probability distributions
The probability distribution is the mathematical description of how the possible outcomes of a random experiment are spread over their values, and it is the single most important idea in the probability and statistics series that this site has been building. A random variable is not a fixed number but a quantity that takes different values with different likelihoods, and the probability distribution assigns a probability to each possible value, or to each interval of values, so that the total probability is always one. Every statistical method in the earlier keywords, from the probability stats row to the inference, bayesian, nonparametric, experimental design, and multivariate rows, is built on a probability distribution: the analysis of variance assumes the normal distribution, the control charts assume a known reference distribution, and the maximum likelihood estimation assumes a particular family of distributions for the data. The engineer who understands the probability distribution knows how likely each measurement is, how much variation to expect, and which statistical model to apply. This document develops the probability mass function and the probability density function, the cumulative distribution function, the mean and the variance, and the key families of discrete and continuous distributions that are used throughout semiconductor statistics.
**The probability distribution is defined through the random variable, which is a function that assigns a number to each outcome of an experiment, and the distribution tells the engineer how that number is spread over its range.** The set of all possible values of the random variable, together with the probabilities attached to them, is the sample space with its distribution, and the distribution must satisfy the two laws of probability: that every probability is between zero and one, and that the probabilities over all the possible values sum to one. A random variable is either discrete, when it takes a countable set of separate values such as the number of defects on a wafer, or continuous, when it takes any value in an interval, such as the thickness of a film. The distribution of a discrete variable is described by its probability mass function, which gives the probability of each separate value, while the distribution of a continuous variable is described by its probability density function, which gives the rate of probability per unit of value. The two types of distribution are handled with slightly different mathematics, but they answer the same question, which is how the values of the variable are spread. The distinction between discrete and continuous variables is the first fork in the study of distributions.
**The probability mass function is the description of a discrete random variable, and it gives the probability that the variable takes each of its possible separate values.** If the discrete variable $X$ can take the values $x_1, x_2, \ldots$, then the probability mass function $p(x)$ satisfies $p(x_i) = P(X = x_i)$ for each value, and it is zero for any value the variable cannot take. The probability mass function must be nonnegative and must sum to one over all the possible values, and the probability that the variable falls in a set is the sum of the probability mass function over the values in that set. A bar chart of the probability mass function shows the relative likelihood of each discrete outcome, and the tallest bars mark the most likely values. The most important discrete distributions are the Bernoulli and binomial distributions for counts of successes, the geometric distribution for the number of trials until a success, and the Poisson distribution for the number of rare events in a fixed interval. In a fab the number of defects on a wafer, the number of particles, and the number of failing dies are modeled by discrete probability mass functions. The probability mass function is the complete description of a discrete random variable.
**The probability density function is the description of a continuous random variable, and it is a nonnegative function whose integral over any interval gives the probability that the variable falls in that interval.** For a continuous variable $X$ with density $f(x)$, the probability that $X$ lies between $a$ and $b$ is the integral of the density from $a$ to $b$, which is the area under the density curve between the two points, and the total area under the entire density curve is one. The density itself is not a probability, because it can be greater than one where the values are very concentrated, and the probability of any single exact value of a continuous variable is zero, which is why the engineer speaks of intervals rather than single points. The density function is the mathematical object that makes it possible to compute probabilities for continuous measurements, and it is the continuous analogue of the probability mass function. The density function is highest where the variable is most likely to fall, and its shape reveals the center, the spread, and the skew of the distribution. In a fab the thickness of a film, the sheet resistance, and the critical dimension are modeled by continuous probability density functions.
