Vector Operators
The Clebsch-Gordan decomposition tells us how a tensor product of irreducibles breaks
apart. The Wigner-Eckart theorem is the payoff: it says that once the irreducible content
is known, the way one set of operators relates to another is forced to be a single number.
The objects on which this rigidity acts are vector operators — triples of
operators that transform under rotations the way the three coordinates of a vector do. We
begin by making that phrase precise.
How Rotations Act on Operators
Fix a representation \(\Pi\) of \(SO(3)\) on a finite-dimensional vector space \(V\), and
let \(\operatorname{End}(V)\) be the space of linear operators \(C : V \to V\). A rotation
\(R\) already acts on the vectors of \(V\) through \(\Pi(R)\); it acts on an operator
\(C\) by transporting the operator along with the space,
\[
R \cdot C = \Pi(R)\,C\,\Pi(R)^{-1}.
\]
The reading is concrete: to apply the rotated operator to a vector, first undo the rotation
with \(\Pi(R)^{-1}\), apply \(C\) in the original frame, then rotate the result back with
\(\Pi(R)\). This is the same conjugation pattern by which a change of basis acts on a
matrix, and a direct check using \(\Pi(R_1 R_2) = \Pi(R_1)\Pi(R_2)\) shows it is itself a
representation of \(SO(3)\), now on the space \(\operatorname{End}(V)\):
\[
R_1 \cdot (R_2 \cdot C) = \Pi(R_1)\Pi(R_2)\,C\,\Pi(R_2)^{-1}\Pi(R_1)^{-1}
= (R_1 R_2) \cdot C.
\]
The Defining Equivariance
Alongside this action on operators sits the most familiar representation of \(SO(3)\) of
all: the rotation of ordinary vectors in \(\mathbb{R}^3\), where \(R\) sends
\(\mathbf{v}\) to \(R\mathbf{v}\). A vector operator is a triple
\(\mathbf{C} = (C_1, C_2, C_3)\) of operators on \(V\) whose three components rotate among
themselves in lockstep with the coordinate directions. To state this cleanly, assemble the
triple into a single operator depending linearly on a direction \(\mathbf{v} \in
\mathbb{R}^3\),
\[
\mathbf{v} \cdot \mathbf{C} = \sum_{j=1}^{3} v_j\, C_j.
\]
Definition: Vector Operator
Let \((\Pi, V)\) be a representation of \(SO(3)\). A triple
\(\mathbf{C} = (C_1, C_2, C_3)\) of operators on \(V\) is a vector
operator if
\[
(R\mathbf{v}) \cdot \mathbf{C} = \Pi(R)\,(\mathbf{v} \cdot \mathbf{C})\,\Pi(R)^{-1}
\qquad \text{for all } R \in SO(3) \text{ and all } \mathbf{v} \in \mathbb{R}^3.
\]
Unwinding the definition shows exactly what it demands. The right-hand side is the action
on operators applied to \(\mathbf{v} \cdot \mathbf{C}\); the left-hand side first rotates
the direction \(\mathbf{v}\) and then forms the combination. The equation says these agree:
rotating the direction and rotating the operator are the same thing. In the language of the
intertwining maps
of representations, the linear map \(\mathbf{v} \mapsto \mathbf{v} \cdot \mathbf{C}\) from
\(\mathbb{R}^3\) into \(\operatorname{End}(V)\) intertwines the rotation action on
\(\mathbb{R}^3\) with the conjugation action on \(\operatorname{End}(V)\). A vector operator
is precisely such an intertwiner; this single observation is what lets the representation
theory take over.
A concrete instance fixes the idea. Let \(V\) be a space of functions on \(\mathbb{R}^3\)
on which \(SO(3)\) acts by rotating the argument, \((\Pi(R) f)(\mathbf{x}) =
f(R^{-1}\mathbf{x})\), and let \(X_j\) be multiplication by the \(j\)-th coordinate,
\((X_j f)(\mathbf{x}) = x_j\, f(\mathbf{x})\). The triple \(\mathbf{X} = (X_1, X_2, X_3)\)
— the position operators of physics — is a vector operator: rotating the coordinate one
multiplies by is the same as conjugating the multiplication operator by the rotation, which
is the defining identity above.
