TL;DR. We prove the Functional Scaling Law for SGD in infinite-dimensional linear regression by tracking coordinate errors. Each update reduces old errors and adds new ones. Under power-law assumptions and sufficiently small learning rates, this gives a loss formula for the whole learning-rate schedule. The key trade-off is simple: larger steps learn faster but add more error, and later steps determine how much of that error remains.

Linear Regression and SGD

Problem setup. We consider an infinite-dimensional linear regression model, with coordinates indexed by $j\in\{1,2,\ldots\}$. Let $\boldsymbol x=(x_j){j\ge1}$ be the input, $\boldsymbol w*$ the target coefficients, and $\boldsymbol w$ the trainable coefficients. Here, the input coordinates are independent, with $x_j\sim N(0,\lambda_j)$ and $\lambda_j$ is the variance of coordinate $j$. Predictions and labels are

$$ f_{\boldsymbol w}(\boldsymbol x)=\sum_j w_jx_j, \qquad y=\sum_j w_{*,j}x_j+\xi. $$

The label noise $\xi\sim N(0,\sigma^2)$ is independent of the input. We allow $\sigma=0$. Assume $\sum_j\lambda_j<\infty$ and $\sum_j\lambda_jw_{*,j}^2<\infty$, so these labels $y$ are well-defined.

Start from $\boldsymbol w_0=\boldsymbol0$. At step $k$, draw a fresh pair $(\boldsymbol x_k,y_k)$. Assume squared loss, the update rule of SGD gives

$$ \boldsymbol w_{k+1} =\boldsymbol w_k-\eta_k \bigl(f_{\boldsymbol w_k}(\boldsymbol x_k)-y_k\bigr)\boldsymbol x_k. $$

Test Loss as Coordinate Errors

In this section, we define the coordinate error and show that the test loss decomposes into the sum of coordinate errors and an irreducible label noise term.

<aside>

Definition 1 (Coordinate error). Let $e_{k,j}$ be the parameter error in coordinate $j$. Let $v_{k,j}$ be its expected square:

$$ e_{k,j}:=w_{k,j}-w_{*,j}, \qquad v_{k,j}:=\mathbb E[e_{k,j}^2]. $$

The expectation averages over the training samples used so far.

</aside>

Clearly, the coordinate error comes from the mismatch between our learned parameters and the target parameter $\boldsymbol{w}^\star$, and thus naturally contributes to the test loss.

<aside>

Lemma 1 (Expected excess test loss). Let $(\boldsymbol x,y)$ be a fresh test sample. The expected excess test loss is given by

$$ \begin{equation} \begin{aligned} L_k &:=\frac12\mathbb E\!\left[ \bigl(f_{\boldsymbol w_k}(\boldsymbol x)-y\bigr)^2 \right]-\frac{\sigma^2}{2}.\\ &=\frac12\sum_j\lambda_jv_{k,j}. \end{aligned} \end{equation} $$

The total expected test loss is $L_k+\sigma^2/2$.

</aside>

Dynamics of Coordinate Errors

In this section, we characterize how the coordinate error evolves during SGD updates.

What one-step SGD changes. We first show that after one SGD update, each coordinate keeps part of its old error and receives a new error from the sampled gradient.

<aside>

Lemma 2 (Dynamics of coordinate error).

$$ \begin{equation} v_{k+1,j} =\underbrace{(1-2\eta_k\lambda_j+2\eta_k^2\lambda_j^2)v_{k,j}}{\text{old error that remains}} +\underbrace{\eta_k^2\lambda_j(2L_k+\sigma^2)}{\text{new error from this step}}. \end{equation} $$

</aside>

What n-step SGD gives. We then extend Lemma 2 to $n$ updates. For simplicity, we define

$$ a_{k,j}:=1-2\eta_k\lambda_j+2\eta_k^2\lambda_j^2, \qquad b_{k,j}:=\eta_k^2\lambda_j(2L_k+\sigma^2). $$