AI Glossary / Statistics and probability / Intermediate

Cumulative distribution

Q: Is a cumulative distribution always smooth?

No. Count distributions and empirical distributions commonly have steps.

Q: Can a cumulative probability exceed one?

No. If it does, the construction or interpretation is wrong for an ordinary probability distribution.

Q: Is a percentile the same thing?

No. A cumulative value gives the probability below a supplied threshold; a percentile identifies a threshold associated with a supplied cumulative proportion, with conventions needed for ties and samples.

Definition

A cumulative distribution describes how probability accumulates up to each value of a variable. Its standard mathematical representation is the cumulative distribution function, F(x) = P(X <= x). Instead of asking how much probability belongs to one value or interval, it asks how much lies at or below a chosen threshold.

What it means

The cumulative view answers questions such as “What proportion of requests finish within two seconds?” or “How likely is daily demand to stay within available capacity?” It applies to counts, measurements, and distributions containing both kinds of outcomes.

The term is commonly used as shorthand for the cumulative distribution function rather than a separate kind of probability model. A cumulative curve can represent an assumed population model or the observed proportions in a dataset. Those interpretations must be distinguished: an observed percentage is evidence about a population, not automatic certainty about future performance.

How it works

For discrete outcomes, add the probability masses at every value up to and including the threshold. For a variable with a probability density, integrate that density up to the threshold. An empirical cumulative distribution instead counts observations at or below the threshold and divides by the sample size.

The resulting curve cannot decrease as the threshold increases. Its values remain between zero and one, approaching zero toward the far left and one toward the far right. Discrete outcomes produce jumps, while intervals containing no probability produce flat stretches.

Subtract cumulative values to obtain interval probabilities: F(b) - F(a) gives P(a < X <= b). The probability above a threshold is 1 - F(x). Endpoint conventions matter when exact values have positive probability, as with counts or rounded measurements.

Three practical examples

A customer-service team reads the share of requests completed within successive waiting-time thresholds, instead of relying solely on the average wait.
A warehouse models the chance that daily order counts fit within staffing capacity, using a cumulative demand distribution.
An AI evaluation team compares the proportions of held-out examples whose absolute prediction errors are no greater than selected tolerances.

Think of it this way

Think of filling a measuring jug with all the probability lying to the left of a moving marker. As the marker moves right, more probability can enter the jug, but none is removed. A large jump means considerable probability is concentrated at one value. A flat stretch means the marker has crossed an interval with no additional probability.

Worked example

Consider an original dataset of 20 service requests, classified by the number of follow-up messages required. Two requests need zero messages, five need one, eight need two, and five need three.

The empirical probabilities at 0,1,2,3 are therefore 0.10,0.25,0.40,0.25. The cumulative proportions are F(0) = 0.10, F(1) = 0.35, F(2) = 0.75, and F(3) = 1.

Thus 75% of observed requests required at most two follow-ups. The proportion requiring more than one but at most two is F(2) - F(1) = 0.40. The proportion requiring at least two is 1 - F(1) = 0.65, whereas the proportion requiring more than two is 1 - F(2) = 0.25.

There is no possible count between one and two, so F(1.5) = 0.35. These calculations describe the 20 observations. Applying the percentages to next month's requests would require considering sampling uncertainty and whether the work has changed.

Illustrative case study

In this fictional case, Beacon Desk compares two AI-assisted support workflows. Their average completion times are similar, so managers initially consider them interchangeable. An analyst plots cumulative completion-time distributions for a fixed set of comparable tasks.

The curves reveal that one workflow finishes more tasks quickly but also leaves a larger share waiting beyond the team's deadline. The curves cross, so neither workflow is uniformly faster at every threshold. Managers focus on the service thresholds that matter to customers and separately investigate the slow cases.

This invented case illustrates how a distribution can reveal tradeoffs hidden by an average. It does not report a measured performance improvement or recommend a particular system.

Common misconceptions and mistakes

A cumulative value is not the probability of exactly that value. For a count variable, exact-value probabilities correspond to jumps, not the total height of the cumulative curve.

Do not confuse a cumulative probability distribution with a running total over calendar time. The horizontal axis here is the possible value of the measured variable, which need not be time at all.

A higher cumulative curve is not universally “better.” For delay or error, higher cumulative probability at a low threshold may be desirable; for revenue, the interpretation changes. Also, grouped data, censoring, and survey weights can require methods beyond simply counting complete records.

Business use

Businesses can use cumulative distributions to translate a statistical model into threshold-based decisions. Capacity limits, tolerances, service deadlines, and customer commitments are often easier to discuss as proportions meeting a defined condition than as an average alone.

State the variable, units, population, time window, and whether the curve is empirical or modeled. Include the sample size and appropriate uncertainty when comparing alternatives. A curve built from completed cases alone can misrepresent long-running unresolved cases, so verify what the dataset excludes before making service claims.

Check your understanding

Check your understanding: In the example, why is F(2) equal to 0.75 rather than 0.40?

Answer: It includes requests requiring zero, one, or two follow-ups.

Why does “at least two” differ from “more than two”?

Answer: The former includes the eight requests with exactly two follow-ups.

Does the observed 75% guarantee the same proportion next month?

Answer: No; it is a sample description, not a future guarantee.

Related terms

Learn next

Construct an empirical cumulative curve from a small dataset containing ties. Practice translating “less than,” “at most,” and “at least” into probabilities, then compare a cumulative curve with a histogram of the same observations.

Sources

Sources: Pishro-Nik, Introduction to Probability, Statistics, and Random Processes, Cumulative Distribution Function, https://www.probabilitycourse.com/chapter3/3_2_1_cdf.php. NIST/SEMATECH e-Handbook, Related Distributions, https://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm. This entry uses cumulative distribution in its standard CDF sense. Calculations and the fictional case are original.

Previous: Cuckoo hashing Next: Cumulative distribution function