The Cramér-Rao inequality is known for providing a lower bound on the variance of an estimator, but appropriate regularity conditions are required for this inequality to hold.

Regularity conditions

Definition (Regularity conditions). For the probability density function $p(\boldsymbol{x}; \theta)$ parametrized by $\theta \in \Theta$,

  1. the support of $p(\boldsymbol{x}; \theta)$ does not depend on $\theta$;
  2. for each $\boldsymbol{x}\in\mathcal{X}$, $p(\bm{x};\theta)$ is differentiable with respect to $\theta$;
  3. for some $\theta \in \Theta$, there exists $\delta > 0$ and non-negative function $G(\boldsymbol{x}; \theta)$,

    \[\begin{aligned} \left|\frac{p(\boldsymbol{x};\eta) - p(\boldsymbol{x};\theta)}{\eta - \theta}\right| &\leq G(\boldsymbol{x}; \theta), \\ \mathbb{E}_{\theta}[G(\boldsymbol{x};\theta)] &< \infty, \end{aligned}\]

    for some $\eta \in (\theta - \delta, \theta + \delta)$;

  4. for some $\theta \in \Theta$, there exists $\Delta > 0$ and non-negative function $H(\boldsymbol{x}; \theta)$,

    \[\begin{aligned} \left|\frac{p(\boldsymbol{x}; \lambda) - p(\boldsymbol{x}; \theta)}{(\lambda - \theta)p(\boldsymbol{x}; \theta)} - \frac{\partial}{\partial\theta}\ln p(\boldsymbol{x}; \theta)\right| \leq H(\boldsymbol{x}; \theta), \end{aligned}\]

    for some $\lambda \in (\theta - \Delta, \theta + \Delta)$

  5. Fisher information $I(\theta)$ satisfies $0 < I(\theta) < \infty$.

Example. Let $x_1,\dots,x_n \sim U(0, \theta)$ for some parameter $\theta \in \Theta = \mathbb{R}_+$. The joint probability density function of $\boldsymbol{x} = (x_1,\dots,x_n)$ is given as

\[p(\boldsymbol{x}; \theta) = \begin{cases} 1 / \theta^n & (0 \leq \min_i x_i, \max_i x_i \leq \theta), \\ 0 & (\text{otherwise}). \end{cases}\]

Then, we can see that the support of $p(\boldsymbol{x}; \theta)$ depends on the parameter $\theta$ as

\[\text{supp}\ p(\boldsymbol{x};\theta) = \lbrace\boldsymbol{x}; p(\boldsymbol{x};\theta) > 0\rbrace = \lbrace \boldsymbol{x}; 0 \leq \min_i x_i,\dots,\max_i x_i \leq \theta \rbrace,\]

and violates the regularity conditions.

Chapman-Robbins bound

The following theorem is a relaxation for the non-regular case of the Cramér-Rao lower bound.

Theorem (Chapman-Robbins bound). For some $\theta \in \Theta$, we assume that there exists $\xi \in \Theta$ satisfying $\text{supp}\ p(\boldsymbol{x}; \xi)\subset \text{supp}\ p(\boldsymbol{x}; \theta)$ and $\xi \neq \theta$. Here, for the unbiased estimator $\hat{g}$ of some $g(\theta)$, we have the following bound.

\[V_{\theta}(\hat{g}) \geq \sup_{\xi} \frac{\lbrace g(\xi) - g(\theta) \rbrace^2}{V_{\theta}(p(\boldsymbol{x};\xi) / p(\boldsymbol{x}; \theta))}.\]

Proof. When $V_{\theta}(\hat{g}) = \infty$, it is clear that the inequality holds. Then, we assume that $V_{\theta}(\hat{g}) < \infty$. For any $\theta \in \Theta$, we have $\mathbb{E}_{\theta}(\hat{g}) = g(\theta)$ and

\[\begin{aligned} \text{Cov}_{\theta}\left(\hat{g}, \frac{p(\boldsymbol{x}; \xi)}{p(\boldsymbol{x}; \theta)} - 1\right) &= \mathbb{E}_{\theta}\left[\left(\hat{g} - g(\theta)\right)\left(\frac{p(\boldsymbol{x}; \xi)}{p(\boldsymbol{x};\theta)} - 1\right)\right] \\ &= g(\xi) - g(\theta). \end{aligned}\]

Thus, from the Cauchy–Schwarz inequality,

\[\begin{aligned} \left\lbrace\text{Cov}_{\theta}\left(\hat{g}, \frac{p(\boldsymbol{x}; \xi)}{p(\boldsymbol{x}; \theta)} - 1\right)\right\rbrace^2 &\leq V_{\theta}(\hat{g})V_{\theta}\left(\frac{p(\boldsymbol{x}; \xi)}{p(\boldsymbol{x}; \theta)} - 1\right) \\ &= V_{\theta}(\hat{g})V_{\theta}\left(\frac{p(\boldsymbol{x}; \xi)}{p(\boldsymbol{x}; \theta)}\right). \end{aligned}\]

Finally, we have

\[V_{\theta}(\hat{g}) \geq \frac{\lbrace g(\xi) - g(\theta) \rbrace^2}{V_{\theta}(p(\boldsymbol{x};\xi) / p(\boldsymbol{x}; \theta))},\]

and from $\text{supp}\ p(\boldsymbol{x}; \xi)\subset \text{supp}\ p(\boldsymbol{x}; \theta)$, the Chapman-Robbins bound is obtained.

Remark. For $\xi = \theta + \delta$ with $\delta \neq 0$ and $g(\theta) = \theta$, we have

\[\begin{aligned} V_{\theta}(\hat{\theta}) &\geq \frac{\delta^2}{\inf_{\delta \neq 0} V_{\theta}\left(\frac{p(\boldsymbol{x}; \theta + \delta)}{p(\boldsymbol{x}; \theta)}\right)} \\ &= \frac{\delta^2}{\inf_{\delta \neq 0}\mathbb{E}_{\theta}\left[\left\lbrace\frac{p(\boldsymbol{x}; \theta + \delta)}{p(\boldsymbol{x}; \theta)} - 1\right\rbrace^2\right]} \end{aligned}\]

from the Chapman-Robbins bound. Here, we can obtain the Cramér-Rao lower bound by $\delta \to \infty$ under the regularity conditions.