AI Glossary / Statistics and probability / Intermediate

Probability density function

Definition

A probability density function, often abbreviated PDF, describes how probability is distributed across the values of an absolutely continuous random variable. It is nonnegative and has total integral one. The probability of falling within an interval is the area under the density over that interval. The density at a single point is not the probability of obtaining exactly that value. In the ordinary continuous-density setting, each exact point has probability zero.

What it means

Density measures probability per unit of the variable. A narrow region can have a high density while containing only a modest probability because its width is small. Consequently, a density can exceed one without violating probability rules. Its numerical height also depends on the measurement units. Probabilities remain unchanged when units are converted correctly, but density heights change to keep the corresponding areas equal. A curve labeled as a density therefore needs meaningful axes and units.

How it works

Specify a distribution with a suitable density or estimate one from observations using an appropriate method. To calculate an interval probability, integrate the density over that interval or subtract cumulative distribution values at the endpoints. Check that the density is nonnegative and integrates to one. For a transformation of the variable, account for the change of scale rather than relabeling the horizontal axis alone. Not every probability distribution has an ordinary density; discrete masses and mixed distributions require different treatment.

Three practical examples

Example 1: A manufacturing analyst uses a continuous density model to estimate the proportion of measurements lying outside tolerance limits. Example 2: A service team models waiting times and calculates the probability of a wait exceeding a threshold. Example 3: A researcher estimates a smooth density from observed measurements to inspect modes and tail behavior. A smooth-looking estimated curve is not proof that the assumed distribution is correct, particularly in sparsely observed tails.

Think of it this way

Think of probability as the amount of material spread along a line and density as how thickly it is spread at each location. A very narrow patch can be thick without containing much total material. The amount in a region depends on both thickness and width. This analogy explains area calculations, but a probability density is mathematical: it need not correspond to a physical substance or a directly observable point frequency.

Worked example

Suppose a duration X is uniformly distributed between 0 and 0.20 seconds. Its density is 1/0.20 = 5 per second throughout that interval and zero outside it. The total area is 5 × 0.20 = 1. The probability that X lies between 0.04 and 0.10 seconds is 5 × 0.06 = 0.30. The density at 0.08 seconds is 5, but the probability of exactly 0.08 seconds is zero in this continuous model. Expressing the same duration in milliseconds gives a range from 0 to 200 and density 1/200 = 0.005 per millisecond. The interval from 40 to 100 milliseconds still has probability 0.005 × 60 = 0.30. The density changed with units; the event probability did not.

Illustrative case study

Fictional case study: Stone Instruments displays a density chart for timing measurements. A reviewer initially interprets a peak above one as an impossible probability. The analyst explains the units and shades the area corresponding to an operational tolerance. The team also separates recorded rounded values from the underlying continuous measurement model. This fictional case illustrates communication rather than empirical validation; the uniform distribution in the worked example is chosen for transparent arithmetic, not asserted as a realistic instrument model.

Common misconceptions and mistakes

Common mistakes include reading curve height as point probability, assuming density must stay below one, and adding density values without accounting for interval widths. A histogram of raw counts is not automatically a normalized density estimate. The probability of an exact continuous value being zero does not mean that observing a recorded rounded value is impossible; the recorded value usually represents an interval. A distribution with a point mass cannot be fully described by an ordinary continuous density alone.

Business use

Business use: Density models help quantify tolerances, waiting-time risk and the spread of continuous measurements. Show relevant probability areas rather than relying on peaks alone. Label units and explain whether the curve is assumed, fitted or estimated nonparametrically. Validate predictions in the parts of the distribution that drive decisions. Avoid giving tail probabilities more precision than the data and model assumptions justify, especially when few extreme observations are available.

Check your understanding

In the example, is 5 the probability of a duration of 0.08 seconds?

Answer: No. It is the density in units of inverse seconds.

Why does converting to milliseconds reduce the density height?

Answer: The horizontal scale becomes 1,000 times larger, so the height becomes 1,000 times smaller to preserve probability areas.

Learn next

Calculate areas under a simple uniform density, then compare them with differences in its cumulative distribution function. Change the measurement units and verify that event probabilities stay fixed. Next, examine a histogram with unequal bin widths to see why density normalization matters. Finally, distinguish the probability of a rounded measurement from that of an exact value in a continuous model.

Cumulative distribution function

Sources

Sources: NIST/SEMATECH e-Handbook, What is a Probability Distribution: https://www.itl.nist.gov/div898/handbook/eda/section3/eda361.htm ; Uniform Distribution: https://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm ; The Aalto Dictionary of Machine Learning: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . These institutional sources support the density, integral and distribution distinctions. The unit-conversion calculation and fictional instrument case are original illustrations.

Previous: Probability calibration Next: Probability distribution