Estimating the probability of success in the Bernoulli scheme

46 Views Asked by At

I'm reading 'Probability' by Shiryayev, in particular the paragraph 'Estimating the probability of success in the Bernoulli scheme'. The author wants to determine if the estimator (estimator of the probability $\theta$ of success in a trial) defined by:

$$S_n/n=\frac{\text{number of success on n trials}}{n}$$

is efficient, in the sense that:

$$V_\theta (S_n/n)=\inf _{T_n}V_\theta (T_n)$$

where $V_\theta(\cdot)$ is the variance (which depends on $\theta$ itself) of the generic estimator $T_n$.

In his calculations he writes:

$$p_\theta(\omega)= \theta^{\sum_{i=1}^n a_i}(1-\theta)^{n-\sum_{i=1}^na_i}=\prod_{i=1}^n\theta^{a_i}(1-\theta)^{1-a_i}$$

where $\omega=(a_1,a_2,...,a_n)$ (and $a_i\in\{0,1\}$) is an elementary event of the space $\Omega$. Then, he defines:

$$L_\theta(\omega)=\log p_\theta(\omega)$$

and calculates:

$$1=\mathbb{E}_\theta(1)=\sum_{\omega}p_\theta(\omega)$$ $$\theta=\mathbb{E}_\theta(S_n/n)=\sum_{\omega}p_\theta(\omega)S_n(\omega)/n$$ $$\Downarrow$$ $$0=\sum_{\omega}\frac{\partial p_\theta(\omega)}{\partial\theta} =\sum_{\omega}\frac{\frac{\partial p_\theta(\omega)}{\partial\theta} }{p_\theta(\omega)}p_\theta(\omega)=\mathbb{E}_\theta\left( \frac{\partial L_\theta(\omega)}{\partial\theta}\right) $$ $$1=\sum_{\omega}S_n(\omega)/n\frac{\partial}{\partial\theta}p_\theta(\omega)=\mathbb{E}_\theta\left( S_n/n\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right)$$

From this, he concludes that:

$$1=\mathbb{E}_\theta\left( (S_n/n-\theta)\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right)$$

but I can't see why, because the only things that I can obtain from the two equations above is, subtracting side by side:

$$1-0=\mathbb{E}_\theta\left( S_n/n\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right)-\mathbb{E}_\theta\left( \frac{\partial L_\theta(\omega)}{\partial\theta}\right)=\mathbb{E}_\theta\left( (S_n/n-1)\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right)$$

Where is the problem here?

Thanks.

1

There are 1 best solutions below

0
On BEST ANSWER

if you subtract a zero anyway, why not take that zero as you need it? Say, $$ 1-0=\mathbb{E}_\theta\left( S_n/n\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right)-{\color{red}\theta}\cdot\mathbb{E}_\theta\left( \frac{\partial L_\theta(\omega)}{\partial\theta}\right)=\mathbb{E}_\theta\left( (S_n/n-{\color{red}\theta})\cdot\frac{\partial L_\theta(\omega)}{\partial\theta}\right) $$ By the way, the equality that you write at the end is right too. And if we replace $1$ by any other constant, it still remains valid. And the $\theta$ is a most suitable constant since we further intend to use Cauchy–Bunyakovsky–Schwarz inequality and get variance of $S_n/n$ exactly as second moment of $S_n/n-\theta$.