Exponential Distribution

原文发布在我的个人小站：here

Story

The Exponential distribution is the continuous counterpart to the [Geometric distribution](file://./Geometric-Distribution.md). The story of the Exponential distribution is analogous, but we are now waiting for a success in continuous time, where successes arrive at a rate of (lambda) successes per unit of time. The average number of successes in a time interval of length (t) is (lambda t), though the actual number of successes varies randomly. An Exponential random variable represents the waiting time until the first arrival of a success.

——adapted from Book BH

Basic

Definition: A continuous r.v. (X) is said to have the Exponential distribution with parameter (lambda) if its PDF is

[f(x) = lambda e^{-lambda x}, quad x > 0 ]

The corresponding CDF is

[F(x) = 1 - e^{-lambda x}, quad x > 0 ]

To calculate the expectation and variance, we first consider (X sim Exp(1)) with PDF (f(x) = e^{-x}), then

[egin{split} E(X) &= int_0^{infty} x e^{-x} dx = 1 \ E(X^2) &= int_0^{infty} x^2 e^{-x} dx \ &= -x^2e^{-x}|_0^{infty} + 2int_0^{infty} x e^{-x} dx \ &= 2E(X) = 2 \ Var(X) &= E(X^2) - E^2(X) = 2-1 = 1 \ M_X(t) &= E(e^{tX}) = int_0^{infty} e^{tx} e^{-x} dx \ &= int_0^{infty} e^{-(1-t)x} dx = frac{1}{1-t} quad ext{for }t<1 end{split} ]

Now let (Y=frac{X}{lambda} sim Exp(lambda)) for

[f_Y(y) = f_X(X(y))frac{dx}{dy} = e^{-lambda y}cdotlambda sim Exp(lambda) ]

[P(Yle y) = P(Xle lambda y) = 1 - e^{-lambda y} sim Exp(lambda). ]

Hence, we can get

(E(Y) = E(X/lambda) = 1/lambda)
(Var(Y) = Var(X/lambda) = 1/lambda^2)
MGF (moment generating function):

[egin{split} M_Y(t) &= E(e^{tY}) =E(e^{tX/lambda}) \ &= E(e^{frac{t}{lambda}X}) = M_X(frac{t}{lambda}) = frac{1}{1-t/lambda} \ &= frac{lambda}{lambda -t} quad ext{for }t<lambda end{split} ]

Memeoryless Property

Memoryless is something like (P(X ge s+t ~|~ X ge s) = P(X ge t)), let (X sim Exp(lambda)), then

[egin{split} P(X ge s+t ~|~ X ge s) &= frac{P(X ge s+t, ~X ge s)}{P(X ge s)} \ &= frac{P(X ge s+t)}{P(X ge s)} \ &= frac{e^{-lambda (s+t)}}{e^{-lambda s}} = e^{-lambda t} \ &= P(X ge t) end{split} ]

Theorem: If (X) is a positive continuous r.v. with memoryless property, then (X) has an exponential distribution. Similarly, if (X) is discrete, then it has a geometric distribution.

Proof idea: use survival function and solve differential equations.

Examples

eg.1 (X_1 sim Exp(lambda_1), ~X_2 sim Exp(lambda_2)), and (X_1 perp X_2). Then (P(X_1 < X_2) = frac{lambda_1}{lambda_1 + lambda_2}).

Proof: By LOTP (law of total probability),

[egin{split} P(X_1 < X_2) &= int_0^{infty} f_{X_1}(x) P(X_2 > X_1 ~|~ X_1=x) dx \ &= int_0^{infty} f_{X_1}(x) P(X_2 > x ~|~ X_1=x) dx \ &= int_0^{infty} f_{X_1}(x) P(X_2 > x) dx quad ext{(independence)} \ &= int_0^{infty} lambda_1 e^{-lambda_1 x} e^{-lambda_2 x} dx \ &= lambda_1 int_0^{infty} e^{-(lambda_1 + lambda_2) x} dx \ &= frac{lambda_1}{lambda_1 + lambda_2} end{split} ]

eg.2 ({X_i}_{i=1}^n) are independent with (X_j sim Exp(lambda_j)). Let (L = min(X_1, cdots, X_n)), then (L sim Exp(lambda_1 + cdots lambda_n)).

