Initial Condition Sensitivity Mapping¶

Why It Matters¶

Not all uncertainties are equally important. Humans waste effort measuring everything equally or fixate on variables that feel important but don't drive outcomes. Sensitivity mapping reveals leverage points - the specific measurements or decisions that most reduce future uncertainty - and helps allocate limited information-gathering resources efficiently.

The Key Move¶

Perturb each initial condition or parameter slightly (say, ±10%) and measure the change in outcome at your forecast horizon. Rank the sensitivities. Focus your measurement and control efforts on the high-sensitivity variables; accept uncertainty in the low-sensitivity ones. When you can't improve initial conditions further, that sensitivity map tells you where your confidence will degrade fastest.

Classic Application¶

Weather model initialization. A 1°C error in sea-surface temperature might have huge effects on hurricane intensity forecasts but minimal effect on mid-latitude winter storms. A 10% error in upper-level wind measurements propagates rapidly in some flow patterns but not others. Sensitivity analysis tells forecasters which observations to prioritize (dropsonde deployments, satellite channels) for each specific forecast scenario.

Surprising Application¶

Startup strategy. Which early decisions have high sensitivity to long-term outcomes? Market choice might be very sensitive (wrong market = dead), while office location might be insensitive (mostly doesn't matter). Product feature set might be highly sensitive in consumer markets but insensitive in enterprise. Mapping these sensitivities helps founders allocate limited attention: obsess over the sensitive variables, use heuristics for the rest. Prevents "startup theater" of optimizing insensitive details.

Failure Modes¶

Linear sensitivity assumptions - many systems have nonlinear sensitivities where small perturbations matter only when they cross thresholds. Time-varying sensitivity - what's sensitive now may not be later, and vice versa. Ignoring interaction effects - two insensitive variables might be highly sensitive in combination. Over-narrowing - focusing only on the top sensitivity can miss combinations of moderate sensitivities that matter collectively. Confusing sensitivity with importance - a sensitive parameter might not be controllable.

Go Deeper¶

Stensrud, Parameterization Schemes: Keys to Understanding Numerical Weather Prediction Models, Chapter 8; Saltelli et al., Global Sensitivity Analysis: The Primer (2008).

Leverage Points - Operations Research & Systems Thinking: A Map of the Territory