**The cumulative distribution function is a single function that describes a random variable completely, whether discrete or continuous, and it gives the probability that the variable takes a value less than or equal to a given number.** If $F(x) = P(X \le x)$, then the cumulative distribution function is nondecreasing, starts at zero for very small values, and reaches one for very large values, and every probability of the distribution can be derived from it. For a discrete variable, the cumulative distribution function is a step function that jumps by the amount of the probability mass at each possible value, while for a continuous variable it is a smooth curve that is the integral of the density from the left tail. The cumulative distribution function is the foundation of the probability integral transform, which turns any continuous random variable into a uniform one, and of the quantile function, which is its inverse and gives the value below which a given proportion of the distribution lies. In a fab the cumulative distribution function answers the question of what fraction of wafers have a metrology reading below a specified limit, and the quantile function gives the value that a target fraction of the population does not exceed. The cumulative distribution function unifies the discrete and the continuous cases.
**The expected value is the center of a probability distribution, and it is the probability-weighted average of the possible values of the random variable.** For a discrete variable, the expected value is the sum of each value multiplied by its probability, while for a continuous variable it is the integral of the value times its density, and in both cases it is the long-run average of the variable over many repeated observations. The expected value of a sum is the sum of the expected values, and the expected value of a constant times a variable is the constant times the expected value, so that the expectation is a linear operation, and these rules make it the workhorse of probability calculations. The expected value is often written as the mean and denoted by the Greek letter mu, and it locates the distribution at the point around which the values balance. The expected value of a function of a random variable is the average of that function over the distribution, and this single idea underlies the calculation of all the other characteristics of a distribution. In a fab the expected value of a metrology reading is the target value that the process aims for, and the difference between the expected value and the specification center is the bias of the process. The expected value is the first and most important summary of a distribution.
**The variance measures the spread of a probability distribution about its mean, and it is the expected value of the squared deviation of the variable from its mean.** If the mean is $\mu$, then the variance is the expected value of $(X - \mu)^2$, which is large when the values are spread widely and small when they are concentrated near the mean, and the standard deviation is the square root of the variance, which is measured in the same units as the variable itself. The variance can also be written as the expected value of $X^2$ minus the square of the mean, and this computing form is often the easiest way to calculate it. The variance of a constant is zero, the variance of a constant times a variable is the constant squared times the variance, and the variance of a sum of independent variables is the sum of their variances, while the variance of a sum of dependent variables includes their covariances. The standard deviation is the natural measure of spread that the engineer uses to set tolerance and control limits, and the empirical rule states that about sixty-eight percent of a normal distribution lies within one standard deviation of the mean. In a fab the variance of a metrology reading measures the uniformity of the process, and reducing it is the central goal of process control. The variance and the standard deviation complete the description of the center and the spread of a distribution.
**The moment generating function is a powerful tool that encodes all the moments of a distribution, and it provides a compact way to compute the mean, the variance, and the higher moments and to derive the distributions of sums.** The moment generating function of a random variable $X$ is the expected value of $e^{tX}$ as a function of a real parameter $t$, and its derivatives at $t = 0$ are the moments of the distribution, so that the first derivative gives the mean and the second gives the second moment from which the variance is computed. The moment generating function has the remarkable property that the moment generating function of a sum of independent variables is the product of their individual moment generating functions, which makes it the natural tool for proving the central limit theorem and for identifying the distribution of a sum. When two random variables have the same moment generating function, they have the same distribution, so that matching a moment generating function to a known form identifies the distribution. The moment generating function exists only when the expected value of $e^{tX}$ is finite in an interval about zero, and where it exists it is a complete description of the distribution. In a fab the moment generating function is mostly a theoretical tool, but it is the engine behind many of the distribution results that the engineer relies on. The moment generating function unifies the calculation of distribution characteristics.
