Differentials

A differential is simply a quick estimate of how much an output changes when you nudge the inputs a little. It's how engineers figure out how much a small measurement error in each input throws off the final answer. The formula $dz = f_x\,dx + f_y\,dy$ looks technical, but it's two simple products added up: $f_x\,dx$ (how sensitive $z$ is to $x$, times how much $x$ moved) and $f_y\,dy$ (the same for $y$). Each input gets its own share of the blame. That's the whole idea.

You've seen this before in one variable: $dy = f'(x)\,dx$, reading the change off the tangent line instead of the curve. This is the same trick one dimension up. Zoom in close enough on a smooth surface and it looks flat, like its tangent plane. On a flat plane, a step $dx$ east changes the height by (slope east) times $dx$, a step $dy$ north changes it by (slope north) times $dy$, and since it's flat, the two effects simply add. For small steps the surface and the plane are almost the same, so $\Delta z \approx dz$.

Error propagation. This is where differentials really earn their keep. If $x$ and $y$ are measured with errors up to $dx$ and $dy$, the worst case is when both errors push $z$ the same way, so $|dz| \le |f_x||dx| + |f_y||dy|$ bounds the error in $z$. It also tells you which measurement to improve: the one with the bigger term.

Relative error works out especially nicely for products of powers. If $V = \pi r^2h$, then dividing $dV$ by $V$ gives $\frac{dV}{V} = 2\frac{dr}{r} + \frac{dh}{h}$. The exponents become weights: a 1% error in $r$ counts twice, because $r$ is squared.

Manufacturing tolerance stacks and the error bars on any computed quantity in a lab come from total differentials.