3

When is the variance of an unbiased estimator lower than the Cramèr-Rao lower bound ??

Xi'an
  • 105,342
P.mcs
  • 33

1 Answers1

5

For distributions that do not satisfy the conditions for the Cramèr-Rao lemma to apply, e.g. for these which support depends on the parameter, there is no reason for the Cramèr-Rao inequality to hold. See for instance this example from Casella & Berger (Example 7.3.13), when $$X_1,\ldots,X_n\sim\mathcal{U}(0,\theta)$$ the estimator $$\frac{n+1}{n}X_{(n)}$$ is unbiased and with variance $\theta^2/n(n+2)$ while the Cramèr-Rao lower bound is $\theta^2/n$. (See this entry on X validated about the Fisher information for Uniforms, which some argue does not exist. I tend to agree.)

StubbornAtom
  • 11,143
  • 1
  • 28
  • 84
Xi'an
  • 105,342
  • 2
    Indeed the Fisher information $I(\theta)$ does not exist here (we can arrive at different expressions for $I(\theta)$ using different formulae). But does the CR inequality really apply in these cases since the regularity conditions are not satisfied? – StubbornAtom Jan 25 '19 at 09:55
  • 1
    @StubbornAtom: The Cramèr-Rao bound lemma applies under regularity conditions that are not satisfied by the Uniform distribution, there is thus no surprise whatsoever that it is not a valid bound on the unbiased estimator in the Uniform case. The OP was apparently asking for cases as such. – Xi'an Feb 11 '19 at 06:13