error propagation

Every measured quantity carries an uncertainty, and when you combine measurements into a derived result, those uncertainties travel into the answer. Error propagation is the calculus of how an uncertainty in the inputs becomes an uncertainty in the output — the systematic way to attach an honest plus-or-minus to a quantity you computed rather than measured directly.

The engine is the total differential. If a result q depends on measured quantities x, y, z, a small input error produces a result error dq = (partial q / partial x) dx + (partial q / partial y) dy + (partial q / partial z) dz, so each variable's contribution is its partial derivative (a sensitivity) times that variable's error. For a worst-case bound you add the absolute values. For independent random errors, however, errors partly cancel, and the standard rule combines them in quadrature: the variance of q is the sum of (partial q / partial x)^2 times the variance of x, and so on — the familiar 'add the squares' formula. Either way, the partial derivatives are exactly the amplification factors telling you which inputs your result is most sensitive to.

This is daily bread in every experimental science and engineering discipline. It tells you that a quantity formed by multiplying or dividing combines relative (percentage) errors, while a sum or difference combines absolute errors; it warns you that subtracting two nearly equal large numbers can blow up the relative error catastrophically (catastrophic cancellation); and it guides experimental design by revealing which measurement to improve for the biggest payoff. The honest caveat: this is a first-order (linear) approximation. When errors are large, or the function is strongly curved, the linear estimate can mislead and you need higher-order terms or a direct simulation.