Poisson binomial distribution explained

In probability theory and statistics, the Poisson binomial distribution is the discrete probability distribution of a sum of independent Bernoulli trials that are not necessarily identically distributed. The concept is named after Siméon Denis Poisson.

p1,p2,...,pn

. The ordinary binomial distribution is a special case of the Poisson binomial distribution, when all success probabilities are the same, that is

p1=p2==pn

.

Definitions

Probability Mass Function

The probability of having k successful trials out of a total of n can be written as the sum[1]

\Pr(K=k)=

\sum\limits
A\inFk

\prod\limitsi\inpi

\prod\limits
j\inAc

(1-pj)

where

Fk

is the set of all subsets of k integers that can be selected from

\{1,2,3,...,n\}

. For example, if n = 3, then

F2=\left\{\{1,2\},\{1,3\},\{2,3\}\right\}

.

Ac

is the complement of

A

, i.e.

Ac=\{1,2,3,...,n\}\setminusA

.

Fk

will contain

n!/((n-k)!k!)

elements, the sum over which is infeasible to compute in practice unless the number of trials n is small (e.g. if n = 30,

F15

contains over 1020 elements). However, there are other, more efficient ways to calculate

\Pr(K=k)

.

As long as none of the success probabilities are equal to one, one can calculate the probability of k successes using the recursive formula [2] [3]

\Pr(K=k)=

n
\begin{cases} \prod\limits
i=1

(1-pi)&k=0\

1
k
k
\sum\limits
i=1

(-1)i-1\Pr(K=k-i)T(i)&k>0\ \end{cases}

where

n
T(i)=\sum\limits
j=1

\left(

pj
1-pj

\right)i.

The recursive formula is not numerically stable, and should be avoided if

n

is greater than approximately 20.

An alternative is to use a divide-and-conquer algorithm: if we assume

n=2b

is a power of two, denoting by

f(pi:j)

the Poisson binomial of

pi,...,pj

and

*

the convolution operator, we have
f(p
1:2b

)=

f(p
1:2b-1
)*f(p
2b-1+1:2b

)

.

More generally, the probability mass function of a Poisson binomial can be expressed as the convolution of the vectors

P0,...,Pn

where

Pi=[1-pi,pi]

. This observation leads to the Direct Convolution (DC) algorithm for computing

\Pr(K=0)

through

\Pr(K=n)

: // PMF and nextPMF begin at index 0 function DC(

p1,...,pn

) is declare new PMF array of size 1 PMF[0] = [1] for i = 1 to

n

do declare new nextPMF array of size i + 1 nextPMF[0] = (1 -

pi

) * PMF[0] nextPMF[i] =

pi

* PMF[i - 1] for k = 1 to i - 1 do nextPMF[k] =

pi

* PMF[k - 1] + (1 -

pi

) * PMF[k] repeat PMF = nextPMF repeat return PMF end function

\Pr(K=k)

will be found in PMF[k]. DC is numerically stable, exact, and, when implemented as a software routine, exceptionally fast for

n\leq2000

. It can also be quite fast for larger

n

, depending on the distribution of the

pi

.

Another possibility is using the discrete Fourier transform.[4]

\Pr(K=k)=

1
n+1
n
\sum\limits
l=0

C-lk

n
\prod\limits
m=1

\left(1+(Cl-1)pm\right)

where

C=\exp\left(

2i\pi
n+1

\right)

and

i=\sqrt{-1}

.

Still other methods are described in "Statistical Applications of the Poisson-Binomial and conditional Bernoulli distributions" by Chen and Liu[5] and in "A simple and fast method for computing the Poisson binomial distribution function" by Biscarri et al.[6]

Cumulative distribution function

The cumulative distribution function (CDF) can be expressed as:

\Pr(K\leqk)=

k
\sum
l=0
\sum\limits
A\inFl

\prod\limitsi\inpi

\prod\limits
j\inAc

(1-pj)

,

where

Fl

is the set of all subsets of size

l

that can be selected from

\{1,2,3,...,n\}

.

