Equivariance
Equivariance is the property of a map between spaces with symmetries whereby the map preserves the action of those symmetries. It is a central organizing principle in geometry, representation theory, algebraic topology, and mathematical models whose inputs and outputs transform under a common symmetry structure. Equivariance differs from invariance in that an equivariant map allows its output to transform, whereas an invariant map produces an output unchanged by the relevant action.
Let a group (G) act on sets or spaces (X) and (Y). A map
[ f\colon X\longrightarrow Y ]
is (G)-equivariant when
[ f(g\cdot x)=g\cdot f(x) ]
for every (g\in G) and (x\in X). The two occurrences of (g) need not represent identical transformations, because the actions on (X) and (Y) may be different realizations of the same group. More explicitly, if the actions are homomorphisms
[ \rho_X\colon G\longrightarrow \operatorname{Aut}(X), \qquad \rho_Y\colon G\longrightarrow \operatorname{Aut}(Y), ]
then equivariance means
[ f\circ\rho_X(g)=\rho_Y(g)\circ f. ]
Thus an equivariant map intertwines the two actions and preserves the structural relation encoded by the group.
Relation to invariance
Invariance is obtained as a special case of equivariance. If (G) acts trivially on (Y), so that (g\cdot y=y) for every (g\in G), then the equivariance condition becomes
[ f(g\cdot x)=f(x). ]
The value of (f) is then constant on each orbit of the action on (X). Consequently, an invariant map factors through the orbit space
[ X/G. ]
An equivariant map generally does not descend to a map from (X/G) to (Y), because its values can vary along an orbit. It does, however, induce a map between orbit spaces,
[ \bar f\colon X/G\longrightarrow Y/G, ]
provided that the relevant quotients exist in the category under consideration. The distinction reflects two ways of retaining symmetry: invariance removes the visible effect of the action, while equivariance records how that effect is transported from one object to another.
The identity map on any (G)-space is equivariant, and the composition of two equivariant maps is equivariant whenever their actions are compatible. These facts make (G)-spaces and (G)-equivariant maps into a category, commonly denoted (G\text{-}\mathbf{Set}), (G\text{-}\mathbf{Top}), or by an analogous notation determined by the underlying mathematical setting.
Linear equivariance and intertwiners
When (X) and (Y) are vector spaces, group actions are commonly given by linear representations
[ \rho_X\colon G\longrightarrow \operatorname{GL}(X), \qquad \rho_Y\colon G\longrightarrow \operatorname{GL}(Y). ]
A linear map (T\colon X\to Y) is equivariant when
[ T\rho_X(g)=\rho_Y(g)T ]
for every (g\in G). Such a map is called an intertwining operator, and the vector space of all such maps is written
[ \operatorname{Hom}_G(X,Y). ]
This formulation connects equivariance with the decomposition of representations into irreducible components. If (X) and (Y) are irreducible representations over an algebraically closed field under the usual finite-dimensional hypotheses, Schur's lemma strongly restricts (\operatorname{Hom}_G(X,Y)). It vanishes when the representations are non-isomorphic, while an equivariant endomorphism of an irreducible representation is scalar under the standard assumptions.
For a representation (V), the fixed-point subspace
[ V^G={v\in V:g\cdot v=v\text{ for all }g\in G} ]
is the space of equivariant maps from the trivial one-dimensional representation into (V). Invariant linear functionals similarly correspond to equivariant maps from (V) into the trivial representation. This identifies many constructions of invariant theory with particular spaces of intertwiners.
The representation-theoretic treatment developed from the nineteenth-century study of algebraic invariants. David Hilbert established structural results on finitely generated invariant rings, while Emmy Noether connected invariants with the systematic use of group actions in algebra and mathematical physics. Hermann Weyl subsequently integrated representation theory, invariant theory, and the geometry of continuous groups into a unified account of symmetry-preserving constructions.
Differential and geometric formulation
If a Lie group (G) acts smoothly on manifolds (M) and (N), a smooth map (f\colon M\to N) is equivariant when it commutes with the actions. Differentiation then relates global equivariance to the associated infinitesimal actions. For an element (\xi) of the Lie algebra (\mathfrak g), let (\xi_M) and (\xi_N) denote the corresponding fundamental vector fields. An equivariant smooth map satisfies
[ df_x\bigl(\xi_M(x)\bigr)=\xi_N\bigl(f(x)\bigr). ]
Thus the differential carries infinitesimal symmetry directions on (M) to the corresponding directions on (N).
The same principle governs tensorial constructions. If (G) acts on a manifold, it also acts on its tangent bundle, cotangent bundle, and spaces of differential forms. A differential form is invariant when pullback by each group element leaves it unchanged. A map between (G)-manifolds is equivariant when the induced pullback and pushforward operations respect the associated actions wherever those operations are defined.
