Can expected value be impossible? Yes, and often is

By Jude Wallis · Published

Yes, and it happens constantly. A fair die has an expected value of 3.5, a value the die cannot show. The expected value is the long-run average over many repetitions, not a forecast of any one trial, so nothing requires it to be an outcome the variable can actually take.

AP Statistics: Unit 2 (topics 2.9 Parameters of Random Variables). Unit 2 topic 2.9 of the Fall 2026 AP Statistics course defines the mean of a discrete random variable as the long-run average outcome and asks you to interpret it in context, which is precisely the reasoning that makes an expected value of 3.5 on a six-sided die correct rather than paradoxical.

The fair die settles it

Roll a fair six-sided die. Each face has probability 1/61/6, so the expected value is

μX=1(16)+2(16)+3(16)+4(16)+5(16)+6(16)=216=3.5\mu_X = 1\left(\tfrac{1}{6}\right) + 2\left(\tfrac{1}{6}\right) + 3\left(\tfrac{1}{6}\right) + 4\left(\tfrac{1}{6}\right) + 5\left(\tfrac{1}{6}\right) + 6\left(\tfrac{1}{6}\right) = \frac{21}{6} = 3.5

The die has no 3.5 on it. It never will. And yet 3.5 is the correct expected value, not an approximation and not a rounding artifact.

This is the fact students trip over most, and the reason is almost always the word "expected". In ordinary English, expecting something means predicting it will happen. In statistics, E(X)E(X) names a weighted average of a whole distribution. The English meaning is the wrong one, and once you drop it the paradox disappears.

Why: it is a long-run average, not a prediction

The expected value answers one question: if you repeated this random process a huge number of times and averaged all the results, what number would that average settle near?

Averaging is exactly the operation that produces values no single trial can produce. Two households with 1 child and 2 children average 1.5 children. No household has 1.5 children. The average is a summary of the collection, not a description of any member of it, so a headline about the average number of children per family is never a claim that some family has a fractional child.

The law of large numbers is the formal version. As the number of trials grows, the mean of the observed results converges toward μX\mu_X. The individual results stay whole, discrete, and legal. Their running average does not have to be any of those things. That difference between the expected value and the sample mean is the whole story.

The balance point picture

Draw the probability distribution as bars sitting on a number line, with each bar's height equal to its probability. The expected value is the point where that line would balance on a fulcrum.

For the fair die, six equal bars over 1 through 6 balance in the middle, at 3.5, which sits in the gap between the 3 bar and the 4 bar. There is nothing strange about a balance point falling where no weight sits: a seesaw with a child on each end balances at the empty middle.

That picture also predicts when the expected value will be impossible. Whenever a distribution puts its weight on separated values, such as whole counts, the balance point will usually land between them. Getting a possible value out is the special case, not the rule.

The picture is worth keeping because it tells you when a computed expected value is wrong. If your answer falls outside the range of possible values, you have made an arithmetic error: a balance point always sits between the smallest and largest value carrying weight.

A second example: a payout that cannot happen

A carnival game pays 5 dollars with probability 0.10, pays 1 dollar with probability 0.30, and pays nothing with probability 0.60. The probabilities add to 1, so the expected payout is

E(X)=5(0.10)+1(0.30)+0(0.60)=0.50+0.30+0=0.80E(X) = 5(0.10) + 1(0.30) + 0(0.60) = 0.50 + 0.30 + 0 = 0.80

The expected payout is 80 cents. The game has exactly three payouts, 5 dollars, 1 dollar, and 0 dollars, and 80 cents is not among them. No player will ever walk away with 80 cents.

What 80 cents does describe is the booth's accounting. Run the game 1000 times and the booth pays out about 800 dollars, because 1000×0.80=8001000 \times 0.80 = 800. That is the practical use of an impossible expected value: it scales up to totals that are entirely real.

What the expected value does let you predict

An impossible expected value is still a useful one. Three things it supports:

  • Totals over many trials. For nn independent repetitions, the expected total is n×μXn \times \mu_X. Roll a die 100 times and the expected sum of the rolls is 100×3.5=350100 \times 3.5 = 350, a perfectly possible total.
  • Comparisons between options. Two games, two strategies, two insurance policies: the one with the higher expected value wins in the long run, whether or not either expected value is achievable.
  • A fairness check. A game is fair when the expected net gain is 0. That is a statement about the average, not a promise that any player breaks even.

What it does not support is a claim about the next trial. "The expected value is 3.5, so I expect to roll a 3.5" and "the expected value is 3.5, so I expect a 3 or a 4" are both wrong, the second more subtly: 3 and 4 are no more likely than 1 or 6 on a fair die.

