6
$\begingroup$

I'm trying to understand $\lambda$'s role in both the Poisson and Exponential Distributions and how it is used to find probabilities (yes, I have read the other post regarding this topic, didn't quite do it for me).

What (I think) I understand:

  1. Poisson Distribution -

    • discrete

    • $\lambda$ is defined as the average number of successes (however "success" is defined, given problem context) per unit of time or space

    • PMF: $~~P(X = k;\lambda) = \frac{ \lambda^ke^{-\lambda} }{k!} $

    • $P(X\leq k) = P(X = 0)~+~P(X = 1)~+~P(X = 2)~+~\ldots~+~P(X = k)$

    • $P(X< k) = P(X = 0)~+~P(X = 1)~+~P(X = 2)~+~\ldots~+~P(X = k~-~1)$

    • $P(X\geq k) = 1~-~P(X<k)$

    • $P(X > k) = 1~-~P(X\leq k)$

  2. Exponential Distribution -

    • continuous

    • $\lambda$ is defined as the average time/space between events (successes) that follow a Poisson Distribution

    • Where my understanding begins to fade:

    • PDF: $~~f(x; \lambda)~=~ \lambda e^{-\lambda x} $

    • CDF: $ P(X \leq k; \lambda)~=~1~-~e^{-\lambda x} $

    • $ P(X > k; \lambda) ~=~ 1 ~-~P(X \leq k; \lambda)~=~e^{-\lambda x}$

Where I think the misunderstanding lies:

As of now I'm assuming $\lambda$ can be interchanged between the two distributions. Is this the case? I have briefly read about, "re-parameterizing" and I think that might be the key, but I don't know what that process is referring to. How do I do this and how does it affect the PMF and CDF of the exponential distribution?

This all came from a problem asking: Given a random variable X that follows an exponential distribution with lambda = 3, find P(X > 8). My approach was $ e^{-3*8} $, which gives a probability that seems far too low.

$\endgroup$
4
$\begingroup$

Suppose I am waiting for a bus at a stop. And suppose that a bus usually arrives at the stop in every 10 mins. Now I define λ to be the rate of arrival of a bus per minute. So, λ = (1/10).

Now I want to compute the probability that no bus arrives in the next minute. I can do it using both the poisson and the exponential distribution.

Poisson

λ = 1/10

Probability of 0 arrivals in the next minute: P(X = 0) = 0.9048

Exponential

λ = 1/10

Probability of my having to wait for more than 1 minute: P(X>1) = 0.9048

Note: Look at the expected values of both the distributions. For Poisson, we get that the average number of buses arriving per minute E(X) = λ = 0.10 buses. For exponential, the average waiting time for a bus to arrive E(X) = (1/λ) = 10 mins

$\endgroup$
5
$\begingroup$

The fact that both of these distributions use the same parameter is probably a coincidence stemming from notational convention. After all, one random variable is discrete, and the other is continuous. They should never model the exact same thing.

However, sometimes it isn't a coincidence. One instance where these parameters can mean similar things, and where you can use either of these distributions in the same modeling task is when you are using a Poisson process. This would be useful in a situation where you're modeling things arriving randomly in time (e.g. you receiving texts on your cell phone). Say you start measuring at time $0$. Fixing any time $t > 0$, you could call the total received by this time $N_t$ and assume $$ N_t \sim \text{Poisson}(\lambda t); $$ $\lambda$ here is a rate measured in calls per unit time.

You could also look at the waiting times between texts $X_1, X_2, \ldots$ and assume that $$ X_i \overset{iid}{\sim} \text{Exponential}(\lambda); $$ here $\lambda$ is also a rate (if you're writing the density the way you are in this question). The arrival times of each text message are the partial sums $S_n = \sum_{i=1}^n X_i$.

The relationship between these two distributions in this particular situation is as follows: given an integer $n$, and a time $t$, $$ P(S_n \le t) = P(N_t \ge n). $$ In the special case where $n = t = 1$, both of these expressions are equal to $1 - e^{-\lambda}$.

