Must every group contain numbers?

No. Its elements can be permutations, matrices or transformations, among other objects.

Must every pair of elements commute?

No. Groups with that extra property are called abelian or commutative groups.

Is every collection of transformations a group?

No. It must satisfy closure, associativity, identity and inverse requirements under the chosen composition rule.

AI Glossary / Mathematics and optimization / Intermediate

Group

Definition

In mathematics, a group is a set together with a binary operation that is closed and associative, has an identity element, and gives every element an inverse. Groups provide a precise language for combining reversible operations, including many symmetries. This mathematical meaning differs from an ordinary collection of people or data points. The operation is essential: the same set can satisfy the group rules under one operation and fail them under another.

What it means

Closure means combining two allowed elements produces another allowed element. Associativity means changing parentheses does not change the result of combining three elements. The identity leaves every element unchanged, and an inverse reverses an element's effect. Groups need not be commutative: changing the order of two operations can change the result. In AI, groups are useful for describing transformations that a model should respect, such as rotations or permutations, when those transformations match the task's real structure.

How it works

Specify a set G and an operation that maps pairs of its elements back into G. Check the axioms for all relevant elements, not only a few examples. For transformation groups, the operation is usually composition: apply one transformation and then another. A group action describes how abstract group elements transform objects such as points or images. A model is invariant when a specified transformation leaves its output unchanged; it is equivariant when transforming the input produces the corresponding transformation of the output. These are model properties relative to a chosen group action, not consequences of the word group alone.

Three practical examples

A shape-recognition study can investigate whether classification should be unchanged under a specified set of rotations.
A model processing an unordered collection can be designed to respect permutations, so arbitrary input ordering does not create unintended changes in its meaning.
A geometry pipeline can use rotation composition and inverses to describe how coordinates move between consistently defined orientations.

Think of it this way

Think of reversible moves of a puzzle that can be performed in sequence, including doing nothing and undoing a move. The group rules describe how those moves combine. The analogy does not mean every puzzle action is reversible or every collection of allowed moves is closed: a restricted menu that omits the result of composing two moves may fail to form a group.

Worked example

Consider rotations of a square about its center by 0, 90, 180 and 270 degrees. Label them r_0, r_1, r_2 and r_3, where the subscript counts quarter-turns. Composition follows r_a composed with r_b = r_((a+b) mod 4). Closure holds because the remainder is always 0, 1, 2 or 3. Associativity follows from addition modulo 4. The identity is r_0. Inverses are r_0 for itself, r_3 for r_1, r_2 for itself and r_1 for r_3. For example, r_3 composed with r_2 equals r_1 because five quarter-turns have the same effect as one. This four-element rotation group is commutative. To see that groups need not commute, also consider planar reflection S(x, y) = (x, −y) and rotation R(x, y) = (−y, x). At point (1, 0), applying S then R gives (0, 1), while applying R then S gives (0, −1). The order matters. Rotations and reflections preserving the square together form a larger group; the differing results do not violate associativity, which concerns parentheses rather than swapping order.

Illustrative case study

In a fictional vision project, Willow Shapes studies images of components that may appear in any quarter-turn orientation. Engineers define the four rotations explicitly and test whether the desired label should remain unchanged. They find that orientation is irrelevant for one inspection task but essential for another that identifies assembly direction. The group description helps clarify the modeling assumption before choosing augmentation or architecture. The case does not claim that imposing invariance always improves accuracy.

Common misconceptions and mistakes

Associative and commutative are different properties. Inverses must belong to the specified set, so nonnegative integers under addition are not a group, while all integers under addition are. Ordinary image cropping is not generally invertible and therefore does not fit a transformation group in the same way as ideal rotations. A mathematical rotation of a continuous image and a resampled pixel operation can differ because discretization loses information. Data augmentation can expose a model to transformed examples without proving exact invariance or equivariance. Finally, a symmetry assumption can be wrong for the intended labels.

Business use

Use group language when transformation structure matters to the task. Ask which transformations are permitted, how they compose and whether they preserve the target or should transform it. Test the assumption with domain experts before building it into a model. Distinguish ideal mathematical actions from their finite-resolution implementation. Evaluate performance on relevant transformed inputs and document exceptions. For nontechnical stakeholders, explain the practical behavior desired, such as treating reorderings consistently, while retaining the formal group definition in the technical specification where it supports verification.

Check your understanding

Why does swapping a rotation and reflection and obtaining different results not disprove that they belong to a group?

Answer: Commutativity is not a group requirement. Associativity concerns how three operations are parenthesized, not whether two operations can be exchanged.

Related terms

Permutations
group action
symmetry
permutation
invariance
equivariance
identity element

Learn next

construct the four-by-four composition table for the square's quarter-turn rotations. Verify the inverse of each element. Then test two simple permutations in both orders and explain why order sensitivity can coexist with the group axioms.

Sources

Sources: David Vogan, MIT 18.704 supplementary notes, Groups and actions, https://math.mit.edu/~dav/groups.pdf ; MIT OpenCourseWare, Modern Algebra, Groups, https://ocw.mit.edu/courses/18-703-modern-algebra-spring-2013/faeeeaf01f838584ce2172cfd85cf74d_MIT18_703S13_pra_l_1.pdf . Vogan's definitions and transformation examples were reviewed. The square calculations and fictional vision case are original; transformation behavior is stated as a modeling choice.

Previous: Grounding Next: Group attribution bias