Chapter 11: The Poisson Distribution

Welcome to Last Minute Lecture.

This free chapter overview is designed to help students review and understand key concepts.

These summaries supplement, not replace, the original textbook and may not be redistributed or resold.

For complete coverage, always consult the official text.

So today's deep dive is pulling from a text that has, honestly, haunted a lot of physics students.

Oh yeah.

For sure.

Right.

It's the Introduction to Error Analysis.

And we are looking at this to answer a really bizarre paradox.

How do you actually go about measuring something that is, by its very nature, completely random?

And I don't mean a situation where your tools just aren't good enough.

I mean something that happens entirely at random.

But somehow, it still manages to have this definite, predictable average rate over time.

It's wild to think about.

It is.

Just think about, I don't know, the number of babies born in a hospital over a three -day weekend or the exact millisecond a radioactive atom just decides to decay.

Yeah.

And it really is a fascinating paradox because normally when we talk about error analysis or uncertainty in science, we're talking about our own human limitations.

Right.

Like being clumsy.

Exactly.

A ruler that isn't precise enough or a stopwatch we clicked a fraction of a second too late.

But when you're dealing with these intrinsically random events, the uncertainty isn't a mistake you made in the lab.

It is, well, it's a fundamental property of the universe itself.

Which is heavy.

And if you are listening to this right now, there is a very good chance you are a college student staring down chapter 11 on the Poisson distribution.

A very likely scenario.

Yeah.

You might be cramming for a lab or, you know, just trying to make sense of your textbook.

So just take a breath.

We've got you.

The mission of this deep dive is to literally walk through this exact chapter step by step.

It's going to be okay.

We're going to go from the basic definition of the Poisson distribution to how it transforms on a graph and all the way down to handling messy real world background radiation in your experiments.

So let's unpack this together.

That is a great place to focus because mastering the Poisson distribution, I mean, it's incredibly empowering for a student.

How so?

Well, it teaches you that uncertainty isn't just about throwing your hands up and admitting what you don't know.

It's about mathematically quantifying the randomness of reality so you can actually work with it confidently.

I love that.

So let's start with the anatomy of randomness, like the foundational concept here.

The textbook defines the Poisson distribution as the probability of getting exactly a certain number of counts.

Right.

Which the book calls nu, that's the Greek letter spelled N -U, looking like a little V.

Yes, nu.

And it looked at that count when your expected average is mu, which is another Greek letter.

Taking that a step further, mu is your expected average, you know, the baseline of what you think should happen.

Okay.

And nu is the specific physical number of events you actually observe when you run a single trial.

And there is a core equation for this.

Equation 11 .2, right?

Exactly.

Equation 11 .2 in your book.

It tells us how to calculate the exact probability of seeing that specific number of events.

Now, if you're looking at equation 11 .2 right now, it looks like a dense wall of algebra.

It really does.

We aren't going to, like,

read every single symbol out loud because that is impossible to follow in an audio format, but let's talk about what the formula is actually doing.

Good idea.

Conceptually, it's a balancing act.

It takes a mathematical constant, e, factors in your expected average, which is mu, and weighs that against the actual number of events you observed, nu.

Finally, it scales the whole thing down by dividing by nu factorial.

And if you are scribbling this down, just as a quick refresher, a factorial...

Oh, yeah, the exclamation point thing?

Right.

Written as a number followed by an exclamation point.

It just means multiplying that number by every whole number below it down to one.

Got it.

So three factorial is just three times two times one, which equals six.

Nice and easy.

Yeah.

But before we even plug numbers into that balancing act, we need to know where that expected average, you know, mu actually comes from.

Yes.

Very important.

The text makes a really crucial point here.

Mu is simply the rate at which the random events happen multiplied by the time you spend measuring them.

Exactly.

So mu equals r times t, rate times time.

Let's anchor this with the specific example the author uses.

It's called counting radioactive decays.

OK, let's visualize it.

Imagine you have a sample of radioactive thorium sitting in front of you.

Careful previous measurements from, say, other scientists.

Tell us it emits alpha particles at a rate of 1 .5 particles per minute.

OK, so that 1 .5 is our rate.

