Robustness: Why Does Stability Not Mean Absence of Change?

Robustness is a system’s capacity to preserve an essential function when it is disturbed. It does not mean remaining unchanged, nor does it promise success under every possible condition. The disturbance and the essential function must both be specified.

Consider speech recognition. Accurate transcription in a quiet room shows that the system works under one condition. If it still preserves the main content when speakers have different accents or mild background noise appears, it is robust to those variations. The output may change slightly without the function being lost.

Robustness differs from accuracy: a high score on one standard test does not show that performance will remain reliable when conditions shift. It is not the same as redundancy either. Redundancy uses spare components to tolerate a local failure, while robustness concerns whether function persists during disturbance. Resilience places greater emphasis on recovery after disruption.

So ‘this system is stable’ is incomplete. We must still ask: stable against which changes, over what range, and with how much acceptable deviation? Robustness is not absolute hardness but a bounded, stable relationship between function and change.

Click to access NIST.AI.100-2e2025.pdf