**The central limit theorem is the deep reason why the normal distribution appears so often, and it states that the sum or the average of many independent random variables is approximately normally distributed, regardless of the distributions of the individual variables, provided their variances are finite.** The approximation improves as the number of terms grows, and it holds for sums of variables from almost any distribution, so that the average of many small independent influences tends to be normal. The theorem is the justification for the normality assumption that underlies the control charts and the analysis of variance, and it explains why so many measured quantities in engineering, which are sums of many small effects, are approximately normally distributed. The theorem also gives the convergence of the standard deviation, so that the distribution of a standardized average approaches the standard normal distribution with mean zero and variance one. The central limit theorem is what makes the normal distribution the most important distribution in the entire subject, and it is the reason the engineer can apply the normal-based tools of the earlier statistics keywords so widely. In a fab the average of many defect measurements across a lot is approximately normal even when the individual measurements are not, so that the normal-based control charts apply to the lot averages. The central limit theorem is the bridge from any distribution to the normal distribution.
**The discrete families of distributions are the building blocks for counts and proportions, and the most important of them are the Bernoulli, the binomial, the geometric, and the Poisson distributions.** The Bernoulli distribution describes a single trial with two outcomes, a success with probability $p$ and a failure with probability $1 - p$, and it is the simplest of all distributions. The binomial distribution describes the number of successes in a fixed number of independent Bernoulli trials, so that its mean is $np$ and its variance is $np(1 - p)$, and it applies whenever the engineer counts how many of $n$ independent items have a given property. The geometric distribution describes the number of trials needed to obtain the first success, and the Poisson distribution describes the number of rare events in a fixed interval of time or space, with a single parameter $\lambda$ that is both its mean and its variance. The Poisson distribution is the natural model for the number of particles on a wafer or the number of defects in a chip, because such counts of rare, independent events follow it closely, and it is also the limit of the binomial distribution for rare events. In a fab the number of fails on a wafer and the number of particles are modeled by the Poisson and the binomial families. The discrete families turn counts into probabilities.
**The normal distribution is the most important of all the continuous distributions, and it is the bell-shaped distribution that is completely described by its mean and its standard deviation.** The normal density is a smooth, symmetric bell curve centered at the mean, with the spread controlled by the standard deviation, and it arises whenever many independent effects combine, by the central limit theorem. The standard normal distribution is the special case with mean zero and standard deviation one, and any normal random variable is converted to the standard normal form by subtracting its mean and dividing by its standard deviation, which is called the z-score. The probability that a normal variable falls within one, two, or three standard deviations of its mean is approximately sixty-eight, ninety-five, and ninety-nine point seven percent, and these figures are the basis of the control limits and the tolerance analysis of a fab. The normal distribution is the foundation of the analysis of variance, the t tests, and the confidence intervals of the inference statistics keyword, and it is assumed by most of the statistical methods of the series. In a fab the normal distribution models the natural variation of well-controlled metrology readings, and the control limits of a chart are set at plus and minus three standard deviations from the mean. The normal distribution is the centerpiece of continuous probability.
**The exponential distribution models the time between events in a Poisson process, and it is the continuous analogue of the geometric distribution, with the memoryless property that its remaining lifetime does not depend on how long the process has already run.** The exponential density has its peak at zero and decays exponentially, and its single rate parameter determines the average time between events, so that the mean and the standard deviation are both the reciprocal of the rate. The memoryless property is unique to the exponential distribution, and it states that the probability of surviving a further interval does not depend on the elapsed time, which makes the exponential the natural model for the time between independent random events. The exponential distribution is closely tied to the Poisson distribution, because the number of events in a fixed interval is Poisson while the times between them are exponential, and it is the basis of the reliability modeling that appears in the reliability keyword of the series. In a fab the exponential distribution models the time between particle adders or between tool failures, and it underlies the reliability and availability calculations. The exponential distribution is the bridge between the discrete Poisson counts and the continuous times of a process.
