Probability does not begin with formulas or distributions. It begins with simplified descriptions of situations in which outcomes are uncertain.
These descriptions do not attempt to capture reality in full detail. They isolate the random mechanism of interest and ignore everything else.
Such idealized descriptions are called probability models. They provide a controlled setting in which randomness can be analyzed, compared, and reused across different problems.
What is the chance of two heads in two tosses of a fair coin?
The question sounds settled before anything else is said. It is not: the answer depends on what counts as an outcome and how probability is spread over the outcomes.
Two reasonable descriptions of the same tosses give 1/3 and 1/4.
description
outcomes
probability of each
P(two heads)
count the heads
0, 1, 2
1/3
1/3
record each toss in order
HH, HT, TH, TT
1/4
1/4
The first description treats "one head" as a single outcome, though it can happen in two ways, HT and TH. Real coins agree with the second row.
A probability is only as definite as the description of outcomes, events and probabilities behind it.
The two rows describe the same situation but choose different outcomes and different probabilities. Writing those choices down explicitly is the job of a probability model.
The definition
A probability model is a mathematical description of a random situation: its sample space of possible outcomes, the events, and the probabilities assigned to them.
A probability model is a mathematical description of a random situation.
It specifies: • the set of all possible outcomes • which collections of outcomes are considered events • how probabilities are assigned to those events
So for two fair tosses, the model with sample space {HH,HT,TH,TT} and probability 41 for each outcome gives P(two heads)=41.
A model describes how the randomness is generated, not what is later measured from it.
The model defines how randomness is generated, not what is later measured from it. Questions about averages, counts, or numerical outcomes arise only after additional structure is placed on the model.
A probability model: three tosses, eight outcomes, each 1/8
This grid is a complete probability model: a list of outcomes, the events formed by grouping them, and a rule that gives each outcome the same probability 1/8. Nothing about the physical coin appears, only the random mechanism. Highlight groups of cells to form events on the coin toss sample space explorer.
A model is a choice, and a poor choice gives confident wrong answers.
Nothing inside the first row of the table looks wrong; only comparison with real tosses shows it. How models are checked against the world is taken up in models vs reality.
Why probability works through such models is the next question.
The rest of the page covers why probability uses models, what all models share, how models relate to random variables and distributions, modeling assumptions, simple discrete models and richer ones, models against reality, and common pitfalls.
Why Probability Uses Models
Real-world situations are too complex to analyze directly. Probability works by replacing reality with a simplified structure that captures only the random mechanism of interest.
Models make this possible by: • stripping away irrelevant details • making assumptions explicit • allowing repeated reasoning under controlled conditions
Any conclusion drawn in probability is therefore conditional on the model being used. The quality of a probabilistic result depends not on how realistic a situation feels, but on how appropriate the chosen model is for the question being asked.
What all Probability Models Have in Common
Probability models may look different on the surface, but they are built from the same basic ingredients.
At the core of every model are: • a list of possible outcomes the experiment may produce • collections of outcomes treated as meaningful events • numerical weights expressing how likely those events are
Nothing in this structure depends on the story behind the model. Coins, dice, cards, measurements, or simulations all fit into the same abstract framework.
This is why probability theory can move freely between different contexts: the interpretation changes, but the underlying machinery does not.
Models, Random Variables, and DistributionsLearn Math Class
Discrete uniform PMF, values 1 to 6
A random variable maps outcomes to numbers and its distribution is what remains once the outcomes themselves are forgotten: six equally likely faces become six bars of height 1/6. The distribution is a summary of the model, not the model itself. Compare the distributions of other families on the PMF visualizer.
The next two sections show that this summary loses information in both directions.
One Distribution, Many Models
RULE
A probability distribution does not uniquely identify how randomness was produced.
The same distribution may arise from: • different outcome spaces • different experimental setups • different probability assignments
Once a distribution is formed, the mechanism that generated it is no longer visible.
Distributions describe results, not processes. Models describe processes, not results.
One Model, Many Distributions
Fix a probability model (S,E,P) — the probability-space triple in its school lettering, S for the sample space and E for the events.
Different random variables can be defined on the same model:
One Model, Many Distributions
X1,X2,⋯:S→R
Each choice produces its own distribution:
One Model, Many Distributions
P(Xk∈A),A⊆R
By changing the mapping (not the model), the resulting distribution may be: • discrete • continuous • mixed
The model remains unchanged. Only the measurement changes.
Fact: One random point, endless coin tosses
The binary digits of a point chosen uniformly from [0,1] are independent fair coin tosses: each pattern of the first k digits marks off an interval of length 2k1, so every pattern is equally likely. One draw from this single model carries infinitely many independent variables.
