@(@\newcommand{\B}[1]{ {\bf #1} }
\newcommand{\R}[1]{ {\rm #1} }
\newcommand{\W}[1]{ \; #1 \; }@)@This is dismod_at-20221105 documentation: Here is a link to its
current documentation
.
Problem
For this case the only data is we a sequence of positive measurements of
Sincidence
which we denote by @(@
y_i1
@)@ for @(@
i = 0 , \ldots , N-1
@)@.
We model @(@
y_i
@)@ as independent and Gaussian with mean equal
to the true value of iota @(@
\iota
@)@,
and standard deviation @(@
\sigma_i
@)@.
The negative log likelihood, up to a constant w.r.t @(@
\iota
@)@, is
@[@
f( \iota ) =
\frac{1}{2} \sum_{i=0}^{N-1} \left( \frac{y_i - \iota}{\sigma_i} \right)^2
@]@
Optimal Solution
The optimal estimator for @(@
\iota
@)@ satisfies the equation
@[@
0 = f'( \hat{\iota} ) =
- \sum_{i=1}^{N-1} \frac{y_i - \hat{\iota} }{\sigma_i^2}
@]@
Weighted Residuals
We use the notation @(@
r_i
@)@ for the weighted residuals
@[@
r_i = \frac{y_i - \hat{\iota}}{\sigma_i}
@]@
If @(@
\hat{\iota}
@)@ were the true value for @(@
\iota
@)@,
the weighted residuals would be mean zero and variance one.
But @(@
\hat{\iota}
@)@ is instead the maximum likelihood estimator and
@[@
0 = \sum_{i=1}^{N-1} \frac{y_i - \hat{\iota} }{\sigma_i^2}
@]@
@[@
0 = \sum_{i=1}^{N-1} \frac{r_i}{\sigma_i}
@]@
Note that if @(@
\sigma_i
@)@ were the same for all @(@
i
@)@,
the sum of the weighted residuals @(@
\sum_i r_i
@)@ would be zero.
CV Standard Deviations
We consider the case were we a coefficient of variation @(@
\lambda
@)@
is used to model the measurement noise; @(@
\sigma_i = \lambda y_i
@)@.
In this case
@[@
0 = \sum_{i=1}^{N-1} \frac{r_i}{y_i}
@]@
Weighted Average of Weighted Residuals
We define the weight @(@
w_i
@)@ by
@[@
w_i = \sigma_i^{-1} / \sum_{i=0}^{N-1} \sigma_i^{-1}
@]@
The corresponding weighted average of the weighted residuals is zero; i.e,
@[@
0 = \sum_{i=1}^{N-1} w_i r_i
@]@