Statistics and probability · Eqora guide
Expected value in math: average, not prediction
Weight each possible value by its probability, then explain what the resulting average does—and does not—tell you.
A model says a workshop will need zero replacement tiles with probability 0.5, two with probability 0.3 and five with probability 0.2. Its expected requirement is 1.6 tiles. Yet no listed outcome is 1.6. That is not a calculation error: expected value in math answers a question about a probability-weighted average, not the exact result of the next workshop.
The difficult part is often deciding what the random variable measures and interpreting the answer after the arithmetic. We will build a complete finite model, calculate the weighted contributions, distinguish the average from the most likely outcome, and test how changing the model changes the result. You can do every calculation on paper; no solver or app is required.
Eqora publishes this explanation as learning support, not a guarantee of correct results or grades. Check the original question, definitions, units and assumptions yourself. The workshop data are invented exact teaching values, not observations or operational advice. The mosaic photographs are conceptual analogies; their fragment counts do not depict the numerical probabilities.
Expected value in math starts by defining the quantity
Let X be the number of replacement tiles needed for one workshop. Capital X names the random variable before the outcome is known; a lowercase value such as x = 2 names one possible realization. It matters that X counts replacements, not all tiles used, broken fragments or attendees. Those would be different quantities with potentially different probability models and different averages.
Our possible values are 0, 2 and 5. They are mutually exclusive: one workshop cannot simultaneously need zero and five replacements. They are also exhaustive within this constructed model: every workshop falls into one of these categories. If four replacements could occur but that possibility was omitted, the table would not describe X completely, and an apparently tidy sum could still be misleading.
Write a sentence defining the experiment and the unit before writing a formula. Here the experiment is one workshop under the stipulated conditions, and the unit of X is tiles per workshop. A probability has no tile unit; it represents a weight. Keeping these roles separate makes it easier to notice when a calculation combines unrelated events or accidentally averages the category labels instead of the measured values.
Check that the probability weights form a complete model
Every probability must lie between zero and one, and the weights for all possible values must sum to one. Our check is 0.5 + 0.3 + 0.2 = 1. These numbers mean fifty, thirty and twenty percent. They do not mean fifty, thirty and twenty tiles. Convert percentages to decimals before multiplying; otherwise the expected value will be one hundred times too large.
The basic calculation is a probability weighted average. Yale's random-variable notes define expectation by weighting the possible values by their probabilities. For a finite list, E(X) = x₁p₁ + x₂p₂ + ⋯ + xₖpₖ. The subscript pairs each value with its own probability. There is no extra division by the number of rows because the probability weights already add to one.
If you are instead given observed counts, first decide whether you are calculating a sample average or constructing an estimated probability model. Our weighted-average guide explains how frequency weights are normalized by their total. Ten observed workshops would not establish exact future probabilities automatically. A model supplied by an exercise and an estimate from data can use similar arithmetic while carrying very different levels of uncertainty.
Calculate each contribution before adding
For our workshop, the contributions are 0 × 0.5 = 0, 2 × 0.3 = 0.6 and 5 × 0.2 = 1. Their sum is E(X) = 1.6 tiles per workshop. Include the zero outcome in the model even though its product is zero. It still consumes half the probability mass and explains why the other outcomes must not be treated as equally likely.
A wrong shortcut would be (0 + 2 + 5)/3 = 7/3. That averages the three distinct values as though each had probability one third. Our model gives zero a larger weight than either nonzero result, so this shortcut answers a different question. A useful handwritten layout has one line per value, its probability and its product; sum the products only after checking the pairs.
You can verify 1.6 by an idealized frequency construction. Imagine a batch of ten workshops containing exactly five zero-replacement cases, three two-replacement cases and two five-replacement cases. That batch uses 5 × 0 + 3 × 2 + 2 × 5 = 16 replacements, or 16/10 = 1.6 per workshop. This is a constructed explanation of the weights, not a promise that every ten workshops will have those counts.
A noninteger average is not a fractional event
The value 1.6 summarizes the distribution; it does not add a new possible outcome to it. An actual workshop in our model needs 0, 2 or 5 replacements. Averaging counts across many workshops can produce a noninteger value even though every individual count is an integer. The same distinction explains why an average household size need not be the size of any particular household.
The most likely outcome here is zero because it has probability 0.5, larger than 0.3 or 0.2. That is the mode, not the expectation. Saying that the expected value is 1.6 does not make two the most likely outcome merely because two is close to 1.6. Rounding the expectation is a numerical operation; it does not change the underlying probability distribution.
