Theoretical vs. Experimental Probability

A probability model might say a fair coin has a one-half chance of landing heads. If you flip it 10 times, you may observe 7 heads, so the experimental probability is 7/10. The difference does not by itself show that the coin is unfair. One number comes from an assumption-based model; the other comes from a particular run of observations.

Theoretical probability starts with a model

Theoretical probability is determined by reasoning about the possible outcomes under stated assumptions. When outcomes are equally likely, use P(A) = number of outcomes in A / total number of outcomes. A fair six-sided die has six equally likely faces. The probability of rolling a number greater than 4 is 2/6, or 1/3, because outcomes 5 and 6 meet the condition.

The assumptions are part of the answer. A die described as fair is modeled with equally likely faces. A coin described as fair is modeled with two equally likely sides. If the object is weighted, damaged, or otherwise not well represented by the model, the theoretical value may not describe its real behavior. Theoretical does not mean guaranteed; it means calculated from the chosen probability model.

Experimental probability comes from observed trials

Experimental probability, also called empirical probability, uses collected results. For an event A, calculate relative frequency: P̂(A) = number of observed occurrences of A / total number of trials. If a die is rolled 60 times and a 6 appears 13 times, the experimental probability of a 6 is 13/60, approximately 0.217 or 21.7%. The theoretical probability for a fair die is 1/6, approximately 0.167 or 16.7%.

The hat over P is sometimes used to distinguish an estimate based on data from the model probability. Your course or exam may simply call both values probabilities, so read the wording carefully. A prompt that gives trial results is asking you to use the observed counts. A prompt that describes equally likely outcomes is usually asking for a theoretical calculation.

Worked comparison: a six-sided die

Suppose 60 rolls produce these results: face 1 appears 8 times, face 2 appears 9 times, face 3 appears 12 times, face 4 appears 8 times, face 5 appears 10 times, and face 6 appears 13 times. The counts total 60. The experimental probability of rolling an even number is (9 + 8 + 13)/60 = 30/60 = 1/2. The theoretical probability is 3/6 = 1/2, because the even faces are 2, 4, and 6.

Here the values happen to match. For rolling a 6 alone, however, the observed fraction is 13/60, while the theoretical probability is 1/6. A match for one event does not force a match for every event, and a mismatch does not automatically invalidate the model. Experimental values depend on the particular sequence of outcomes.

Why observed results fluctuate

Randomness permits variation. If a fair coin is flipped twice, the outcomes could be HH, HT, TH, or TT under the usual independent-trial model. One of those four equally likely sequences contains two heads; two contain exactly one head; one contains no heads. A short experiment can therefore look far from an even split without contradicting the fair-coin model.

As the number of independent repetitions grows, the relative frequency of an event tends to settle near its theoretical probability. This long-run behavior is often described by the law of large numbers. It does not say that every new trial makes the observed fraction move closer. After a run with many heads, a tail is not “due” on the next fair flip, and one result does not change the coin’s modeled chance on the next trial.

Also separate absolute count from relative frequency. If a fair coin is flipped 20 times, the expected number of heads under the model is 10, but exactly 10 is not promised. With 2,000 flips, the count will usually be much larger than 10, while the proportion of heads is expected to be closer to 1/2 than it typically is in a very short run. The claim concerns the proportion over repeated trials, not a fixed count.

The sample size and the quality of the trials both matter

A larger number of trials generally reduces the amount of random fluctuation in relative frequency, assuming the process remains stable and the trials meet the model’s conditions. It cannot repair biased data collection. If a survey only asks people at one location, many more responses from that same location may produce a precise description of that group while still failing to represent a broader population.

For an experiment, ask whether the same process was used each time, whether outcomes were recorded consistently, and whether one trial affected another. For a survey or observational data set, ask who could be selected and who actually responded. A large convenience sample is not automatically representative. The probability formula is simple; deciding whether the trials support the intended conclusion requires attention to how the data were generated.

When each approach is useful

A theoretical model is useful when outcomes and assumptions can be described well enough to calculate a probability. It lets you make predictions before collecting data, compare possible outcomes, and solve questions where all relevant outcomes are known. A standard deck problem, for example, often assumes a properly shuffled deck and draws from a known number of cards.

Experimental probability is useful when observations are available or when the actual process is too complicated to model exactly. A company can estimate a defect rate from inspected products; a weather service can evaluate past forecast outcomes; a student can estimate the chance of a bus arriving late from recorded trips. These estimates describe the data collected and may be used to inform future expectations if the process remains sufficiently similar.

Read probability tables and graphs

A frequency table can turn experimental counts into relative frequencies. Divide each event’s count by the total number of observations. For a table of 40 days in which a bus was late on 9 days, on time on 28 days, and early on 3 days, the observed probability of late arrival is 9/40 = 0.225, or 22.5%. The relative frequencies should sum to 1, apart from small rounding differences, if the outcomes are mutually exclusive and cover every recorded trial.

When a graph uses a relative-frequency axis, its bar heights represent proportions rather than raw counts. Check the axis labels before interpreting the display. A bar at 0.25 means one-quarter of the observed cases in that category, not 0.25 observations. If a question asks for experimental probability from a display, identify the relevant count or proportion and confirm the denominator.

Common mistakes

• Calling an observed fraction theoretical. If it comes from recorded outcomes, it is experimental or empirical probability.
• Assuming experimental and theoretical values must match after a particular number of trials. Random outcomes can differ.
• Treating a larger sample as protection against bias. More observations help with random variability only when data collection is appropriate.
• Using the number of favorable outcomes over possible outcomes for data that were actually observed, or using observed counts when the prompt asks for an equally likely model.
• Interpreting the law of large numbers as a guarantee that the next trial will correct an earlier imbalance. Each independent trial retains its own probability.
• Forgetting to divide an observed count by the total number of trials, or using the wrong denominator when some observations are missing.

A quick decision process

1. Look for the basis of the information: stated equally likely outcomes or recorded trial results.
2. For equally likely outcomes, count the outcomes in the event and divide by all possible outcomes.
3. For observed results, divide the event’s observed frequency by the number of recorded trials.
4. Convert fractions to decimals or percentages only if the question asks for them; keep adequate precision while calculating.
5. If comparing the values, describe a difference as sampling variation unless the evidence supports another conclusion.
6. Check the assumptions: fairness, independence, consistent procedure, and representative data where relevant.

Exam takeaway

Theoretical probability is a prediction from a model; experimental probability is a relative frequency from data. Both range from 0 to 1, and both can be written as fractions, decimals, or percentages. A short run can differ noticeably from a model. Over many suitable repetitions, the observed proportion tends to stabilize near the model value, while biased collection or changing conditions can keep an experimental estimate from answering the intended question.