Asymptotic information loss from grouping (source code)

= Asymptotic information loss from grouping

Suppose $T_1,\ldots,T_n$ are <independent and identically distributed> with an <exponential distribution> of rate $\theta>0$, but only the bin indices $N_i=\lfloor T_i/\Delta\rfloor$ are observed, for known bin width $\Delta>0$. These are <geometric random variables> on the nonnegative <integers>, with
$$
\mathbb P_\theta(N_i=k)=(1-e^{-\theta\Delta})e^{-\theta\Delta k},\qquad
\mathbb EN_i=\frac1{e^{\theta\Delta}-1}.
$$
When $\overline N>0$, maximizing the grouped <likelihood> gives the rate <maximum-likelihood estimator>
$$
\widehat\theta=\frac1\Delta\log\left(1+\frac1{\overline N}\right).
$$
If all bin indices are zero, the <likelihood> increases toward its supremum as $\theta\to\infty$, so there is no finite rate maximizer.

Using $\operatorname{var}(N_i)=e^{\theta\Delta}/(e^{\theta\Delta}-1)^2$, the <central limit theorem> and <delta method> give
$$
\sqrt n(\widehat\theta-\theta)\xrightarrow{d}
N\left(0,\frac{(e^{\theta\Delta}-1)^2}{\Delta^2e^{\theta\Delta}}\right).
$$
For the ungrouped observations, $1/\overline T$ has limiting <variance> $\theta^2$ on this $\sqrt n$ scale. The grouped limiting <variance> equals $4\sinh^2(\theta\Delta/2)/\Delta^2$, which exceeds $\theta^2$ for positive $\theta\Delta$. Thus grouping reduces <asymptotic relative efficiency> by the ratio $\{\theta\Delta/[2\sinh(\theta\Delta/2)]\}^2$. This ratio tends to one as $\Delta\downarrow0$: sufficiently fine bins lose little information, whereas a fixed bin width causes a genuine persistent precision loss. Setting $\Delta=1$ recovers $(e^\theta-1)^2/e^\theta$.