The Infinitesimal Form
Differentiating the defining identity at the identity rotation converts it into a bracket
condition on the Lie algebra \(\mathfrak{so}(3)\), often the more practical test. Using the
basis \(E_1, E_2, E_3\) of \(\mathfrak{so}(3)\)
with \([E_j, E_k] = \sum_{l} \varepsilon_{jkl} E_l\), where \(\varepsilon_{jkl}\) is \(+1\)
on a cyclic permutation of \((1,2,3)\), \(-1\) on a non-cyclic permutation, and \(0\) when
an index repeats, the vector-operator condition becomes
\[
[\pi(E_j), C_k] = \sum_{l} \varepsilon_{jkl}\, C_l,
\]
where \(\pi\) is the Lie algebra representation associated to \(\Pi\). The components
\(C_k\) rotate under the infinitesimal generators exactly as the basis vectors of
\(\mathbb{R}^3\) rotate under the cross product — the same structure constants
\(\varepsilon_{jkl}\) appear, which is the algebraic shadow of the
correspondence between the cross product and the bracket on \(\mathfrak{so}(3)\).
Endomorphisms as a Representation
The proof of the Wigner-Eckart theorem turns the geometric statement about vector operators
into an algebraic one about a representation, and then lets Schur's lemma finish the job.
The bridge is an identification of the operator space \(\operatorname{End}(V)\) with a
tensor product, plus one special feature of the rotation group: its irreducibles are their
own duals. We establish both here.
Operators as a Tensor Product
Let \(V\) be a finite-dimensional representation of \(SO(3)\) and \(V^*\) its
dual representation.
There is a natural way to build an operator on \(V\) out of a covector and a vector: given
\(\varphi \in V^*\) and \(v \in V\), the rule
\[
w \;\longmapsto\; \varphi(w)\, v
\]
is a linear operator on \(V\) — it reads off the \(\varphi\)-component of its input and
returns that multiple of \(v\). This assignment \((\varphi, v) \mapsto \varphi(\,\cdot\,)v\)
is bilinear, so by the
universal property of the tensor product
it factors through a unique linear map
\[
\Psi : V^* \otimes V \longrightarrow \operatorname{End}(V),
\qquad \Psi(\varphi \otimes v)(w) = \varphi(w)\, v.
\]
Proposition (Endomorphisms as a Tensor Product)
Let \(V\) be a finite-dimensional representation of \(SO(3)\). The map \(\Psi\) above is
an isomorphism of representations
\[
\operatorname{End}(V) \;\cong\; V^* \otimes V,
\]
where \(SO(3)\) acts on \(\operatorname{End}(V)\) by conjugation,
\(R \cdot C = \Pi(R)\,C\,\Pi(R)^{-1}\), and on \(V^* \otimes V\) by the
tensor product of the dual and standard representations.
Proof:
First, \(\Psi\) is a vector-space isomorphism. Pick a basis \(v_1, \dots, v_n\) of \(V\)
with dual basis \(\varphi_1, \dots, \varphi_n\) of \(V^*\), so that
\(\varphi_i(v_k) = \delta_{ik}\). Then \(\Psi(\varphi_i \otimes v_j)\) is the operator
sending \(v_k\) to \(\delta_{ik} v_j\) — that is, the elementary matrix \(E_{ji}\) with
a \(1\) in row \(j\), column \(i\). As \((i, j)\) ranges over all pairs, these are
exactly the \(n^2\) elementary matrices, a basis of \(\operatorname{End}(V)\). So
\(\Psi\) carries a basis of \(V^* \otimes V\) to a basis of \(\operatorname{End}(V)\)
and is an isomorphism of vector spaces.
Next, \(\Psi\) intertwines the two actions. On the tensor product, a rotation acts by
\((\Pi^* \otimes \Pi)(R) = \Pi^*(R) \otimes \Pi(R)\) with \(\Pi^*(R) =
[\Pi(R^{-1})]^\top\). Applying \(\Psi\) to a rotated elementary tensor and evaluating on
\(w \in V\),
\[
\begin{align*}
\Psi\big((\Pi^*(R)\varphi) \otimes (\Pi(R) v)\big)(w)
&= (\Pi^*(R)\varphi)(w)\, \Pi(R) v \\\\
&= \varphi\big(\Pi(R)^{-1} w\big)\, \Pi(R) v \\\\
&= \Pi(R)\,\big[\varphi(\Pi(R)^{-1} w)\, v\big] \\\\
&= \Pi(R)\,\Psi(\varphi \otimes v)\big(\Pi(R)^{-1} w\big) \\\\
&= \big(\Pi(R)\,\Psi(\varphi \otimes v)\,\Pi(R)^{-1}\big)(w).