Even a very plain model supports a great many measurements. In the model where a single point is chosen uniformly from the interval [0,1], so that every subinterval has probability equal to its length, the point itself has a continuous distribution. Reading its first decimal digit gives a discrete variable with ten equally likely values. Replacing every point below 21 by 21 gives a mixed variable, with probability 21 sitting on the single value 21 and the rest spread evenly above it.
One model, two random variables
The model is fixed: two fair dice, 36 equally likely ordered outcomes. Measure the sum and you get the blue PMF, a triangle on 2 to 12 peaked at 7. Measure the larger of the two faces instead and you get the amber PMF on 1 to 6, rising as (2k − 1)/36 up to P(max = 6) = 11/36. Nothing about the dice changed; only the function applied to each outcome did.
A third choice, such as whether the sum is even, would give a third distribution on the same model.
Modeling Assumptions
Every probability model rests on explicit choices.
Typical assumptions include: • outcomes are treated as symmetric • trials do not influence each other • the same mechanism is repeated each time • the outcome space is finite or infinite
Pitfall: Assuming independence for draws without replacement
Treating two draws without replacement as independent gives P(two aces)=(524)2=1691. After one ace, 3 of the 51 cards left are aces, so the answer is 524⋅513=2211.
These assumptions are not conclusions. They are inputs to the model.
Changing an assumption changes the model, and therefore changes all results derived from it.
Card draws show the second assumption at work. If each card is put back and the deck shuffled before the next draw, every draw faces the same 52 cards. If the card is kept, the second draw faces only 51 cards, and the deck it draws from is no longer the one the first draw saw.
A model's answers are only as trustworthy as its assumptions, and those are chosen, not proved.
Simple Discrete Models
Some probability models have a finite or countable set of outcomes.
In such models: • all outcomes can be listed one by one • events are formed by grouping outcomes • probabilities are assigned directly to each outcome
These models are often used as starting points because their structure is fully visible.
Examples of this class include: • two-outcome mechanisms • finite multi-outcome mechanisms
Concrete realizations of these appear on the following pages: • Coin Toss Model • Dice Roll Model
Pitfall: Listing the sums as equally likely outcomes
Taking the 11 sums of two dice as equally likely outcomes gives P(sum=7)=111. The equally likely outcomes are the 36 ordered pairs, 6 of which add to 7, so P(sum=7)=366=61.
In the dice model, the list of outcomes depends on what is recorded for each roll. Recording both faces in order gives the pairs from (1,1) to (6,6); recording only the total gives the numbers from 2 to 12. Both lists describe the same roll of two dice, and each total corresponds to a group of pairs, an event in the sense of the second bullet above.
Two dice: a finite model with 36 outcomes
A simple discrete model lists every outcome explicitly: 36 ordered pairs, each with probability 1/36, and every event is a group of cells whose probability is a count over 36. Everything about the model can be checked by counting. Highlight events and count their cells on the dice roll sample space explorer.
Richer models grow out of this one by repetition and dependence.
From Simple Models to Richer Ones
Simple discrete models are only the starting point.
By extending or modifying a basic model, one can describe: • repeated experiments • dependent outcomes • infinite outcome spaces • measurements that vary continuously
The underlying idea remains the same: outcomes and probabilities are fixed first, and additional structure is built on top of them.
More advanced models reuse the same foundations, but allow for greater complexity in how randomness is represented.
Models vs Reality
A probability model is not a description of the world itself.
It is a constructed object that isolates a specific random mechanism while ignoring everything else.
Because of this: • models can be useful even when they are unrealistic • realistic detail does not guarantee correctness • conclusions are valid only within the chosen model
Fact: Seven riffles to trust the card model
Four riffle shuffles cannot even reach every order: one riffle can produce at most 252 orders, and 2208 is far below 52!≈8×1067. Bayer and Diaconis showed in 1992 that the deck comes close to the equally likely model only after about 7 riffles.
Card games give a familiar example. The standard card model takes all 52! orders of a shuffled deck as equally likely, and every probability for a dealt hand rests on that assumption. A real shuffle starts from whatever order the deck was in and changes it a little at a time, so whether the model fits depends on how much shuffling was actually done.
When a model does not match the situation it is applied to, probability calculations remain correct mathematically but become irrelevant in practice.
How This Section Is Organized
This page introduces probability models at a conceptual level.
Each model is then developed on its own page, using the same internal structure: • description of the random mechanism • definition of outcomes and events • probability assignment • possible measurements and distributions
This organization allows new models to be added without changing the overall framework, and makes it easy to compare different modeling choices across examples.
Model Pages
The following pages present concrete probability models built using the framework introduced above.
Each page focuses on a single random mechanism and shows how outcomes, events, and probabilities are specified within that model.