Our r.

Right.

And let's say we decide to turn on our lab detector and just count particles for exactly two minutes.

That's our time, t.

So rate times time?

Yeah, 1 .5 particles per minute multiplied by two minutes gives us an expected average.

Our mu.

Right, our mu of exactly three.

We expect an average of three particles.

But here is the million dollar question.

What is the actual probability that we turn on the machine for two minutes, sit there, and see exactly three particles?

Oh, this is where equation 11 .2 comes to the rescue.

Exactly.

We plug in mu equals three for our average and nu equals three for our observed count.

And if you run that balancing act through a calculator, the result it spits out is 0 .22 or 22 percent.

Now think about what that number actually represents in the real world.

Three is our expected average.

It is statistically the most likely outcome.

But even so, if you run this two minute experiment over and over again, you will only see exactly three particles about 22 percent of the time.

Wait, really?

Yeah.

You'll only see the quote unquote expected number roughly one in every five tries.

Oh, wow.

Which totally brings to mind the farmer analogy from the textbook's quick check section.

It's such a great visual for this.

It is a classic.

Imagine a farmer who knows his hens lay an average of exactly one egg a day.

Just because the expected average is one, it doesn't mean he won't occasionally walk into the hen house and find zero eggs or, you know, maybe find two eggs on a random Tuesday.

That analogy highlights the most important conceptual hurdle you need to clear right now.

What's that?

If the farmer finds zero eggs or if our nuclear physics student counts four particles instead of the expected three, nobody made a mistake.

It is not a counting error.

The farmer didn't lose an egg.

The variation is simply the intrinsic unavoidable randomness of the process itself.

OK, so if the farmer's egg count is naturally bouncing all over the place, how do we report a single measurement with any real confidence?

That is the big question.

If it's all random, I mean, we need a way to figure out how much variation we should actually expect.

We need to talk about standard deviation and what the book calls the square root rule.

Absolutely.

So in any statistical distribution, you want to know the standard deviation represented by the Greek letter sigma.

Yes, sigma.

It tells you how spread out your data is around the mean.

Now in most statistics classes, you have to calculate the mean and then do a whole separate totally miserable calculation to find the standard deviation.

Oh, I remember those pages of math.

Right.

But the Poisson distribution has a truly magical property, which is proven in equation 11 .8.

The standard deviation of a Poisson distribution is simply the square root of the expected mean.

Sigma equals the square root of mu.

Wow, that is incredibly elegant.

It really is.

The average itself dictates the spread of the randomness.

And this mathematically justifies that famous square root rule that you, the listener, have probably been blindly using earlier in your lab course without really knowing why.

Almost definitely.

The rule says if you make a single measurement of new random events, your best estimate for the true expected count is simply nu -nu,

and your uncertainty is just the square root of nu.

Spot on.

So your final reported answer for the count is nu plus or minus the square root of nay.

Let's see this in action using the textbook's more radioactive decays example.

Let's do it.

A student monitors a thorium sample for 30 minutes.

She counts exactly 49 alpha particles.

Okay, so applying the rule, her raw count is 49.

The uncertainty is the square root of 49, which is a nice clean 7.

So she reports the total number of particles emitted in 30 minutes as 49 plus or minus 7.

But the lab isn't over.

Now she needs to find the rate of emission in particles per minute.

Okay, hold on.

If I'm a student taking notes right now, my instinct is to take a shortcut here.

Uh -oh.

Let's hear it.

Well, if the count is 49 in 30 minutes, to find the rate per minute, can I just divide 49 by 30 to get roughly 1 .63 and then, you know, just take the square root of that new number to find my uncertainty?

That seems way faster.

You'd think so, but doing that completely breaks the math.

Oh, it does.

I'm so glad you brought that up because it is the single most common mistake made in Chapter 11.

Wow.

Okay.

You cannot take the square root of the calculated rate.

The square root rule applies only to the physical things you actually counted.

It only applies to nu.

Ah, I see.

So the magic square root only applies to the literal number of quicks you heard on the Geiger counter, or like the literal physical number of eggs the farmer's holding in his hands.

Precisely.

