Equal means can hide different risks. Density height is not point probability.
By the end: Interpret variance, standard deviation, quantiles, and density areas without calculus.
Variance is the average squared distance from the mean, weighted by probability. Squaring avoids cancellation of positive and negative deviations. Standard deviation is the square root of variance, so it uses the original variable’s units. Neither quantity is a maximum possible error.
A sure payoff of ten has mean ten and standard deviation zero. The zero-or-100 gamble with probabilities 0.9 and 0.1 has the same mean, variance 900, and standard deviation thirty. Its rare large payout drives spread. Means and standard deviations alone still do not determine every tail probability.
Consider a delivery charge of four units with probability three quarters and twelve units with probability one quarter. The mean charge is three plus three, or six units. Relative to that mean, four is two units below and twelve is six units above. Weighting the signed deviations gives negative 1.5 plus positive 1.5, or zero. This cancellation is inevitable for deviations from the mean. It cannot tell us whether charges stay close to six or vary widely around it.
Variance avoids that cancellation by squaring each deviation before weighting it. The squared deviations are four and thirty-six. Their weighted contributions are three quarters times four, which is three, and one quarter times thirty-six, which is nine. Variance is their sum, twelve squared charge units. The rarer high charge supplies three quarters of this variance because its squared distance is nine times larger. A low probability does not prevent a distant outcome from making a large contribution to spread.
Standard deviation is the square root of variance, about 3.46 charge units here. Squaring 3.46 gives approximately twelve. Taking the square root restores the original unit, which makes standard deviation easier to compare with the mean. It is not the expected signed deviation, which is zero. It is also not a maximum distance from the mean. The twelve-unit charge is six units above the mean, exceeding one standard deviation while remaining a permitted outcome with probability one quarter.
| Charge | Probability | Deviation | Squared distance | Weighted square |
|---|---|---|---|---|
| 4 | 3/4 | −2 | 4 | 3 |
| 12 | 1/4 | 6 | 36 | 9 |
| Weighted total | 1 | 0 | Not summed directly | 12 |
Compare a charge that is always six. It has the same mean but every deviation is zero, so its variance and standard deviation are zero. Now compare equally likely charges of two and ten. Their mean is also six, and their distances from it are both four. Variance is sixteen and standard deviation four. Those two outcomes are less extreme on the high side than twelve, yet their greater weight at a substantial distance gives more variance than the four-or-twelve distribution.
Changing units must change these summaries coherently. If a charge measured in whole units is converted to hundredths of a unit, every value and the mean multiply by one hundred. Every squared deviation multiplies by ten thousand, so variance becomes one hundred twenty thousand squared hundredths. Standard deviation multiplies by one hundred, becoming about 346.41 hundredths. The underlying uncertainty has not increased. Only the numerical scale changed. Comparing variances without checking units can therefore produce meaningless conclusions.
Try adding a fixed handling fee of five units to both original charges. The outcomes become nine and seventeen, and their mean becomes eleven. Their distances from the new mean remain negative two and positive six, so variance stays twelve and standard deviation stays about 3.46. In contrast, doubling both charges doubles every distance and quadruples variance. Shifting a distribution changes its location; stretching its numerical scale changes its spread. Neither operation by itself changes the probability attached to each outcome.
In a continuous idealization, probability belongs to intervals. A uniform wait between ten and twenty minutes assigns probability in proportion to interval length. Waiting between twelve and fifteen minutes has probability 3/10. A density plot has height 0.1 per minute over this ten-minute range, so total area is one.
A single exact value has zero probability in this continuous model, even though some value will occur. A recorded wait of fifteen minutes usually means a rounded interval, not an infinitely precise point. Density can exceed one: uniform waiting over half a minute has density two per minute, with area still one.
A discrete charge distribution places probability mass at individual values. Asking for the probability of a charge exactly four makes sense in that model. A continuous wait model works differently. Suppose a shuttle arrives at a time spread uniformly between zero and four minutes after you start waiting. Every interval of equal length inside this range receives equal probability. Since four minutes must contain total probability one, each minute contributes one quarter. The density is therefore 0.25 per minute throughout the range.
A wait between one and three minutes covers two of the four minutes. Its probability is two times 0.25, or one half. A wait between one and one and a half minutes covers half a minute, with probability 0.125. On the density drawing, horizontal width measures minutes and vertical height measures probability per minute. Multiplying them cancels the time unit and gives probability. The whole rectangle has width four and height one quarter, so its total area is one.
