sensitivity analysis
A model has many knobs — input parameters — and a prediction that comes out the other end. A natural and important question is: if I wiggle this knob, how much does the answer move? And which knobs matter most? Sensitivity analysis is the systematic study of how the output of a model responds to changes in its inputs. It tells you which inputs your result is delicate to (so you should measure those carefully) and which it barely notices (so you can stop worrying about them) — turning a black box into something you understand and can prioritise.
There are two broad styles. LOCAL sensitivity analysis asks how the output changes for a tiny nudge around one specific operating point — it is essentially a partial derivative: sensitivity of output Y to input x_i is the partial derivative dY/dx_i, often estimated by a finite difference (perturb x_i by a small amount, see how much Y changes, divide) or, more cleanly, by automatic differentiation or an adjoint method. Local analysis is cheap and exact-at-a-point but only describes the immediate neighbourhood. GLOBAL sensitivity analysis instead explores the whole range of plausible inputs at once, asking how much each input contributes to the output's total variability — variance-based 'Sobol indices' are the standard tool, apportioning the output variance among the inputs and their interactions. Global methods are more honest for nonlinear models with large uncertainties but cost many model runs (often via Monte Carlo sampling).
Sensitivity analysis is a close partner of uncertainty quantification: UQ tells you HOW uncertain the output is, while sensitivity analysis tells you WHY — which inputs to blame, and therefore where reducing input uncertainty would most reduce output uncertainty. It is invaluable for model simplification (fix the inputs that don't matter), experimental design (measure the inputs that do), and debugging (an output that is wildly sensitive to a parameter it physically shouldn't depend on hints at a bug). The honest cautions: LOCAL sensitivity at one point can badly mislead for a nonlinear model — a parameter that looks unimportant near your operating point may dominate elsewhere — so do not over-trust a single derivative. And a finite-difference sensitivity inherits the usual step-size dilemma: too large a perturbation and you measure nonlinearity instead of the slope; too small and round-off swamps the tiny difference.
A chemical-yield model has inputs temperature T, pressure P, and catalyst amount c. A local analysis perturbs each by 1% and finds yield changes by 8% for T, 0.5% for P, 3% for c — so yield is most sensitive to temperature. The team then measures T precisely and stops fussing over P. A global Sobol analysis later confirms T explains most of the output variance, with a notable T-by-c interaction a single-point derivative had missed.
Ranking how much each input moves the output shows where to measure carefully and what to ignore.
Local sensitivity at one point can mislead for nonlinear models — a parameter unimportant near your operating point may dominate elsewhere. Finite-difference sensitivities also face the step-size dilemma: too large measures curvature, too small drowns in round-off.