You must lock in the uncertainty of the raw count first.

As we said, the count is 49 plus or minus 7.

To find the rate, you have to use standard error propagation.

You divide your main value by the time interval, and then you also divide your uncertainty by that exact same time interval, assuming, of course, your stock watch measurement has no meaningful uncertainty.

Okay, let's slow down and do that math out loud for everyone listening.

Good idea.

We take the raw count, 49, and divide it by 30 minutes, which gives us a rate of about 1 .6 particles per minute.

Then we take the raw uncertainty, the 7, and we also divide it by 30 minutes, which gives us roughly 0 .2.

Exactly.

So the final mathematically correct rate is 1 .6 plus or minus 0 .2 particles per minute.

Put a massive star next to that in your notes.

Remember the golden rule?

Only ever take the square root of the raw counts.

Yes.

If you try taking the square root of a calculated rate, your entire error analysis will be wrong.

All right.

I'm feeling really good about this so far.

But I am looking ahead at the math, and I feel like I'm hitting a wall.

What's the issue?

As much as equation 11 .2 is great for small numbers, doing factorials for large numbers gets completely out of hand.

Oh, absolutely.

I mean, if I'm trying to calculate the odds of seeing exactly 72 events,

I have to figure out 72 factorial.

So my calculator is going to burst into flames.

It really will.

Standard calculators just give up.

And this is where we introduce the shapeshifter.

The shapeshifter.

The Gaussian approximation.

Let's think about how the Poisson distribution looks visually.

If you flip to figures 11 .1, 11 .2, and 11 .3 in the text, you can literally see this evolution happen.

OK, I'm picturing it.

When your expected average, mu, is very small, like 0 .75 or 3, the graph is a bar chart that is highly asymmetrical.

It's lumped up on the left near zero with a long trailing tail stretching out to the right.

It doesn't look anything like a standard bell curve.

Yeah, it looks like a lopsided staircase.

And that makes logical sense because you obviously can't have fewer than zero events, like the farmer can't find negative eggs in the hen house.

Exactly.

But theoretically, there's a tiny microscopic chance he could find 9 or 10, so the graph has to stretch infinitely to the right.

But here is the amazing shapeshifting trick.

As your expected average, mu, gets larger, the Poisson distribution starts to morph.

Really?

Yeah.

By the time mu equals 9, as seen in figure 11 .3, the curve has magically become symmetrical.

It smooths out.

In fact, it becomes practically indistinguishable from a standard Gaussian distribution, what most people just call a bell curve.

Right, right.

And specifically, it perfectly matches a Gaussian curve where the mean is mu and the standard deviation is the square root of mu.

That is incredible.

So the math actually gets easier the more data you collect.

The lopsided staircase just smooths out into a perfect tail.

It really does.

And this isn't just a neat visual trick for textbooks, you know, it is a vital mathematical shortcut.

Okay, give me an example.

The textbook is a great one.

The Gaussian approximation to a Poisson distribution.

Imagine an experiment where your expected mean, mu, is 64.

Okay, 64.

And you want to know the probability of getting exactly 72 counts.

Okay, if we stubbornly use the exact Poisson formula, equation 11 .2, we have to calculate 64 to the 72nd power divided by 72 factorial.

This is a massive unwieldy nightmare.

It is.

But if you do somehow manage to force a computer to calculate it, the exact Poisson probability is 2 .9%.

Okay, 2 .9%.

Now what if we use our new shortcut, the Gaussian approximation?

Because 64 is a fairly large number, we can just treat this as a standard bell curve with a mean of 64 and a standard deviation of 8.

Right, because the square root of 64 is 8.

Exactly.

The mathematics of a Gaussian curve are much simpler to process.

If you evaluate the standard Gaussian function for 72, you get 3 .0%.

Wait, 2 .9 % versus 3 .0%.

That is an incredibly close approximation for a fraction of the computational effort.

It is.

And if we step back and look at why this matters in a real lab setting, it's about ranges.

What do you mean by ranges?

Often, you don't just want to know the probability of exactly one single number, you want to know the probability of a whole range of numbers.

Like, what is the probability of getting 72 or more counts?