**The gamma distribution generalizes the exponential distribution to the time until a specified number of events occur, and it has two parameters that separately control its shape and its scale.** The gamma density is skewed to the right, and its shape parameter controls the skew, so that for small shape values it looks like the exponential and for larger shape values it becomes more symmetric and bell-shaped. The exponential distribution is the special case of the gamma distribution with a shape parameter of one, and the chi-square distribution is a further special case in which the shape parameter is half the degrees of freedom and the scale is two. The chi-square distribution arises as the distribution of the sum of squares of independent standard normal variables, and it is the basis of the tests of variance, of the goodness of fit, and of the analysis of contingency tables. The t distribution and the F distribution are also built from the chi-square and the normal distributions: the t distribution describes the ratio of a normal variable to an independent estimate of its standard deviation, and the F distribution describes the ratio of two independent chi-square variables divided by their degrees of freedom. In a fab the chi-square, t, and F distributions provide the reference tables for the hypothesis tests and the analysis of variance of the series. The gamma family is the engine of the sampling distributions of statistics.
**The beta distribution is a flexible distribution on the interval between zero and one, and it is the natural model for proportions and for the uncertainty about a probability.** The beta density has two shape parameters that control its form, and by changing them it can be flat, bell-shaped, J-shaped, or U-shaped, so that it can represent almost any pattern of a variable confined to a unit interval. The uniform distribution is the special case of the beta distribution with both shape parameters equal to one, and it gives every value in the interval equal probability. The beta distribution is the conjugate prior for the binomial proportion in the bayesian statistics of the series, which makes it central to the bayesian updating of a probability, and it also models random quantities that are constrained to a proportion, such as a yield. In a fab the beta distribution models the uncertainty about a yield or a proportion of good dies, and it summarizes the prior and posterior knowledge of a binomial success probability. The beta distribution rounds out the set of continuous families used in the series.
**The standard normal distribution provides the scale on which all normal probabilities are compared, and it is the reference distribution that turns any normal measurement into a comparable z-score.** The z-score of a measurement is the number of standard deviations that the measurement lies above or below the mean, and it is computed by subtracting the mean and dividing by the standard deviation, so that a positive z-score is above the mean and a negative one is below it. The z-score converts a measurement from the original units into a unitless scale on which the standard normal distribution with mean zero and variance one applies, and the probability of the measurement can then be read from the standard normal table. The z-score is the basis of the process capability analysis, because the fraction of the distribution outside the specification limits is computed from the z-scores of the upper and lower limits. In a fab the engineer converts a metrology reading to its z-score to find the fraction of the population that would fail a specification, and to compare measurements that are made in different units. The standard normal distribution is the universal reference for all normal probability calculations.
**The support of a distribution is the set of values that the random variable can actually take, and it is one of the first features that distinguishes one family from another. The support of the binomial and the Poisson distributions is the nonnegative integers, the support of the normal and the exponential distributions is the positive reals for the exponential and all reals for the normal, and the support of the beta distribution is the unit interval between zero and one. The support determines which family can model a given measurement, because a quantity that can never be negative cannot be modeled by a distribution with negative support, and a quantity confined to a proportion cannot be modeled by a distribution over the whole line. The parameters of a distribution control the details of its shape within its support, such as the center, the spread, and the skew, and changing a parameter changes the distribution in a predictable way. The engineer chooses a family by matching its support and its flexibility to the range and the shape of the measured quantity. The support and the parameters are the two features that identify a distribution family.
**The percentiles and the quartiles of a distribution are the values that divide the distribution into specified proportions, and they are the direct tools for tolerance and specification analysis.** The 100p-th percentile is the value below which a fraction p of the distribution lies, so that the median is the fiftieth percentile and the quartiles are the twenty-fifth and the seventy-fifth percentiles, and the interquartile range is the difference between the third and the first quartiles. The percentiles are the values of the inverse cumulative distribution function at the desired probabilities, and they are read directly from the cumulative distribution function or from the standard normal table for the normal family. The percentiles give the engineer the boundary values that contain a specified fraction of the population, which is exactly what a tolerance limit is, and the interquartile range is a robust measure of spread that is not distorted by outliers. In a fab the engineer sets the upper and lower specification limits at the percentiles that correspond to an acceptable fraction of out-of-spec wafers. The percentiles turn the cumulative distribution function into practical limits.