Available models: • Coin Toss Model • Dice Roll Model
Additional models will be added over time, following the same structure and conventions.
Probability Models at a Glance
The page has built up the probability model as the foundational layer beneath random variables and distributions, distinguishing what a model specifies from what is later measured on it. The table below collects the core ideas into a single reference card, pairing each aspect of probability modeling with its concise statement and a concrete example or symbol.
a model is a constructed object isolating a mechanism, not a description of the world
a coin model is useful even if real coins are slightly biased
Probability Models at a GlanceLearn Math Class
Common Pitfalls
Each of the two mistakes below is marked where it happens in the sections above; here each is worked through in full, with the wrong and the right working side by side.
1Assuming independence for draws without replacement
Two cards are drawn from a standard deck of 52, and the first card is not put back. The assumption that trials do not influence each other is applied anyway, and the probabilities of the two draws are multiplied as if the deck were unchanged.
Wrong turn
Error here
P(two aces)=524⋅524
=270416=1691≈0.0059
Right path
After an ace is drawn, 51 cards remain and 3 of them are aces.
P(two aces)=524⋅513
=265212=2211≈0.0045
Why it happensCoin and dice models, the first ones met, do assume independent trials, and the habit of multiplying carries over to draws where the first result changes what is left.
How to checkCount outcomes directly. There are 52⋅51=2652 ordered pairs of different cards and 4⋅3=12 ordered pairs of two aces, which gives 265212=2211. In a deck of two cards with one ace, the wrong method gives 21⋅21=41 for two aces, though drawing two aces from it is impossible.
2Listing the sums as equally likely outcomes
For two dice, the totals 2 to 12 are listed as the outcomes of the model, and each total is given the same probability.
Wrong turn
Error here
There are 11 totals, each with probability 111, so
P(sum=7)=111≈0.091
Right path
The equally likely outcomes are the 36 ordered pairs of faces. Six of them add to 7: (1,6), (2,5), (3,4), (4,3), (5,2), (6,1).
P(sum=7)=366=61≈0.167
Why it happensThe question is about the total, so the totals look like the natural outcomes. The symmetry of a fair die holds for its faces, not for the totals that two dice produce.
How to checkCount how many pairs give each total. The total 2 comes only from (1,1), while 7 comes from six pairs, so the two cannot be equally likely. The counts 1,2,3,4,5,6,5,4,3,2,1 add up to 36.
Both slips put the wrong outcomes into the model. Draws without replacement are not independent, and the sums of two dice are not equally likely, because they come from different numbers of pairs. Counting the outcomes directly catches both.
Skills You Should Learn
Each skill opens its step-by-step method page.
Probability methods
Solve coin toss probability problemsWord problemsHeads and tails over several tosses: count the sequences, or use the binomial formula for a biased coin.
Can you recover the experiment from its distribution?
+
No. Once a distribution is formed, the mechanism that produced it is no longer visible. The same distribution can arise from different outcome spaces, different experimental setups and different probability assignments. Distributions record results, not processes — which is why knowing a variable is, say, uniform tells you nothing about how it was generated.Read more →
Does changing the random variable change the model?
+
No — only the measurement changes. The model is the outcome space, the events and the probability assignment; a random variable is a mapping defined on top of it. Define a different mapping and you get a different distribution while the underlying model sits untouched. Confusing the two makes distributions look like properties of experiments rather than of measurements.Read more →
Can the same model produce both discrete and continuous distributions?
+
Yes. Fix one probability model and define different random variables on it: each mapping produces its own distribution, and those can be discrete, continuous or mixed. The randomness is fixed by the model; what varies is what you choose to measure. So the discrete/continuous split describes the variable, not the experiment behind it.Read more →
Does a more realistic model give better answers?
+
Not necessarily. Realistic detail does not guarantee correctness, and deliberately unrealistic models are often the useful ones because they isolate the mechanism you care about. What matters is whether the model matches the situation you are applying it to. When it does not, the calculations stay mathematically correct and become practically irrelevant.Read more →
How many shuffles does it take to mix a deck of cards?
+
About seven riffle shuffles. Card probabilities assume all 52! orders of the deck are equally likely, but a real shuffle starts from the deck's last order and changes it a little at a time. One riffle can produce at most 252 orders, so four can reach at most 2208, far below 52!≈8×1067. Bayer and Diaconis showed in 1992 that the deck comes close to the equally likely model only after about 7 riffles.Read more →
How can one random number contain infinitely many coin tosses?
+
Through its binary digits. Pick a point uniformly from [0,1] and write it in base 2: each pattern of the first k digits marks off an interval of length 2k1, so all 2k patterns are equally likely. That is exactly what k independent fair tosses give, for every k, so the digits of a single draw behave as an endless sequence of independent tosses.Read more →