1
$\begingroup$

I am trying to prove certain relations between certain values of the so called Brier inaccuracy measure (Brier scoring rule).

Given a vector $p = (p_1, \ldots p_n)$, where $p_1 + \ldots p_n = 1$ and $p_i \geq 0$ for all $i \leq n$ we can compute the Brier inaccuracy of the vector $p$ (assuming we are given $n$ propositions and $p_i$'s denote our degrees of belief in the corresponding propositions if the proposition $i$, corresponding to $p_i$ is actually true):

$B_i(p) = (1-p_i)^2 + \sum_{j=1, j\neq i}^n p_j^2$.

If we find out that e.g. the proposition $n$ is false, we can conditionalize on it and obtain a posterior inaccuracy:

$Q_i(p) = (1 - \frac{p_i}{1-p_n})^2 + (\frac{1}{(1-p_n)^2})\sum_{j=1, j\neq i}^{n-1} p_j^2$.

In general, it is possible that by such conditionalization the measure of inaccuracy can increase. However, I conjecture that if $p_i = max\{p_j: j<n\}$, then the inaccuracy has to decrease (and actually, if all $p_i$'s are equal for $i <n$ then I already know that it also decreases).

That is, there is the following problem:

Assume that:

(1) $0<p_i<1$ for all $i\leq n$,

(2) $p_1 + \ldots + p_n = 1$,

(3) $p_1 \geq p_2 \geq \ldots \geq p_{n-1}$,

(4) $p_1 > p_{n-1}$.

Then consider: $B_1(p) = (1-p_1)^2 + \sum_{j=2}^n p_j^2$

and:

$Q_1(p) = (1 - \frac{p_1}{1-p_n})^2 + (\frac{1}{(1-p_n)^2})\sum_{j=2}^{n-1} p_j^2$.

What I want to demonstrate is that for any vector $p$ satisfying conditions (1)-(4) it holds that:

$B_1(p) > Q_1(p)$.

I did try to compare how much of a decrease in inaccuracy one gets by the conditionalization with how much increase one gets from it.

What I have concluded is that I need to prove the folloqing inequality:

$2p_1 + \sum_{i=1}^{n-1}p_i - (\sum_{i=1}^{n-1}p_i)^2 > (\frac{1}{1-p_n})\cdot \sum_{i=1}^{n-1}p_i^2 + \sum_{i=1}^{n-1}p_i^2$.

Is there now an elementary way to use our assumptions (1)-(4) to prove the inequality?

Any hint would be helpful.

$\endgroup$
2
$\begingroup$

Clearly $\sum_{i=1}^{n-1}p_i - (\sum_{i=1}^{n-1}p_i)^2\geq 0$, so we can ignore those terms in the inequality. Since $(\frac{1}{1-p_n})\sum_{i=1}^{n-1}p_i^2 \geq \sum_{i=1}^{n-1}p_i^2$ it suffices to prove $$p_1 > (\frac{1}{1-p_n})\cdot \sum_{i=1}^{n-1}p_i^2$$

which can be written as the obvious inequality $$p_1(p_1+\cdots+p_{n-1}) > p_1^2+\dots+p_{n-1}^2$$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.