1
$\begingroup$

Consider the following CDF of a binomial distribution with $p\in(0,1)$ and ${\lfloor k\rfloor}\in [0,n]$ \begin{equation*} F(k;n,p) = \sum_{i=0}^{\lfloor k\rfloor} \binom{n}{i} p^i (1-p)^{n-i} \end{equation*}

How does $F(k;n,p)$ change in the following circumstances

  • $p$ increase with $k,n$ fixed
  • $n$ increase with $k,p$ fixed

and how to prove these monotonicity conclusions?

Some numerical simulation results: $F(1,3,0.3)=0.784$, $F(1,3,0.1)=0.9720$, $F(1,4,0.1)=0.9477.$

$\endgroup$
  • 3
    $\begingroup$ Intuitively: $F(k;n,p)$ is essentially $P(X\le k)$, where $X\sim Binomial(n,p)$. That is, it is the probability that "the number of successes in $n$ Bernoulli trials is small" (where "small" means $\le k$). So if $p$ is increased (and all else is fixed), I would expect this probability to decrease, since increasing $p$ increases the success probability of each trial, so you would expect more successes. Similarly, if $n$ is increased (all else fixed), I would expect more successes to occur, since there are more trials (more opportunities for success), so $F(k;n,p)$ should decrease. $\endgroup$ – Minus One-Twelfth Aug 10 at 1:20
0
$\begingroup$

(Edited)

For the dependence on $p$, notice that $$ \frac{\partial \, p^i (1-p)^{n-i}}{\partial p}= \frac{i}{p(1-p)}p^i (1-p)^{n-i} -\frac{n}{1-p} \, p^i (1-p)^{n-i} \tag 1$$

Also, if $X$ is a $(n,p)$ Binomal (hence $E[X]=np$), let $X^{(k)}$ be $X$ truncated to $[0, \lfloor k \rfloor]$

Then $$P(X^{(k)}=i)=\frac{1}{F(k,n,p)} \binom{n}{i} p^i (1-p)^{n-i} \,[0\le i \le\lfloor k \rfloor] \tag 2$$ and

$$ \begin{align}\frac{\partial \, F(k,n,p)}{\partial p} &= \frac{1}{p(1-p) }F(k,n,p) E[X^{(k)}]-\frac{n}{1-p} F(k,n,p) \\ &= \frac{F(k,n,p)}{p(1-p)} ( E[X^{(k)}] - E[X] ) \tag 3 \end{align} $$
But $E[X^{(k)}] < E[X]$ (except for the trivial case $\lfloor k \rfloor = n$). Hence the derivative is negative and $F(k,n,p)$ decreases with $p$.


Corrected: I had the wrong sign in $(3)$ (confirmation bias!). Now it's correct (checked numerically). And, yes, $(2)$ is right, it's the distribution of a truncated Binomial (which of course corresponds to the distribution of $X$ conditioned on $X\le k$).


Added: For the dependence on $n$: letting $X_{n,p}$ be a Binomial $(n,p)$, and with $k$ integer, we have

$$F(n,k,p)= P(X_{n,p}\le k)= \sum_{i=0}^k \binom{n}{i}p^i(1-p)^{n-i} $$

and $$F(n,k-1,p) = F(n,k,p) - \binom{n}{k}p^k(1-p)^{n-k}$$ Hence $$\begin{align} F(n+1,k,p)&= P(X_{n+1,p} \le k)\\ &= P(X_{n+1,p} \le k \mid X_{n,p} < k) P(X_{n,p} < k) + P(X_{n+1,p} \le k \mid X_{n,p} = k) P(X_{n,p} = k) \\ &= 1 \times F(n,k-1,p) + (1-p) \binom{n}{k}p^k(1-p)^{n-k}\\ &= F(n,k,p) - \binom{n}{k}p^k(1-p)^{n-k} + (1-p) \binom{n}{k}p^k(1-p)^{n-k}\\ &= F(n,k,p) - \binom{n}{k}p^{k+1}(1-p)^{n-k} \end{align} $$ Then $F(n+1,k,p)<F(n,k,p)$