It can be computed by invoking the DC function above, and then adding elements

0

through

k

of the returned PMF array.

Properties

Mean and Variance

Since a Poisson binomial distributed variable is a sum of n independent Bernoulli distributed variables, its mean and variance will simply be sums of the mean and variance of the n Bernoulli distributions:

\mu=

n
\sum\limits
i=1

pi

\sigma2

n
=\sum\limits
i=1

(1-pi)pi

Entropy

There is no simple formula for the entropy of a Poisson binomial distribution, but the entropy is bounded above by the entropy of a binomial distribution with the same number parameter and the same mean. Therefore, the entropy is also bounded above by the entropy of a Poisson distribution with the same mean.[7]

The Shepp–Olkin concavity conjecture, due to Lawrence Shepp and Ingram Olkin in 1981, states that the entropy of a Poisson binomial distribution is a concave function of the success probabilities

p1,p2,...,pn

.[8] This conjecture was proved by Erwan Hillion and Oliver Johnson in 2015.[9] The Shepp–Olkin monotonicity conjecture, also from the same 1981 paper, is that the entropy is monotone increasing in

pi

, if all

pi\leq1/2

. This conjecture was also proved by Hillion and Johnson, in 2019.[10]

Chernoff bound

The probability that a Poisson binomial distribution gets large, can be bounded using its moment generating function as follows (valid when

s\geq\mu

and for any

t>0

):

\begin{align} \Pr[S\ges] &\le\exp(-st)\operatornameE\left[\exp\left[t\sumiXi\right]\right] \\&=\exp(-st)\prodi

t
(1-p
i+e

pi) \\&=\exp\left(-st+\sumilog\left(pi(et-1)+1\right)\right) \\&\le\exp\left(-st+\sumilog\left(

t-1))\right)\right) \\&=
\exp(p
i(e

\exp\left(-st+\sumi

t-1)\right) \\&=
p\exp\left(s-\mu-slog
i(e
s
\mu

\right), \end{align}

where we took t=\log\left(s/\mu\right). This is similar to the tail bounds of a binomial distribution.

Approximation by Binomial Distribution

A Poisson binomial distribution

PB

can be approximated by a binomial distribution

B

where

\mu

, the mean of the

pi

, is the success probability of

B

. The variances of

PB

and

B

are related by the formula
n
Var(PB)=Var(B)-style\sum
i=1
2
\displaystyle(p
i-\mu)

As can be seen, the closer the

pi

are to

\mu

, that is, the more the

pi

tend to homogeneity, the larger

PB

's variance. When all the

pi

are equal to

\mu

,

PB

becomes

B

,

Var(PB)=Var(B)

, and the variance is at its maximum.

Ehm has determined bounds for the total variation distance of

PB

and

B

, in effect providing bounds on the error introduced when approximating

PB

with

B

. Let

\nu=1-\mu

and

d(PB,B)

be the total variation distance of

PB

and

B

. Then

d(PB,B)\le(1-\mun+1-\nun+1)

n
\sum
2
\displaystyle(p
i-\mu)
i=1
((n+1)\mu\nu)

d(PB,B)\geCmin(1,

1
n\mu\nu
n
)style\sum
i=1
2
\displaystyle(p
i-\mu)

where

C\ge1
124
.

d(PB,B)

tends to 0 if and only if

Var(PB)/Var(B)

tends to 1.[11]

Approximation by Poisson Distribution

A Poisson binomial distribution

PB

can also be approximated by a Poisson distribution

Po

with mean
n
λ=\sum
i=1

\displaystylepi

. Barbour and Hall have shown that
1min(
32
1
λ
n
,1)style\sum
i=1

\displaystyle

2\le
p
i

d(PB,Po)\le

1-\epsilon
λ
n
\sum
i=1

\displaystyle

2
p
i

where

d(PB,B)

is the total variation distance of

PB

and

Po

.[12] It can be seen that the smaller the

pi

, the better

Po

approximates

PB

.