**The approximation relationships among the distributions let the engineer substitute a simpler family when the situation permits, and they are among the most useful facts about the families.** The Poisson distribution approximates the binomial distribution when the number of trials is large and the success probability is small, and the normal distribution approximates both the binomial and the Poisson distributions when their means are large, which is a consequence of the central limit theorem. The t distribution approaches the standard normal distribution as its degrees of freedom grow, and the F distribution and the chi-square distribution are connected through the fact that a chi-square variable with one degree of freedom is the square of a standard normal variable. These approximations are used throughout the series, because they let the engineer use the normal-based tools when the sample is large enough, and they explain why the normal distribution appears even in settings that begin with a discrete count. The approximations are justified mathematically, and they hold to a good degree of accuracy when the stated conditions are met. In a fab the engineer uses the normal approximation to a Poisson count of defects when the count is large, so that a normal control chart can monitor it. The approximation relationships bind the families into a coherent web.
**The skewness and the kurtosis are the higher moments that describe the shape of a distribution beyond its center and spread, and they distinguish distributions that have the same mean and variance but different shapes.** The skewness measures the asymmetry of a distribution, so that a positive skewness indicates a long right tail and a negative skewness a long left tail, and the exponential and the gamma distributions have positive skewness while the normal distribution has none. The kurtosis measures the weight of the tails relative to the normal distribution, so that a distribution with a high kurtosis has heavier tails and a higher peak, and the t distribution has a higher kurtosis than the normal distribution, which is why the t tests are more conservative for small samples. The skewness and the kurtosis are computed from the third and the fourth central moments, and they are the usual way to compare the shape of an observed histogram with the shape of a candidate family. In a fab a metrology distribution that is skewed suggests a process with a drift or a bounded response, and a distribution with heavy tails suggests occasional large excursions that the normal model would underestimate. The higher moments complete the description of a distribution's shape.
**The fitting of a distribution to data is the process of estimating its parameters so that the fitted family matches the observed histogram, and it is the step that turns data into a probability model.** The method of moments estimates the parameters by setting the theoretical moments of the distribution equal to the observed moments of the data, such as matching the sample mean to the theoretical mean and the sample variance to the theoretical variance. The maximum likelihood estimation estimates the parameters by choosing the values that make the observed data as probable as possible, and it is the preferred method because it has good statistical properties and it connects directly to the inference statistics keyword. The goodness of fit is then assessed by comparing the observed histogram with the fitted density, either graphically or with a formal test such as the chi-square test or the Kolmogorov-Smirnov test. In a fab the engineer fits a normal distribution to the metrology data to establish the baseline for a control chart, and fits a Weibull distribution to the lifetime data in the reliability analysis. The fitting of a distribution connects the observed data to the theoretical families.
**The Weibull distribution is the standard model for the lifetimes and the failure times of components, and it is the most important distribution in the reliability analysis of a fab.** The Weibull distribution has a shape parameter that controls whether the failure rate is decreasing, constant, or increasing over time, and it includes the exponential distribution as the special case of a constant failure rate. The Weibull distribution models the time to failure of tools, components, and films, and its hazard function describes how the risk of failure changes with age, which is the basis of the bathtub curve that describes the failure rate of electronic components over their lifetime. The Weibull distribution is fitted to failure data, and its shape parameter is estimated to reveal whether failures are dominated by infant mortality, random failures, or wear-out. In a fab the Weibull distribution underlies the reliability, availability, and maintainability analysis of the tools, and it is closely related to the reliability keyword of the series. The Weibull distribution extends the exponential model to the realistic case of a changing failure rate.