$\endgroup$
  • $\begingroup$ I think, from simulations, both of them are actually decreasing functions. For fixed k,p while varying n, you also need to consider the effect from the polynomial $p^i(1−p)^{n−i}$ which is decreasing as n increases. Example: $F(1,3,0.3)=0.784$, $F(1,3,0.1)=0.9720$, $F(1,4,0.1)=0.9477$. $\endgroup$ – Mathexx Aug 7 at 3:05
  • $\begingroup$ In (2), it does not look like a truncated binomial distribution but a conditional probability. Hence, we might not have $E[X^{(k)}]<E[X]$. Also, please check the simulation result in the last comment. $\endgroup$ – Mathexx Aug 10 at 0:29
  • $\begingroup$ You are right in $F$ decreasing with $p$, Sign fixed. But the rest (including the truncated mean) is right. $\endgroup$ – leonbloy Aug 10 at 1:48
  • $\begingroup$ You are right. Do you have any hints on the second case: $k,p$ fixed but $n$ varies? From the simulation, $F$ is a decreasing on $n$ does not seem to depend on the selection of $k,p$. $\endgroup$ – Mathexx Aug 10 at 21:36
0
$\begingroup$

Restrict the problem to $k\geq 0$ since when $k<0$ $F(k;n,p)$ is the sujm of binomials with negative bottom numbers, that is, zero.

Let me give heuristic arguments that provide ideas that can easily be made into proofs.

Consider a random variate $P$ which is the sum of $n$ independent Bernoulli trials with success probability $p$. Then $F(k;n,p)$ is the probability that $P \leq k$. And for $\Delta p >0$, $F(k;n,p+\Delta p)$ is the probability that $P \leq k$ for a sum involving the slightly higher Bernoulli value $p+\Delta p$.

Now picture each trial to be done by taking a uniform $X_i$ on $(0,1)$ and accepting if it is less than $p$. For every vector $\vec{X}$ of trial outcomes going into $F(k;n,p)$ there is a matching vector of $\vec{X}$ of trial outcomes going into $F(k;n,p+\Delta p)$. And there are also some outcomes contributing to $F(k;n,p+\Delta p)$ that were not in $F(k;n,p)$, namely, those cases where for some $i$, $p < X_i \leq p+\Delta p$.

So $F(k;n,p+\Delta p) > F(k;n,p)$ and $F$ is monotonic increasing with $p$.

Now fix $k$ and $p$ and let $n$ increase to $n+1$. Well, for all $i$,

$$\binom{n+1}{i} = \frac{n+1}{n+1-i} \binom{n}{i} > \binom{n}{i} $$ so each term in the sum is monotonic increasing, so the sum is also monotic increasing with $n$.


CORRECTION

$$\binom{n+1}{i} p^{i+1} (1-p)^{n-{i+1}} = \frac{n+1}{n+1-i} \frac{p}{1-p} \binom{n}{i} p^i(1-p)^{n-i} $$ If $p\geq \frac12$ this is always greater than $\binom{n}{i}$ so each term in the sum is monotonic increasing, so the sum is also monotonic increasing with $n$.

If $p\geq \frac12$ then this is monotonic decreasing until $i$ is large enough that $(n+1) p > (n+1-i) (1-p)$. After that, in changes to monotonic decreasing.

The change point comes at

$$ i > (n+1) \frac{1-2p}{1-p}$$

$\endgroup$
  • $\begingroup$ I think, from simulations, both of them are actually decreasing functions. For fixed $k,p$ while varying $n$, you also need to consider the effect from the polynomial $p^i(1-p)^{n-i}$ which is decreasing as $n$ increases. $\endgroup$ – Mathexx Aug 5 at 19:34
  • $\begingroup$ In the first equation of the CORRECTION, should it be $\binom{n+1}{i} p^{i} (1-p)^{n+1-i}$ instead? $\endgroup$ – Mathexx Aug 9 at 23:25
  • $\begingroup$ The problem with the "change point" is that it only applies to the terms of the sum, but it doesn't say much about the whole sum. $\endgroup$ – leonbloy Aug 11 at 17:05

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.