10 Binomial Approximation

Consider I1,⋯,In∼Bernoulli(p), and Sn∼Binomial(n,p), so Sn=dI1+⋯+In. We know P(Sn=k)=(nk)pk(1−p)n−k=n!k!(n−k)!pk(1−p)n−k,0≤k≤n.
We want to estimate the Binomial distribution, because factorials are troublesome to work with. The key tool is to use Stirling Approximation e112n+1<n!(ne)n2πn<e112n.

1 Entropy Approximation

We can use entropy to approximate Binomial distribution. To be more specific, KL divergence.

Theorem (Entropy Approximation)

Let Sn∼Binomial(n,p) and define f=kn. Then for k=1,⋯,n−1,12πnf(1−f)e−nKL(f||p)>P(Sn=k)>[1−112nf(1−f)]12πnf(1−f)e−nKL(f||p), where KL(f||p)=−flog⁡(pf)−(1−f)log⁡(1−p1−f).

2 Normal Approximation

Taylor expansion of KL(f||p) (see as a function about f) about f=p gives KL(f||p)=(f−p)22p(1−p)+2g−16g2(1−g)2(f−p)3 for some g between f,p.

To obtain a normal approximation, we need the remainder R quantity to be small. Notice that R=|n(2g−1)6g2(1−g)2(f−p)3|≤n|f−p|36[min(f,p)min(1−f,1−p)]2.

Good news is that when I1,⋯,In∼i.i.dBernoulli(p), by SLLN, Snn=d1n(I1+⋯+In)→a.s.p.
So normal approximation is that P(Sn=k)∼12πnp(1−p)e−(k−np)22np(1−p).
This is accurate if n|f−p|36[min(f,p)min(1−f,1−p)]2≪1.
However, in general, normal approximation is less accurate than the entropy approximation.

We can also use CLT:

Theorem (CLT for Binomial Distribution)

Let Sn∼Binomial(n,p) for 0<p<1. Then ∀a,b∈R where a<b, limn→∞P(a≤nSnn−pp(1−p)≤b)=∫ab12πe−t22dt.

2.1 Application of CLT for Binomial(n,p)

Let p be the proportion of the population supporting Trump.
n be the number of people polled u.a.r from the population.
Sn be the number in the sample who support Trump.
p^n=Snn be the estimator of p.
How large should n be, s.t. P(p∈[p^n−ε,p^n+ε])≥1−α for given ε>0 and 0<α<1?

This is the #ConvidenceLevel . For each trial, p^ moves, and this moving interval should cover p at least 100(1−α)% of the time.

Denote Ij={1,j-th person polled supported Trump,0,otherwise.
Then I1,⋯,In∼i.i.dBernoulli(p),Sn=I1+⋯+In∼Binomial(n,p). By CLT, np^n−pp(1−p)→dZ∼N(0,1). so P(p^n−ε≤p≤p^n+ε)=P(−ε≤p^n−p≤ε)=P(−εnp(1−p)≤np^n−pp(1−p))≈Φ(z)−Φ(−z)=1−2Φ(−z),
so

1−2Φ(−z)≥1−α⇒Φ(−z)≤α2⇒n≥[−Φ−1(α/2)ε]2p(1−p).

Since p(1−p)≤14,∀p∈[0,1], this yields n≥14[−Φ−1(α/2)ε].

If ε=0.05,α=0.05, we have n≥385.
If ε=0.01,α=0.05, we have n≥9604.