-sample scenarios—a fact that underscores the power of stochastic gradient approaches in real-world machine learning, where data can indeed be extremely large or even effectively unlimited.
+Hence, even if the sample space is continuous (e.g., $x_i\sim\mathcal{N}(\mu, \sigma^2)$), and we draw a fresh $x_{i_k}$ at every iteration from that continuous distribution, the same proofs hold. We never used anything special about the discrete nature or the finite size $n$. Our SGD update formulas and convergence in mean argument relied only on the unbiasedness of $\nabla \ell_{i_k}(w_k)$ and the boundedness of its variance. As a result, the entire analysis generalizes seamlessly to large-scale or infinite-sample scenarios—a fact that underscores the power of stochastic gradient approaches in real-world machine learning, where data can indeed be extremely large or even effectively unlimited.