Oh, I see.

With the exact Poisson formula, you'd have to calculate that messy factorial equation for 72 and 73 and 74 all the way to infinity and add them all up.

Exactly.

It's practically impossible.

But with the Gaussian approximation, you can just use the standard Gaussian probability tables found in appendix A or B of your textbook.

Yes.

You just figure out how many standard deviations away your number is, look it up in the table, and you have your percentage instantly.

That turns an impossible infinite sum into a quick glance at the back of the book.

That's the power of the approximation.

It bridges the gap between chaotic randomness and practical workable statistics.

So how do we actually use these tables to prove someone's lab equipment is broken?

I like where your head is at.

Let's take all these statistical tools and apply them to experimental reasoning.

We're calling this the cosmic ray lie detector.

The textbook explores this beautifully with the cosmic ray counting example.

Picture an introductory physics lab where students are using a Geiger counter to detect cosmic rays hitting a specific area of their workbench.

All right, let's set up the drama.

Student A goes first.

He claims he has made careful repeated measurements all morning and he confidently asserts that the expected rate of cosmic rays is exactly nine particles per minute.

Okay.

Now comes student B to check his work.

Student B turns on the counter for one minute and she gets a count of 12.

So the question is, does student B's result of 12 cast serious doubt on student A's claim that the true average is nine?

Well, let's let the math decide.

If student A is right and the expected average, Mo, is nine, then the standard deviation is simply the square root of nine, which is three.

Correct.

Student B got 12.

12 is exactly one standard deviation away from the mean of nine because nine plus three is 12.

And in the world of statistics, being one standard deviation away is perfectly normal.

It happens all the time.

So student B's result does not contradict student A at all.

Student A's expected rate of nine is completely plausible.

But then student C steps up.

Student C is more thorough.

She decides to count particles for 10 straight minutes.

Oh, okay.

Let's scale it up.

If student A's rate of nine per minute is correct, student C should expect an average of 90 particles over those 10 minutes.

Right.

Nine times 10.

But when she checks her counter, she actually has 120 counts.

So is student A wrong now?

Let's run the numbers again.

If the expected Mo is 90, the standard deviation is the square root of 90.

The square root of 90 is approximately 9 .5.

Now student C got 120.

That is a difference of 30 counts from the expected 90.

Exactly.

If one standard deviation is 9 .5, then being 30 counts off means student C's result is more than three standard deviations away from student A's prediction.

Precisely.

And because these are large numbers, we can use our Gaussian approximation and look up three standard deviations in appendix A.

And what does the table say?

The probability of a result being more than three standard deviations away from the mean is roughly 0 .3 percent.

Wait, 0 .3 percent?

Which means student C's result isn't just slightly off.

Statistically it's a giant red flag.

Absolutely.

A 0 .3 percent chance means it is extremely improbable that both student A's expected rate is correct and student C's measurement is just a random fluctuation.

Exactly.

So either student A's initial measurement of 9 was totally wrong or student C's Geiger counter is malfunctioning and introducing systematic errors.

Or I guess maybe student A made his measurements on a day when the actual flux of cosmic rays from space was truly anomalous.

Also possible.

And this is how scientists actually reason experimentally.

The math doesn't just give you a number.

It acts as an alarm bell.

It tells us when to stop, question our assumptions, and check our equipment.

Which perfectly sets us up for the final boss of Chapter 11, subtracting a background.

Ah, yes.

The final boss.

This is the most complex procedure in the chapter because it forces you to combine everything we've learned, counting, converting to rates, and error propagation, all to isolate a weak signal hidden inside noisy data.

This is an incredibly common real world problem.

Imagine you are studying a radioactive sample with a detector.

The problem is your detector doesn't just see your sample.

It also registers background radiation, like cosmic rays hitting the building or trace amounts of radiation in the concrete walls.

That makes sense.

You have to subtract that background out.

It's like trying to measure the exact volume of a friend whispering to you while you are both standing in the middle of a rock concert.

That is a perfect analogy.

You have the total norms you're hearing, but you need to mathematically subtract the rock concert to figure out how loud the whisper is.