$\endgroup$
  • $\begingroup$ So can $\lambda$ be interchanged between the two distributions? Reason I ask is because, from this source, there is a proposition that reads: "The number of occurrences of an event within a unit of time has a Poisson distribution with parameter $\lambda $ if the time elapsed between two successive occurrences of the event has an exponential distribution with parameter $\lambda $ and it is independent of previous occurrences," This makes me think I can interchange $\lambda$ between the two equations. $\endgroup$ – Marshall McQuillen Jan 29 '18 at 12:52
  • $\begingroup$ @MarshallMcQuillen this is what I’ve answered. You cannot “interchange” lambda in the case of a Poisson process, but the two random variables share a formula for different probabilities $\endgroup$ – Taylor Jan 29 '18 at 14:12
3
$\begingroup$

The lambdas are interchangeable in certain contexts. Suppose that I'm measuring radioactivity with a Geyger counter. In one case, $\lambda=2$ means that on average I get 2 clicks per second, and the average time between clicks is $1/2$ seconds. The number of clicks per second is from a Poisson distribution, and the time between clicks is from an Exponential distribution, with both of these having $\lambda=2$.

$\endgroup$
1
$\begingroup$

As stated in Taylor's answer, there is a connection between the exponential distribution and the Poisson distribution through the equivalence of certain probability statements relating these in a Poisson process. Mathematically, this equivalence comes from a recurrence property of the incomplete gamma function, which can be used to show the probabilistic equivalence at issue.

In a Poisson process we have events occurring at some specified rate $\lambda > 0$ and we can analyse the process by looking either at the time between events, or the number of events in a given time. To do the former, let $X_1, X_2, X_3, ... \sim \text{IID Exp} (\lambda)$ be the time between events in the process, and define the partial sums $S_n \equiv \sum_{i=1}^n X_i$, which represent the time taken for the first $n$ events. Then we have $S_n \sim \text{Ga} (n, \lambda)$ so that:

$$\begin{equation} \begin{aligned} \mathbb{P}(S_n \leqslant t) = 1 - \mathbb{P}(S_n >t) &= 1 - \int\limits_t^{\infty} \text{Ga} (s|n, \lambda) ds \\ &= 1-\frac{\lambda^n}{\Gamma (n) } \int\limits_t^{\infty} s^{n-1} \exp (- \lambda s) ds \\ &= 1-\frac{1}{\Gamma (n)} \int\limits_t^{\infty} (\lambda s) ^{n-1} \exp (- \lambda s) \lambda ds \\ &= 1-\frac{1}{\Gamma (n)} \int\limits_{\lambda t}^{\infty} r^{n-1} \exp \left( - r \right) dr \\ &= 1-\frac{\Gamma(n, \lambda t)}{\Gamma (n)}. \\ \end{aligned} \end{equation}$$

Using integration by parts, the upper incomplete gamma function follows the recurrence:

$$\begin{matrix} \Gamma (n, x) = (n-1) \Gamma(n-1, x) + x^{n-1} \exp (-x) & & \Gamma (1, x) = \exp(-x). \end{matrix}$$

For integer $n$, repeated application of this recurrence yields:

$$\Gamma (n, x) = \Gamma (n) \exp (-x) \sum_{k=0}^{n-1} \frac{x^k}{k!}.$$

So, letting $N_t \sim \text{Pois} (\lambda t)$ we have:

$$\begin{equation} \begin{aligned} \mathbb{P}(S_n \leqslant t) &= 1- \exp (-\lambda t) \sum_{k=0}^{n-1} \frac{(\lambda t)^k}{k!} \\ &= 1 - \sum_{k=0}^{n-1} \text{Pois} (k|\lambda t) \\ &= \sum_{k=n}^{\infty} \text{Pois} (k|\lambda t) \\ &= \mathbb{P} (N_t \geqslant n). \end{aligned} \end{equation}$$

This gives us a basic intuitive result for the Poisson process. If the time taken for the first $n$ events is no greater than $t$ then the number of events that have occurred at time $t$ is at least $n$. If the time between events follows an exponential distribution then the number of events at a given time follows a Poisson distribution.

$\endgroup$
1
$\begingroup$

Does this, from Wikipedia sum up the relationship in a simple way?

If for every t > 0 the number of arrivals in the time interval [0, t] follows the Poisson distribution with mean λt, then the sequence of inter-arrival times are independent and identically distributed exponential random variables having mean 1/λ.

Reference: S. M. Ross (2007). Introduction to Probability Models (ninth ed.). Boston: Academic Press. ISBN 978-0-12-598062-3. pp. 307–308.

$\endgroup$

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.