A narrower uniform model can have a taller density without violating any probability rule. Suppose an express shuttle arrives uniformly between zero and one quarter of a minute. Its density is four per minute because one quarter times four equals one. An interval of 0.05 minute inside that range has probability 0.2. Height four is not a 400% event probability. The height describes how much probability accumulates per minute, while an event probability needs the width of an interval as well.
Changing time units changes density height too. The express range is fifteen seconds wide. In seconds its density is one fifteenth per second, rather than four per minute. A three-second interval then has probability three times one fifteenth, again 0.2. The event did not change when its labels changed units. This gives a useful check on any density interpretation. A claimed probability should survive unit conversion, even though the numerical density height usually changes with the horizontal scale.
An exact point has no width and therefore zero probability in these continuous models. This does not mean that no waiting time occurs. The model distributes total probability across intervals containing infinitely many points, without giving positive mass to a particular point. A recorded value is different because instruments round. If a display rounds to the nearest tenth of a minute, a displayed 1.2 usually represents actual times from 1.15 up to 1.25 minutes, with a boundary convention for exact ties.
In the four-minute uniform model, that displayed value has probability 0.1 times 0.25, or 0.025. At a boundary, the rounding interval may be shorter because waits cannot be negative or exceed four. The displayed zero represents only zero through 0.05 minute, giving probability 0.0125 under this convention. Use the comparison figure to predict what doubling display precision does to an interior reading's probability. Its represented width halves, so its mass halves, although the underlying continuous waiting model stays unchanged.
| Display increment | Displayed value | Actual interval, minutes | Probability mass |
|---|---|---|---|
| 0.1 minute | 1.2 | 1.15 to 1.25 | 0.1 × 0.25 = 0.025 |
| 0.1 minute | 0 | 0 to 0.05 | 0.05 × 0.25 = 0.0125 |
| 0.05 minute | 1.20 | 1.175 to 1.225 | 0.05 × 0.25 = 0.0125 |
| Infinite precision | Exactly 1.2 | One point, zero width | 0 |
Squared distances are used, so signs do not matter. Variance has squared payoff units.
| Distribution | Mean | Standard deviation |
|---|---|---|
| Always ten | 10 | 0 |
| Zero or 100, with 90%/10% weights | 10 | 30 |
Suppose a repair takes one day with probability 0.6, three days with probability 0.3, and eight days with probability 0.1. Its mean duration is 0.6 plus 0.9 plus 0.8, or 2.3 days. That number does not directly answer a customer asking how long to allow for 90% of repairs. Add probability in time order instead. By one day the accumulated probability is 0.6. By three days it is 0.9. By eight days it is one.
Under the convention that a quantile is the smallest value where accumulated probability reaches the requested level, the median is one day and the 90th percentile is three days. The 95th percentile is eight days. A small change in the requested percentage can therefore produce a large jump in the deadline for a discrete model. There is no actual repair duration between three and eight in this construction. Interpolating a duration without stating a different convention would obscure that discrete gap.
Pay attention to whether a deadline includes its endpoint. The probability of taking more than three days is 0.1. The probability of taking at least three days is 0.4 because it includes the mass at exactly three. In a continuous density model, an individual endpoint has probability zero, so these inequalities have the same probability at a fixed boundary. Discrete masses make the difference visible. A contract that says 'within three days' needs an operational definition of the deadline and its measurement precision.
Mean and standard deviation still do not specify every tail. Consider two nonnegative cost distributions. The first assigns equal probability to costs zero and two. The second assigns probability one third to zero, one half to one, and one sixth to three. Both have mean one and variance one, so both have standard deviation one. Yet the first has no probability of cost above two, while the second places one sixth of its mass there. Matching two summaries does not match the distribution.
You can verify the second variance without a formula shortcut. Distances from its mean of one are negative one, zero, and two. Their squares are one, zero, and four. Weighting gives one third plus zero plus four sixths, which equals one. Its mean is zero plus one half plus three sixths, also one. For the first distribution, both distances from one have magnitude one, so the weighted squared distance is one. The equal summaries conceal different possible costs and different frequencies.