Mistakes to avoid

  • Do not round the expected value to a possible outcome. Reporting 3.5 as "about 4" for a die destroys the answer. Keep the decimal.
  • Do not confuse it with the most likely outcome. That is the mode. A distribution with P(0)=0.6P(0) = 0.6 and P(5)=0.4P(5) = 0.4 has mode 0 and expected value 2, and 2 has probability 0.
  • Do not read it as a prediction for one trial. Write "in the long run, over many repetitions, the average is 3.5", which is the phrasing that earns the interpretation point.
  • Do check that it lands inside the range of possible values. An expected value above the maximum or below the minimum is always an arithmetic mistake, usually probabilities that do not add to 1. The steps are in how to find expected value.

The fair die, and what 3.5 is good for

A fair six-sided die is rolled. Find μX\mu_X for the number showing, explain why it cannot occur, and find the expected total for 100 rolls.

  1. Each face has probability 1/61/6, and the six probabilities add to 6×16=16 \times \frac{1}{6} = 1.

  2. Multiply and add: μX=1+2+3+4+5+66=216=3.5\mu_X = \frac{1 + 2 + 3 + 4 + 5 + 6}{6} = \frac{21}{6} = 3.5.

  3. Check the range: 3.5 lies between the smallest value 1 and the largest value 6, so the answer at least clears the balance-point check.

  4. The possible outcomes are the whole numbers 1 through 6. Since 3.5 is not a whole number, no single roll can produce it, and P(X=3.5)=0P(X = 3.5) = 0.

  5. For 100 independent rolls the expected total is 100×3.5=350100 \times 3.5 = 350.

  6. That total is achievable, and any single roll landing on 3.5 is not. The expected value is a property of the distribution, not of one outcome.

μX=3.5\mu_X = 3.5. No face shows 3.5, so the expected value is impossible as an outcome, yet it is exactly right as a long-run average: over 100 rolls, expect the numbers to total about 350, an average of 3.5 per roll.

A carnival game where no player can win the expected amount

A game costs 1.50 dollars to play. It pays 5 dollars with probability 0.10, 1 dollar with probability 0.30, and 0 dollars with probability 0.60. Find the expected payout and the expected net gain per play, and say whether either is a possible result for one player.

  1. Confirm the probabilities add to 1: 0.10+0.30+0.60=10.10 + 0.30 + 0.60 = 1.

  2. Expected payout: 5(0.10)+1(0.30)+0(0.60)=0.50+0.30+0=0.805(0.10) + 1(0.30) + 0(0.60) = 0.50 + 0.30 + 0 = 0.80 dollars.

  3. The three possible payouts are 5, 1, and 0 dollars, so 0.80 dollars cannot happen on any single play.

  4. Build the net gain by subtracting the 1.50 dollar cost from each payout: 5−1.50=3.505 - 1.50 = 3.50, 1−1.50=−0.501 - 1.50 = -0.50, and 0−1.50=−1.500 - 1.50 = -1.50.

  5. Expected net gain: 3.50(0.10)+(−0.50)(0.30)+(−1.50)(0.60)=0.35−0.15−0.90=−0.703.50(0.10) + (-0.50)(0.30) + (-1.50)(0.60) = 0.35 - 0.15 - 0.90 = -0.70 dollars.

  6. Cross-check with the shortcut E(payout)−cost=0.80−1.50=−0.70E(\text{payout}) - \text{cost} = 0.80 - 1.50 = -0.70, which agrees.

  7. The three possible net results are 3.50, -0.50, and -1.50 dollars. None of them is -0.70, so the expected net gain is also impossible as a single result.

Expected payout 0.80 dollars and expected net gain -0.70 dollars, and neither can occur on any one play. The reading that is correct: over many plays, players lose about 70 cents per play on average, so the booth takes in roughly 70 dollars per 100 plays. The game is not fair, since a fair game has an expected net gain of 0.

Frequently asked questions

Is the expected value the most likely outcome?

No. The most likely outcome is the mode. If P(0)=0.6P(0) = 0.6 and P(5)=0.4P(5) = 0.4, the mode is 0 while E(X)=0(0.6)+5(0.4)=2E(X) = 0(0.6) + 5(0.4) = 2, and 2 has probability 0. Expected value is a balance point, not a peak.

Should I round an impossible expected value to a whole number?

No. Rounding 3.5 to 4 changes a correct answer into a wrong one and loses the point. Report the decimal and interpret it as a long-run average. Round only at the end, and only as far as the context or the question calls for.

Do binomial distributions have this problem too?

Yes. The binomial mean is μX=np\mu_X = np, so 5 flips of a fair coin give μX=5(0.5)=2.5\mu_X = 5(0.5) = 2.5 heads. You cannot observe 2.5 heads. Over many sets of 5 flips, the average number of heads approaches 2.5.

Can an expected value fall outside the possible values entirely?

No, and that is a useful check. The expected value is a weighted average, so it always lies between the smallest and largest values that carry probability. Landing outside that range means an arithmetic slip, most often probabilities that do not add to 1.

How should I word an interpretation on the AP exam?

Use long-run language with units and context: "if this process were repeated many times, the average number of goals per game would be about 1.1". Avoid "we expect 1.1 goals next game", which reads as a prediction for one trial and does not earn the point.