When interpreting a question, separate an average, a likely individual result and a requirement that must always be met. They are not interchangeable. An exercise asking for E(X) can be answered with 1.6 and a careful sentence about the average. A question about enough stock for every possible modeled workshop would involve the largest listed requirement, five, and perhaps other conditions not supplied by the exercise.
Long-run interpretation needs stable assumptions
Under repeated independent trials with the same finite distribution, the sample average approaches the expected value in the law-of-large-numbers sense. This is not a fixed deadline or an exact equality after a chosen number of trials. A short sequence can differ substantially from 1.6. Even an unusually long run of zero outcomes does not make the next workshop owe the model extra replacements.
The underlying probabilities must still describe the situation you are repeating. Changing materials, workshop size or the definition of replacement can change the distribution. Combining observations from different conditions may estimate an average of a mixture rather than the original model. A correct expectation formula does not repair unsupported assumptions about where the probabilities came from or whether they remain relevant.
To check your interpretation, name the repeated quantity and the assumptions alongside the number: under this stipulated stable model, replacement demand averages 1.6 tiles per workshop over repeated trials. Avoid claims such as the next workshop will need about two or every group of ten needs sixteen. Those are stronger statements than the model's expectation supports and require different information.
Changes of scale and sums have different rules
Suppose each replacement contributes three practice points and every workshop contributes one fixed point. Define Y = 3X + 1. The possible point totals are 1, 7 and 16, with the same probabilities. Direct calculation gives E(Y) = 1 × 0.5 + 7 × 0.3 + 16 × 0.2 = 5.8. Alternatively, 3E(X) + 1 = 3 × 1.6 + 1 = 5.8.
This illustrates linearity of expectation: E(aX + b) = aE(X) + b for this finite model. The fixed amount is added once to the expectation, not once per unweighted row. For two finite random variables, E(X + Z) = E(X) + E(Z), even without independence. Independence matters for many probability and variance calculations, but it is not required for this expectation-of-a-sum rule.
Do not extend that rule to arbitrary products or nonlinear transformations. In our model E(X²) = 0² × 0.5 + 2² × 0.3 + 5² × 0.2 = 6.2, while [E(X)]² = 1.6² = 2.56. Squaring before averaging is different from squaring the average. Write the transformed possible values explicitly whenever you are uncertain which operation belongs inside the expectation.
The same expectation can hide different variability
Consider model A, which always returns two tiles, and model B, which returns zero or four with probability one half each. Both have expectation two. Model A never varies; model B never returns two at all. Thus the expectation alone does not describe how spread out the possible results are, nor does it determine the chance of exceeding a particular threshold.
For the workshop distribution, the variance is E(X²) − [E(X)]² = 6.2 − 2.56 = 3.64 square tiles. Its standard deviation is √3.64, approximately 1.91 tiles. The squared unit belongs to variance; taking the square root restores the original unit. These spread measures complement the mean, but they still do not predict a specific next outcome or replace the full distribution.
You do not need to calculate variance unless the task asks for it. The important independent check is recognizing that an average cannot answer every probability question. To find the probability of needing more than one replacement in our model, sum the probabilities for two and five: 0.3 + 0.2 = 0.5. Do not divide the expectation by the threshold or treat 1.6 as a probability.
Solve a changed model without copying the first average
For a fresh practice model, let Z count replacement pieces with values 1, 3 and 6 and probabilities 0.4, 0.4 and 0.2. Before looking at the arithmetic, define the unit, check the probability sum and pair every value with its weight. The expectation is 1 × 0.4 + 3 × 0.4 + 6 × 0.2 = 0.4 + 1.2 + 1.2 = 2.8 pieces.
The result lies between the minimum one and maximum six, as every finite probability-weighted average must. That bound is necessary but not sufficient for correctness. The modal outcomes are tied at one and three, while 2.8 is not a possible individual result. For an idealized ten-case frequency check, four cases of one, four of three and two of six yield twenty-eight pieces, averaging 2.8.
Now explain the answer without a formula: the stipulated probabilities imply a long-run average of 2.8 pieces per trial, not a prediction of the next trial. Our independent answer-checking guide can help structure that final comparison with the original conditions. Keep the distribution, products, unit and interpretation together in your notes; if any one of them changes, recompute rather than reuse the earlier number.
