This approximation has vanishing error from a linear function (as the interval decreases) and (most importantly!) we only have finite noisy samples; so it’s essentially indistinguishable (in a statistical way) for a small enough interval given some variance.