**The lognormal distribution models a variable whose logarithm is normally distributed, and it is the natural model for quantities that are positive and right-skewed, such as particle sizes and some resistances.** A variable has a lognormal distribution when its natural logarithm is normal, so that the variable itself is always positive and has a long right tail, and its parameters are the mean and the standard deviation of the underlying normal logarithm. The lognormal distribution arises as the product of many small independent positive factors, just as the normal distribution arises as the sum, and it is used when a quantity can grow multiplicatively. In a fab the lognormal distribution models the size distribution of particles and the distribution of some electrical parameters that are inherently positive. The lognormal distribution and the Weibull distribution are the two most important skewed positive families beyond the gamma, and together they cover the reliability and the particle distributions of a fab.
**The sampling distribution of the sample mean is the distribution of the average of a random sample, and it is the bridge from the probability distributions to the inference of the series. If the population is normally distributed, the sample mean is also normally distributed, with the same mean and a standard deviation equal to the population standard deviation divided by the square root of the sample size, and this is true exactly. If the population is not normal, the central limit theorem guarantees that the sample mean is approximately normal when the sample is large, and this approximation is the reason that the normal-based tools apply to the averages of the series. When the population standard deviation is unknown and estimated from the sample, the standardized sample mean follows the t distribution rather than the normal distribution, and the difference is important for small samples. The sampling distribution of the sample mean is the foundation of the confidence intervals and the hypothesis tests of the inference statistics keyword, because every test statistic is a function of the sample whose distribution is known. In a fab the sample mean of the metrology readings across a lot has a sampling distribution that is narrower than the individual readings, so that the lot average is a more precise estimate of the process mean. The sampling distribution is the probability distribution applied to statistics.
**The process capability analysis is the application of the normal distribution to the question of whether a process can meet its specifications, and it is one of the most direct uses of the probability distributions in manufacturing.** The capability index Cp is the ratio of the specification width to the natural variation of the process, and it measures how much room the process has relative to its limits, while the capability index Cpk also accounts for the centering of the process relative to the target. Both indices are computed from the z-scores of the specification limits, so that the fraction of the distribution outside the limits is found from the standard normal distribution, and a capability index of one corresponds to a process whose specification limits are six standard deviations apart. The probability distribution is what makes the capability analysis possible, because it converts the specification limits and the process spread into the expected fraction of defective wafers. In a fab the engineer uses the capability indices to decide whether a process can hold its specifications, and the Cp and the Cpk values are reported for every critical process. The capability analysis is the normal distribution made practical.
**The history of the probability distributions is the story of the statisticians whose names the families carry, and it connects the mathematics to the pioneers of the field.** Abraham de Moivre introduced the normal approximation to the binomial distribution and the normal curve itself in the eighteenth century, and Carl Friedrich Gauss applied the normal distribution to the theory of errors in astronomy, so that the normal is often called the Gaussian distribution. Siméon-Denis Poisson introduced the distribution that bears his name for the number of rare events, and Jacob Bernoulli developed the Bernoulli and the binomial families, while Adolphe Quetelet applied the normal distribution to social data. William Gosset, publishing under the name Student, developed the t distribution for the small-sample inference of the brewery where he worked, and Ronald Fisher developed the F distribution and the analysis of variance, while Karl Pearson developed the chi-square distribution and the family of curves that bears his name. The names on the distributions are the names of the scientists who understood the mathematics and the data of their time. The history shows that the probability distributions were built to answer real questions about real variation.
**The distribution families are organized by whether the variable is discrete or continuous and by what each family models, and the following table collects the principal distributions with their parameters, their support, and the quantities they describe.** The table makes it easy to choose the distribution that matches a given type of measurement, and it shows the symmetry between the discrete families for counts and the continuous families for measurements. The engineer reads the table by matching the kind of quantity to the distribution that models it.
