Help understand a proof in the paper "Likelihood Ratio Tests for Monotone Functions"

48 Views Asked by At

I'm reading the paper Likelihood Ratio Tests for Monotone Functions by Moulinath Banerjee and Jon A. Wellner. I got stuck at equality 2.8, where the authors used summation by parts to derive 2.8 from 2.7.

Suppose we have $\Delta_i \sim Bernoulli(\omega_i)$, we want to obtain the $\hat{\omega_i}$ that maximize the log-likelihood

$$\Phi(\omega_1, \omega_2, ..., \omega_m) = \sum_{i=1}^m \bigg\{\Delta_i\log\omega_i + (1 - \Delta_i)\log (1-\omega_i)\bigg\}\ ,$$

under the constaint that $$0 < \omega_1 \leq \omega_2 \leq ... \leq \omega_m \leq \theta_0\ .$$

For such $\hat{\omega_i}$ to exist, the authors claimed that

$$\sum_{i=1}^m \left[\frac{\Delta_i}{\hat{\omega_i}} - \frac{1 - \Delta_i}{1 - \hat{\omega_i}}\right]\hat{\omega_i} = \sum_{i=1}^m \left[\frac{\Delta_i}{\hat{\omega_i}} - \frac{1 - \Delta_i}{1 - \hat{\omega_i}}\right] \theta_0$$ has to hold.

To prove the claim, the authors started from the concavity of $\Phi$, by defining $$S_i = \sum_{j\leq i} \left[ \frac{\Delta_j}{\hat{\omega_j}} - \frac{1 - \Delta_j}{1 - \hat{\omega_j}} \right]\ , i=1, ..., m\ ,$$

then

\begin{align} \frac{d}{dt}\Phi((1-t)\hat{\omega} + t\omega)\bigg\rvert_{t=0} &= \sum_{i=1}^m \left[\frac{\Delta_i}{\hat{\omega_i}} - \frac{1 - \Delta_i}{1 - \hat{\omega_i}}\right](\omega_i - \hat{\omega_i}) \tag{2.7}\\ &= -\sum_{i=1}^m S_i\left[\omega_{i+1} - \omega_i - (\hat{\omega_{i+1}} - \hat{\omega_i})\right] \tag{2.8}\\ &\leq 0\ . \end{align}

What makes me confused was how to derive equation (2.8) from (2.7). The authors said they used summation by parts, with $\hat{\omega}_{i+1} = \omega_{i+1} = \theta_0$. Could anyone help me to understand how this summation by parts works here?

1

There are 1 best solutions below

0
On BEST ANSWER

Define $\sigma_N:=\sum_{n=0}^Na_nb_n$ and $B_n:=\sum_{k=0}^nb_k$. General summation by parts looks like:

$$\sigma_N=a_NB_N-\sum_{n=0}^{N-1}B_n(a_{n+1}-a_n).$$

So let $b_i:=\frac{\Delta_i}{\hat{\omega}_i}-\frac{1-\Delta_i}{1-\hat{\omega}_i}$ and $a_i:=\omega_i-\hat{\omega}_i$.

Are you able to see the result from here?