Learning to Read Equations

Here is why this matters for reading research. The same Σ appears in a physics paper summing energies over particles, in a genetics paper summing over positions in a DNA sequence, and in a machine-learning paper summing the loss over training examples. The operator is identical; what is being summed is not. That distinction (shared machinery, different subject matter) is the backbone of this whole curriculum.

When you see $\sum_i$ with no bounds written, the convention is "over every value of $i$ that makes sense here", usually every item in a dataset or every element of a vector. Authors drop the bounds when they think the context makes them obvious. Your first question should be: what is $i$ ranging over?

An equation is a claim about how quantities relate. Before reading it, sort the symbols into three bins:

• Variables: quantities that change within the story the equation tells (model size $N$, time $t$, position $x$).
• Parameters: numbers that are fixed for one system but differ between systems, often fitted from data ($A$, $\alpha$ in a scaling law; a spring constant $k$).
• Constants: fixed by nature or mathematics ($c$, the speed of light; $e \approx 2.718$; $\pi$).

The line between parameter and constant is a modelling choice, and good papers tell you which is which. When they don't, that is a question to ask.

Function notation is where programmers have a head start. $f(x)$ is a function call. $L(N)$ is "loss as a function of $N$". $P(A \mid B)$ is a probability function evaluated on an event with a condition attached. When a paper writes $L(\theta)$, it is telling you the loss depends on the parameters $\theta$. That is a type signature as much as a formula.

Subscripts and superscripts are the other everyday notation. A subscript is usually an index or label: $x_i$ is the $i$-th item of the array $x$, and $\theta_t$ is the parameters at step $t$. $L_\infty$ is a label meaning "$L$ in the limit". A superscript is usually a power ($x^2$), but sometimes a label ($W^{Q}$ in attention means "the query weight matrix", not $W$ raised to the power $Q$). If a superscript is a letter that couldn't sensibly be an exponent, it is a label.

Rearranging is the one algebra skill that matters for reading. It lets you ask a different question of the same relationship. From $\lambda = c/f$ you can get $f = c/\lambda$ by doing the same operation to both sides. The rule is that both sides must stay equal: multiply both by $f$, then divide both by $\lambda$. Papers do this silently all the time. When a line seems to come from nowhere, it is usually a rearrangement of the previous one.

A graph is a function drawn. Before looking at the curve, read the axes: what is on each, what are the units, and is the spacing linear (equal steps add the same amount) or logarithmic (equal steps multiply by the same factor)?

Almost every important plot in a scaling-law paper is log-log. On a linear axis the ticks go 0, 10, 20, 30; on a log axis they go 1, 10, 100, 1000. This one choice changes what shapes mean. On log-log axes a power law appears as a straight line, and its slope is the exponent. That is why "the line is straight" is the headline finding of so many AI papers.

Negative exponents mean division. $N^{-\alpha}$ is $1 / N^{\alpha}$. As $N$ grows, $N^{-\alpha}$ shrinks toward zero, faster for larger $\alpha$. This is how papers write "decreases with scale" compactly. A fractional exponent is a root: $x^{1/2} = \sqrt{x}$, and $x^{3/4}$ is between linear and square-root growth.

Which log? Natural log ($\ln$, base $e$) dominates physics and ML math. Base 2 dominates information theory because the answer comes out in bits. Base 10 shows up on plot axes. Many ML papers write $\log$ and mean $\ln$. Changing base only multiplies the result by a constant, so the shape of every relationship is unchanged; only the units differ (bits vs "nats").

Scientists often care more about how one thing scales with another than about the exact numbers. The symbol $\propto$ ("is proportional to") discards the constant and keeps the shape. $y \propto x^2$ says "double $x$, quadruple $y$" without committing to units or a prefactor.

This is a powerful reading habit. Faced with a long equation, ask: if I double this input, what happens to the output? For $y \propto x$ it doubles. For $y \propto x^2$ it quadruples. For $y \propto x^{-1}$ it halves. For $y \propto \log x$ it adds a fixed amount. That is the difference between something that runs away and something that crawls.

Two more symbols live here. $\approx$ means "approximately equal". Notice it, because it tells you an approximation has been made, and every approximation has conditions under which it fails. $\Delta$ (capital delta) means "change in": $\Delta x = x_{\text{after}} - x_{\text{before}}$. You will also see $\Delta$ as a difference between two conditions ("$\Delta$ accuracy") in ML results tables.

Now put it all together on an equation you have seen in AI writing:

Use this five-pass routine whenever an equation stops you in a paper. It takes about two minutes and works without derivation.

1. Inventory. List every symbol. Mark each as operator, variable, parameter or constant. Find where the paper defines it (usually within a few lines, or in a notation table).
2. Type signature. What goes in, what comes out? Write it as a function call: "loss(params, data) → number".
3. Shape. Identify the skeleton: sum, product, ratio, power, exponential, log. Ask "double this input, what happens?"
4. Status. Is this a definition, an identity, a model assumption, a fitted empirical form, or a derived result? The surrounding words ("we define", "we assume", "we find", "it follows") tell you.
5. Limits. Under what conditions was it shown? What range of data? What would break it?