| Distribution | Type | Parameters | Mean | Models |
|---|---|---|---|---|
| Bernoulli | discrete | p | p | single success/failure |
| Binomial | discrete | n, p | np | successes in n trials |
| Geometric | discrete | p | 1/p | trials until first success |
| Poisson | discrete | λ | λ | rare events in interval |
| Uniform | continuous | a, b | (a+b)/2 | equal chance on interval |
| Exponential | continuous | rate λ | 1/λ | time between events |
| Normal | continuous | μ, σ | μ | combined many effects |
| Gamma | continuous | shape, scale | shape·scale | time until k events |
| Chi-square | continuous | df | df | sum of squares of normals |
| t | continuous | df | 0 | normal / sd estimate |
| F | continuous | df1, df2 | df2/(df2−2) | ratio of variances |
| Beta | continuous | α, β | α/(α+β) | proportion on [0,1] |
**The choice of a probability distribution for a set of data is guided by the type of the variable and by the shape of its distribution, and the following flowchart routes a new measurement to the distribution family that most likely models it.** The first question is whether the variable is discrete or continuous, the second is the support of the variable, and the third is the shape of its observed histogram. Working through these questions selects a distribution family that matches the data.
```flowchart
A([New variable]) --> B{Discrete or continuous?}
B -- discrete count --> C{What is counted?}
C -- successes in n trials --> D[Binomial]
C -- rare events in interval --> E[Poisson]
C -- trials until first success --> F[Geometric]
B -- continuous measurement --> G{Where can it take values?}
G -- any real value --> H{Near bell shape?}
H -- yes --> I[Normal]
H -- no, skewed --> J{Right-skewed?}
J -- time between events --> K[Exponential]
J -- waiting time to k events --> L[Gamma]
G -- between 0 and 1 --> M[Beta]
G -- sum of squares of normals --> N[Chi-square]
```
**The connection between the probability distribution and the other keywords of the series is direct, and it completes the foundational layer that the entire statistics series builds upon.** The probability stats keyword supplies the rules of probability and the events that the distributions describe, while the statistics basics keyword supplies the descriptive measures of center and spread that the expected value and the variance generalize. The inference statistics keyword relies on the sampling distributions, such as the t, chi-square, and F distributions, to compute the p values and the confidence intervals, and the bayesian statistics keyword uses the beta and normal families as conjugate priors. The nonparametric statistics keyword applies when the data do not match any of the parametric families, and the experimental design keyword assumes the normal distribution in its analysis of variance, while the multivariate statistics keyword extends the normal distribution to the joint setting. The probability distribution is the common language in which all of these methods are expressed, because every statistical test and every control chart is a statement about a probability distribution. The engineer who masters the probability distribution can understand the assumptions behind every statistical tool in the series.
**A concrete example ties the distributions together and shows how they work in a fab, and the example of yield analysis illustrates the complete workflow.** The engineer counts the number of failing dies on a sample of wafers and models the count with the Poisson distribution, finding that the observed variation matches the Poisson mean, and uses the binomial distribution to model the yield, the fraction of good dies, on a wafer. The engineer measures the thickness of the film on many wafers, finds that the histogram is approximately bell-shaped, and models the thickness with the normal distribution, computing the z-scores of the specification limits to find the expected fraction of out-of-spec wafers. The engineer uses the t distribution to construct a confidence interval for the true mean thickness and the chi-square distribution to test whether the variance is within tolerance. The example shows that the probability distributions are not an abstract list but the working tools that quantify counts, yields, measurements, and uncertainties on every product. This single example shows how the choice of a probability distribution turns raw fab data into probabilities, tolerances, and decisions.
**The closing lens for probability distributions is that a distribution is not a list of formulas but the complete description of how a random quantity behaves, and the value of the subject is in matching the right family to the right measurement.** With this lens the engineer sees every count as a binomial or a Poisson mass, every measurement as a normal, exponential, or gamma density, every test statistic as a t, chi-square, or F distribution, and every proportion as a beta distribution, and sees the central limit theorem as the reason the normal distribution sits at the heart of it all. The mastery of the probability distributions is the mastery of translating raw variability into probability, which is precisely the translation that every control chart, every specification, and every statistical test in the series must make. Read probability distributions through a family-matching lens rather than a formula-collection lens.