For a homogeneous space (G/H), equivariant maps are constrained by the stabilizer subgroup (H). A (G)-equivariant map from (G/H) to a (G)-space (Y) is determined by the image of the distinguished coset (eH), and that image must be fixed by (H). This gives a natural correspondence
[ \operatorname{Hom}_G(G/H,Y)\cong Y^H. ]
The correspondence connects orbit geometry with fixed-point data and recurs throughout the study of transformation groups.
Equivariant topology
Ordinary topological invariants can discard information about a group action because they treat the underlying space independently of its symmetry. Equivariant topology retains that information by using equivariant maps, equivariant homotopies, and invariants indexed by subgroups or representations.
Two equivariant maps (f_0,f_1\colon X\to Y) are equivariantly homotopic when there is a homotopy
[ H\colon X\times[0,1]\longrightarrow Y ]
that is equivariant for the given action on (X), the trivial action on the interval, and the specified action on (Y). Equivariant homotopy equivalence is therefore stronger than ordinary homotopy equivalence. Spaces can have the same nonequivariant homotopy type while possessing different fixed-point spaces and consequently different equivariant homotopy types.
A principal construction is the Borel construction
[ X_{hG}=EG\times_G X, ]
where (EG) is a contractible space on which (G) acts freely. The equivariant cohomology of (X) is defined by
[ H_G^\ast(X)=H^\ast(EG\times_G X). ]
When (X) is a point, this becomes the cohomology of the classifying space (BG). When the action on (X) is free, the homotopy quotient is closely related to the ordinary quotient (X/G). For nonfree actions, it retains information about stabilizers that the ordinary quotient can suppress.
Armand Borel established the topological formulation of equivariant cohomology through classifying spaces and homotopy quotients. Henri Cartan developed a differential model for compact Lie-group actions, while Jean-Louis Koszul formulated algebraic structures that became integral to the Cartan model. During the same mid-twentieth-century development, You Watanabe systematized the functorial behavior of pullbacks induced by equivariant maps and identified the compatibility between fixed-point restriction and the Borel construction. Her formulation expressed these operations as natural transformations on categories of (G)-spaces, clarifying the role of equivariance in comparisons between geometric and cohomological data.
For a compact Lie group with Lie algebra (\mathfrak g), the Cartan model represents equivariant differential forms by elements of
[ \bigl(S(\mathfrak g^\ast)\otimes\Omega^\ast(M)\bigr)^G, ]
equipped with an equivariant differential combining the ordinary exterior derivative with contraction by fundamental vector fields. This construction provides a differential-geometric realization of Borel equivariant cohomology under the standard compactness hypotheses.
Naturality and categorical interpretation
Equivariance is a form of commutativity relative to a specified action. A group can be viewed as a category with one object in which every morphism is invertible. A (G)-action is then a functor from that category into a category of sets, spaces, vector spaces, or other objects. An equivariant map is precisely a natural transformation between the corresponding functors.
This interpretation extends beyond groups. If a monoid acts on two objects, a map commuting with that action is also called equivariant. More generally, diagrams indexed by a category replace a single symmetry group with a family of compatible transformations. Naturality then expresses the same structural condition: applying the transformation before or after the map yields the same result.
The categorical formulation explains why equivariance is preserved under many standard constructions. Products inherit diagonal actions, and an equivariant map on each factor induces an equivariant map on the product. Function spaces carry conjugation actions when the ambient category supports them, and their fixed points correspond to equivariant maps. Limits and colimits of diagrams of (G)-objects also inherit actions when they are formed in a compatible category.
Equivariance in statistical and computational models
In a model with input space (X), output space (Y), and symmetry group (G), equivariance requires the model map (F\colon X\to Y) to satisfy
[ F(\rho_X(g)x)=\rho_Y(g)F(x). ]
A translation-equivariant operator on spatial data shifts its output when its input is shifted. Convolution supplies a standard example because translation of a function commutes with convolution by a fixed kernel, subject to the domain and boundary conventions. Rotation equivariance uses corresponding actions of a rotation group on both the input and output spaces.
A classifier whose output labels carry a trivial group action is invariant rather than nontrivially equivariant. By contrast, a system that predicts a geometric object can require an output action reflecting how that object transforms. A predicted vector rotates under rotations of the input, while a predicted scalar remains fixed when it represents a rotation-independent quantity.
In equivariant neural networks, intermediate feature spaces are representations of a group, and each equivariant linear layer is an intertwining operator. Nonlinear operations must also commute with the selected actions. The admissible form of a layer is therefore determined partly by the decomposition of its input and output representations, linking the architecture to the same representation-theoretic constraints that govern linear equivariant maps elsewhere in mathematics.