As

n
Var(Po)=λ=\sum
i=1

\displaystylepi

and

Var(PB)

n
=\sum\limits
i=1

pi-\sum\limits

n
i=1
2
p
i
,

Var(Po)>Var(PB)

; so a Poisson binomial distribution's variance is bounded above by a Poisson distribution with
n
λ=\sum
i=1

\displaystylepi

, and the smaller the

pi

, the closer

Var(Po)

will be to

Var(PB)

.

Computational methods

The reference [13] discusses techniques of evaluating the probability mass function of the Poisson binomial distribution. The following software implementations are based on it:

See also

References

  1. 3. 2. 295–312. Wang. Y. H.. On the number of successes in independent trials. Statistica Sinica. 1993.
  2. 27. 3. 123–124. Shah. B. K.. On the distribution of the sum of independent integer valued random variables. American Statistician. 1994. 2683639.
  3. 81. 3. 457. Chen. X. H.. A. P. Dempster . J. S. Liu . Weighted finite population sampling to maximize entropy. Biometrika. 1994. 10.1093/biomet/81.3.457.
  4. 46. 2. 803–817. Fernandez. M.. S. Williams. Closed-Form Expression for the Poisson-Binomial Probability Density Function. IEEE Transactions on Aerospace and Electronic Systems. 2010. 10.1109/TAES.2010.5461658. 2010ITAES..46..803F. 1456258.
  5. 7. 875–892. Chen. S. X.. J. S. Liu. Statistical Applications of the Poisson-Binomial and conditional Bernoulli distributions. Statistica Sinica. 1997.
  6. Biscarri . William . Zhao . Sihai Dave . Brunner . Robert J. . 2018-06-01 . A simple and fast method for computing the Poisson binomial distribution function . Computational Statistics & Data Analysis . 122 . 92–100 . 10.1016/j.csda.2018.01.007 . 0167-9473.
  7. 47 . 5. 2039–2041. Harremoës. P.. Binomial and Poisson distributions as maximum entropy distributions. IEEE Transactions on Information Theory. 2001. 10.1109/18.930936.
  8. Encyclopedia: J.. Gani. V.K.. Rohatgi. Shepp. Lawrence. Olkin. Ingram. Entropy of the sum of independent Bernoulli random variables and of the multinomial distribution. Contributions to probability: A collection of papers dedicated to Eugene Lukacs. 201–206. Academic Press. New York. 1981. 0618689. 0-12-274460-8 .
  9. Hillion. Erwan. Johnson. Oliver. 2015-03-05. A proof of the Shepp–Olkin entropy concavity conjecture . Bernoulli. 23. 4B. 3638–3649 . 1503.01570 . 10.3150/16-BEJ860. 8358662.
  10. Hillion. Erwan. Johnson. Oliver. 2019-11-09. A proof of the Shepp–Olkin entropy monotonicity conjecture . Electronic Journal of Probability. 24 . 126 . 1–14 . 10.1214/19-EJP380. free. 1810.09791.
  11. Ehm . Werner . 1991-01-01 . Binomial approximation to the Poisson binomial distribution . Statistics & Probability Letters . 11 . 1 . 7–16 . 10.1016/0167-7152(91)90170-V . 0167-7152.
  12. Web site: Barbour . A.D. . Hall . Peter . 1984 . On the Rate of Poisson Convergence . Zurich Open Repository andArchive . Mathematical Proceedings of the Cambridge Philosophical Society, 95(3) . 473–480.
  13. Hong . Yili . On computing the distribution function for the Poisson binomial distribution . Computational Statistics & Data Analysis . March 2013 . 59 . 41–51 . 10.1016/j.csda.2012.10.006.