Is discriminant analysis limited to two classes?

No. Standard methods can handle multiple predefined classes.

Is it the same as logistic regression?

No. Logistic regression directly models conditional class probabilities, while Gaussian discriminant analysis models class-conditional predictor distributions and priors.

Does linear mean there is only one predictor?

No. Linear boundaries can separate classes in a space with many predictors.

AI Glossary / Statistics and probability / Intermediate

Discriminant analysis

Definition

Discriminant analysis is a family of statistical methods for distinguishing predefined groups using measured predictors. It can construct classification rules and, in some forms, identify combinations of variables that separate groups. Classical linear discriminant analysis models class-specific means with a shared covariance matrix; quadratic discriminant analysis allows separate covariance matrices. Despite the historical source glossary's two-group description, these methods also extend to multiple groups. They require labeled groups and are distinct from unsupervised clustering.

What it means

The central problem is to use known examples from different classes to classify a new observation. A manufacturer might distinguish product types using measurements, or a researcher might compare patterns across predefined populations. In Gaussian discriminant models, the method estimates how predictors are distributed within each class and combines those distributions with class prior probabilities. It is therefore a generative classification approach in statistical terminology, even though its purpose is discrimination among classes. The phrase does not refer to social unfairness or to every possible discriminative machine-learning model.

How it works

Define the classes, predictors and intended population. Estimate class means, covariance structure and prior probabilities from an appropriate training dataset or justified external information. Under the Gaussian formulation, combine the class-conditional densities and priors through Bayes' rule to obtain posterior class probabilities. Choose a class using a decision rule that reflects the task and error costs. Equal covariance matrices lead to linear boundaries; allowing different matrices generally yields quadratic boundaries. Assess estimation stability, especially when predictors are numerous or strongly correlated. Validate all preprocessing and model choices on data excluded from fitting, and evaluate classification quality and probability calibration separately.

Three practical examples

Manufacturing: classify material types using measured physical properties and verified training labels.
Pattern recognition: use linear discriminant analysis as a baseline for distinguishing labeled categories with numerical features.
Exploratory analysis: examine supervised projections that summarize separation among predefined groups, while checking out-of-sample behavior.

Think of it this way

Think of drawing maps of where different groups tend to lie in a measurement space. A new point is compared with those maps and with how common the groups are expected to be. The decision depends on both location and spread, so the nearest group mean is not always the most likely class.

Worked example

Suppose a hypothetical single predictor follows a normal distribution within each of two classes. Class A has mean two, class B mean four, both have variance one, and their prior probabilities are equal. The log posterior odds of B versus A are 2x - 6. At x = 3.5, the log odds equal one and the model's posterior probability of B is 1/(1 + exp(-1)), approximately 0.7311. The equal-cost classification boundary is x = 3. If the priors instead become 0.75 for A and 0.25 for B, the log odds gain log(1/3), so the boundary becomes [6 + log(3)]/2, approximately 3.5493. At 3.5 the preferred class then changes to A. This demonstrates the effect of priors under specified distributions, not fitted performance on real data.

Illustrative case study

Fictional case: Linden Materials compares a linear discriminant classifier with other approaches for identifying labeled material batches. Analysts inspect class-specific distributions, check the number of observations available to estimate covariance and reserve separate batches for validation. When the operational class mix differs from the training sample, they review prior probabilities rather than assuming the sample proportions apply unchanged. This invented case illustrates model evaluation and deployment context. It does not establish that Gaussian assumptions hold or that a discriminant model is the best classifier for the task.

Common misconceptions and mistakes

Do not assume the pooled dataset must be normally distributed: the classical Gaussian assumption concerns predictors within classes. Quadratic analysis is more flexible but estimates more covariance parameters and can be unstable with limited data. A visually strong separation in training data may reflect overfitting or leakage. LDA can also mean latent Dirichlet allocation in another context, so expand the abbreviation. Classification probabilities rely on model and prior assumptions, and a method's ability to distinguish groups does not prove that the predictors cause group membership.

Business use

Businesses use discriminant analysis as an interpretable classification baseline and sometimes as a supervised dimension-reduction tool. The approach can be useful when numerical measurements and reliable class labels are available, but its suitability depends on data geometry and sample size. Regularized covariance estimation may help in some settings, without eliminating the need for validation. Assess errors that matter operationally, including rare classes and unequal costs, and revisit performance when populations shift. A single accuracy figure does not establish reliable probabilities or equitable performance across relevant groups.

Check your understanding

Why does changing the priors move the boundary in the example even though the class means and variances stay fixed?

Answer: Posterior probabilities combine the likelihood of the measurement within each class with prior class probabilities. Making A more common requires stronger measurement evidence before choosing B under the same equal-cost rule.

Related terms

Linear discriminant analysis
Classification
Covariance matrix
Logistic regression
Regularization
Probability calibration
quadratic discriminant analysis
class prior
Bayes' rule
generative classifier

Learn next

derive the two-class Gaussian decision boundary and explore the effect of unequal priors and error costs. Compare shared and separate covariance models on held-out observations. Then distinguish supervised discriminant projections from principal components, which seek variance rather than separation using class labels.

Decision boundary

Sources

David Howell, University of Vermont, statistics glossary: https://www.uvm.edu/~statdhtx/fundamentals9/Glossary/Glossary.html Penn State, Discriminant Analysis: https://online.stat.psu.edu/stat505/Lesson10 scikit-learn, Linear and Quadratic Discriminant Analysis: https://scikit-learn.org/stable/modules/lda_qda.html

Previous: Discretization Next: Discrimination