For guided exploration, choose a threshold of one and compare the chance of exceeding it in these two cost models. The first gives one half; the second gives one sixth. Move the threshold to 2.5 and the first falls to zero while the second stays at one sixth. Neither model has a universally smaller tail at every threshold. A manager concerned about exceeding a reserve needs the probability at that reserve, not an unsupported assumption that equal standard deviations imply equal chances of running out.
| Cost or event | Distribution A | Distribution B |
|---|---|---|
| Cost 0 | Probability 1/2 | Probability 1/3 |
| Cost 1 | Probability 0 | Probability 1/2 |
| Cost 2 | Probability 1/2 | Probability 0 |
| Cost 3 | Probability 0 | Probability 1/6 |
| Cost > 1 | 1/2 | 1/6 |
| Cost > 2.5 | 0 | 1/6 |
| Mean; variance | 1; 1 | 1; 1 |
A quantile is a threshold below which a specified proportion lies. In the continuous uniform ten-to-twenty model, the 90th percentile is nineteen minutes. For discrete distributions, quantiles can jump and need a convention; one common definition takes the smallest value whose cumulative probability reaches the requested proportion.
A bell-shaped normal distribution is a useful model for some quantities and averages, not a universal law of data. A standard deviation does not automatically imply normal tail probabilities. Skew, heavy tails, truncation, and mixtures can make a normal approximation misleading, especially for rare-event decisions.
A single bell-shaped curve can hide a mixture of operating conditions. Imagine that 80% of deliveries use a nearby route with times uniformly spread between two and four hours. The other 20% use a distant route with times uniformly spread between eight and ten hours. The nearby route's mean is three hours, halfway across its range. The distant route's mean is nine hours. Weighting the route means gives 0.8 times three plus 0.2 times nine, or 4.2 hours overall.
The overall mean lies in a gap where the model permits no delivery time. The combined density is 0.4 per hour between two and four, since that two-hour interval must receive weight 0.8. It is 0.1 per hour between eight and ten, giving area 0.2. It is zero in between. Replacing this shape with a bell curve centered at 4.2 would invent probability for times between four and eight and might alter the very late tail that a scheduling decision needs.
A deadline of five hours is exceeded by every distant-route delivery and by no nearby-route delivery. Its exceedance probability is therefore 20%. A deadline of nine hours is exceeded by half the distant route's uniform range. Multiplying its route weight 0.2 by that conditional half gives 10% overall. This multiplication uses a conditional probability, not independence between route and duration. Route directly determines which duration range is possible. The weighted partitions let you calculate the event without pretending that all delivery times follow one common shape.
Changing route proportions changes the overall distribution even if each route's behavior stays fixed. If the distant share rises to 40%, the mean becomes 0.6 times three plus 0.4 times nine, or 5.4 hours. The chance of exceeding five hours becomes 40%, and of exceeding nine hours becomes 20%. A higher average here does not show that drivers became slower on either route. It comes entirely from a different mixture. Comparing only overall averages can misidentify the source of a change.
Conditioning on route can improve prediction because it removes uncertainty about which range applies. That is different from establishing the causal effect of assigning a delivery to a route. Distant deliveries may carry different goods or require different stops. Observational differences between groups do not establish what would happen if the same delivery were reassigned. A representative sample helps describe the mixture actually encountered. A causal comparison needs evidence about assignment and other differences, not merely a more detailed probability distribution.
Finally separate outcome spread from uncertainty about the model itself. The two uniform ranges and their weights are stipulated here. If you estimated them from a few deliveries, their values would also be uncertain. A range describing future delivery times is a prediction statement; a confidence interval concerns a procedure's repeated-sample coverage for a fixed parameter. A posterior interval instead assigns probability to parameter values under a prior and likelihood. None can be read directly from the density's height or from its standard deviation alone.
A variable is zero or four, each with probability one half. Find its mean, variance, and standard deviation.
First find the mean, then square each distance from it.
| Outcome | Distance from mean | Squared distance |
|---|---|---|
| 0 | −2 | 4 |
| 4 | 2 | 4 |
| Weighted average | 4 |
The mean is two. Both outcomes lie two units from it, so variance is 0.5 × 4 + 0.5 × 4 = 4 squared units. Standard deviation is two units. The variable never equals its mean in this example.
A wait is uniform from zero to half a minute. What are its density and the probability of waiting at most 0.1 minute?
Area, not height alone, is probability.
| Interval width | Density | Probability area |
|---|---|---|
| 0.1 minute | 2 per minute | 0.2 |
| 0.5 minute | 2 per minute | 1 |
The density is two per minute. The interval from zero to 0.1 minute has area 0.1 × 2 = 0.2. Height two is valid because the entire support has width 0.5 and total area one.
The explanations and exercises are complete without these optional links. The references distinguish probability identities from inference procedures and reporting guidance. The examples are constructed teaching models, not empirical estimates or professional decision advice.
Textbook definitions of experiments, sample spaces, events, and probability. The course’s numerical examples are original constructions.
Continuous density and cumulative probabilities. Our waiting-time example rescales the uniform model.