And here is where the mechanical application is so crucial.

You can't just subtract the total time from the background time.

To make the math work, you have to convert everything into a standardized rate first.

That standardization is the key to the whole process.

The textbook outlines a strict three -step procedure to handle this.

Lay on me.

Step one, you measure the total events, so source plus background over a certain time, and calculate a total standardized rate.

Step two, you completely remove your radioactive source, measure just the background events over a different time, and calculate a background rate.

And step three, you subtract the background rate from the total rate to find the pure source rate.

Let's walk through the book's final example, radioactive decays with a background.

If you are taking notes, follow along carefully.

A student puts her sample in a liquid scintillation detector for 10 minutes.

She registers 2 ,540 total counts.

Then she removes the sample, runs the detector for three minutes, and gets 95 background counts.

Okay, let's take it slow.

Step one, total rate.

The raw count is 2540.

The uncertainty is the square root of 2540, which is about 50.

Correct.

So her total count is 2540 plus or minus 50.

And then we divide both of those numbers by the 10 minutes to get the standardized rate.

Let's do that division.

Yeah.

2540 divided by 10 is 254, and 50 divided by 10 is 5, so the total rate is 254 plus or minus 5 counts per minute.

Perfect.

Step two, background rate.

Raw count is 95.

Uncertainty is the square root of 95, roughly 10.

Okay.

So the count is 95 plus or minus 10, divide by the three minutes.

95 divided by 3 is roughly 32, 10 divided by 3 is roughly 3.

So the background rate is 32 plus or minus 3 counts per minute.

Notice we followed your golden rule.

We did.

We only took the square root of the raw counts, the 2540 and the 95, before doing any of that division.

I'm so glad you caught that.

Never take the square root of the rate.

Now, step three, subtracting to find the source rate.

Right.

We subtract the rates.

254 minus 32 gives us 222 counts per minute for the source alone.

Okay, but here's where my brain breaks a little.

How do we handle the uncertainties?

We have a plus or minus 5 from the total and a plus or minus 3 from the background.

Do I just add them together to get 8?

You'd think so, but no.

We don't just add them linearly.

Because these are independent and random uncertainties, we have to combine them in what's called quadrature.

Why quadrature?

I mean, why can't I just add 5 and 3?

Because random errors don't always push in the exact same direction.

Sometimes one error makes your result a little too high, while the other error makes it a little too low, and they partially cancel each other out.

That makes sense.

Adding them in quadrature, which is mathematically similar to using the Pythagorean theorem for a right triangle, accounts for this random multidirectional nature.

Okay.

How does that work mechanically?

You square both uncertainties, add those squares together, and then take the square root of the total.

Okay, that makes so much more sense.

Let's do that math.

5 squared is 25, 3 squared is 9, 25 plus 9 is 34, and the square root of 34 is approximately 6.

Exactly.

So our final combined uncertainty is plus or minus 6.

Meaning the final definitive answer for the activity of the radioactive source alone is 222 plus or minus 6 counts per minute.

You nailed it.

It is incredibly satisfying to see all those rules come together to solve a messy real world problem.

It really is.

We started with a basic formula for predicting random events, the Poisson equation.

We discovered that the uncertainty is magically just the square root of the count.

We learned that if we gather enough data, the lot -sided staircase morphs into a smooth, predictable Gaussian bell curve.

And finally, we used all of that to standardize our rates, strip away background noise, and find the true signal hidden underneath.

It's a complete toolkit for dealing with a random universe.

It allows you to look at a chaotic system and extract precise mathematically rigorous truth from it.

We spent so much time in science labs trying to eliminate uncertainty.

We calibrate our tools, we stabilize our tables, we try to make everything perfect.

Yeah, we do.

But the Poisson distribution shows us that true randomness has its own perfect mathematical fingerprint.

In fact, if your lab data is ever too perfect.

If it doesn't show that natural square root of must spread that we've been talking about today,

that is when you should really be suspicious.

That is a profound takeaway.

If the randomness isn't there, you aren't looking at reality.

Keep that in mind next time you are analyzing your data.

A warm thank you for listening from the Last Minute Lecture Team.