Proof:

[egin{split} P(L > t) &= Pleft(min(X_1,cdots,X_n) > t ight) \ &= P(X_1 > t, cdots, X_n >t) \ &= P(X_1 > t) cdots P(X_n >t) quad ext{indep.} \ &= e^{-lambda_1 t}cdots e^{-lambda_n t} \ &= e^{-(lambda_1 + cdots lambda_n)t} sim Expleft(sum_j lambda_j ight) end{split} ]

The intuition of this result is that if you consider (n) Poisson processes with rate (lambda_j),

(X_1) as the waiting time for a green car
(X_2) as the waiting time for a red car
...

Then (L) is the waiting time for a car of any color (i.e., any car). So it makes sense, the rate is (lambda_1 + cdots + lambda_n).

eg.3 (Difference of two exponetial) Let (X sim Exp(lambda)) and (Y sim Exp(mu)), (X perp Y). Then what is the PDF of (Z=X-Y)?

Solution:
Recall the story of exponential, one can think of (X) and (Y) as waiting times for two independent things. For example,

(X) as the waiting time for a red car passing by
(Y) as the waiting time for a blue car

If we see a blue car passing by, then the further waiting time for a red car is still distributed as same distribution as (Y), for the memoryless property of exponential. Likewise, if we see a red car passing by, then the further waiting time is distributed as same as (X). The further waiting time is somehow what we are interested in, say (Z).

The above intuition says that, the conditional distribution of (X-Y) given (X > Y) is the distribution of (X), and the conditional distribution of (X-Y) given (X le Y) is the distribution of (-Y) (or in other words, the conditional distribution of (Y-X) given (Y ge X) is same as the distribution of (Y)).

To make full use of our intuition, we know that

If (X>Y), which means (Z>0), then (Z~|~X>Y = X) a.s. holds, that is

[egin{gathered} f_Z(z~|~X>Y) = lambda e^{-lambda z} \ ext{and since }P(X<Y) = 0 \ implies f_Z(z) = f_Z(z~|~X>Y)P(X>Y) \ = frac{mu}{lambda + mu}lambda e^{-lambda z}. end{gathered} ]

If (X < Y), which means (Z < 0), then (Z~|~X<Y = -Y) a.s. holds, that is

[egin{gathered} f_Z(z~|~X<Y) = f_Y(y(z))left|frac{dy}{dz} ight| = mu e^{mu z} \ implies f_Z(z) = f_Z(z~|~X<Y)P(X<Y) \ = frac{lambda}{lambda + mu} mu e^{mu z} end{gathered} ]
However, this is just a sketch. Later we will see how to derivate the form mathematically.

From the above point of view, the PDF of (Z) had better be discussed by the sign of (Z).

If (Z > 0), which implies $X > Yimplies P(X < Y) = 0 $, then

[egin{split} P(Z > z) &= P(X-Y>z ~|~ X>Y)P(X>Y) + P(Z>z~|~X<Y)P(X<Y) \ &= P(X>z)P(X>Y) + 0 quad ext{(memoryless)} \ &= frac{mu}{lambda + mu} e^{-lambda z} quad ext{(by eg.1)} \ implies f_Z(z) &= frac{lambdamu}{lambda + mu} e^{-lambda z} quad ext{for }z>0 end{split} ]

If (Z le 0), which implies (X le Y), then

[egin{split} P(Z < z) &= P(Z<z ~|~ X>Y)P(X>Y) + P(X-Y<z~|~X<Y)P(X<Y) \ &= 0 + P(Y-X > -z ~|~ Y>X)P(Y>X) \ &= P(Y>X)P(Y > -z) quad ext{(memoryless)} \ &= frac{lambda}{lambda + mu}e^{mu z} quad ext{(by eg.1)} \ implies f_Z(z) &= frac{lambdamu}{lambda + mu}e^{mu z} quad ext{for }z<0 end{split} ]

Therefore, the PDF of (Z) has the form

[f_Z(z) = frac{lambdamu}{lambda + mu} egin{cases} e^{-lambda z} &quad z>0 \ e^{mu z} &quad z<0 end{cases} ]

Note: (P(X=Y)=0) since the integral domain is a line ((y=x)) whose measure is 0. That is (P(Z=0) = 0). This is why we can give no care of the case (X=Y).