\end{align*}
\]
The second line uses the definition of the dual action, namely that
\((\Pi^*(R)\varphi)(w) = \varphi(\Pi(R)^{-1} w)\); the third pulls the scalar
\(\Pi(R)\) through. The result is the conjugation action on \(\operatorname{End}(V)\)
applied to \(\Psi(\varphi \otimes v)\). Thus \(\Psi\) intertwines, and being an
invertible intertwiner, it is an isomorphism of representations. \(\blacksquare\)
The Irreducibles Are Self-Dual
The second ingredient is special to the rotation group. The
irreducible representations of \(SO(3)\)
are, through the correspondence with \(\mathfrak{sl}(2;\mathbb{C})\), the spaces \(V_m\),
one in each odd dimension \(2\ell + 1\). Each is isomorphic to its own dual.
Proposition (Self-Duality of the Irreducibles)
Every finite-dimensional irreducible representation \(V\) of \(SO(3)\) is isomorphic to
its dual: \(V \cong V^*\).
Proof:
The dual \(V^*\) has the same dimension as \(V\), and passing to the dual
preserves irreducibility:
an invariant subspace of \(V^*\) is the annihilator of an invariant subspace of \(V\),
so \(V^*\) is irreducible exactly when \(V\) is. Passing to the complexified picture,
\(V\) and \(V^*\) are then both irreducible representations of
\(\mathfrak{sl}(2;\mathbb{C})\) of the same dimension. But the
classification of irreducibles
says any two irreducibles of \(\mathfrak{sl}(2;\mathbb{C})\) of equal dimension are
isomorphic. Hence \(V^* \cong V\). \(\blacksquare\)
Combining the two propositions gives the form in which the operator space enters the proof.
For an irreducible \(V\), self-duality lets us replace \(V^*\) by \(V\), so that
\[
\operatorname{End}(V) \;\cong\; V^* \otimes V \;\cong\; V \otimes V.
\]
The right-hand side is a tensor product of two irreducibles, and its irreducible content
is exactly what the
Clebsch-Gordan decomposition
computes. This is the lever: a question about operators has become a question about how the
standard three-dimensional representation sits inside \(V \otimes V\), which the previous
page already answered.
The Wigner-Eckart Theorem
Everything is now in place. On an irreducible representation, the operator space contains
only one slot into which a vector operator can map, and Schur's lemma forces any two vector
operators to fill that slot in proportion. The result is a rigidity statement of unusual
economy: a single complex number relates one vector operator to another.
Theorem (Wigner-Eckart, Irreducible Case)
Let \((\Pi, V)\) be an irreducible finite-dimensional representation of \(SO(3)\), and
let \(\mathbf{A}\) and \(\mathbf{B}\) be two vector operators on \(V\), with
\(\mathbf{A}\) nonzero. Then there is a constant \(c \in \mathbb{C}\) such that
\[
\mathbf{B} = c\, \mathbf{A}.
\]
Before the proof, it is worth saying what the statement buys. A vector operator on an
\(n\)-dimensional space is, on its face, three \(n \times n\) matrices — a great many
numbers. The theorem says that once one nonzero vector operator \(\mathbf{A}\) is
known on an irreducible \(V\), every other vector operator on \(V\) is a scalar multiple of
it. All the apparent freedom collapses to one number.
Proof
Proof:
Recall that a vector operator is, by definition, an
intertwining map
\[
\phi_{\mathbf{A}} : \mathbb{R}^3 \longrightarrow \operatorname{End}(V),
\qquad \phi_{\mathbf{A}}(\mathbf{v}) = \mathbf{v} \cdot \mathbf{A},
\]
from the rotation action on \(\mathbb{R}^3\) to the conjugation action on
\(\operatorname{End}(V)\), and likewise \(\phi_{\mathbf{B}}\) for \(\mathbf{B}\). Both
extend complex-linearly to intertwiners out of the complexification \(\mathbb{C}^3\),
which carries the unique three-dimensional irreducible representation of \(SO(3)\) —
the irreducible \(V_2\) in our labelling. To compare \(\phi_{\mathbf{A}}\) and
\(\phi_{\mathbf{B}}\) we examine where such an intertwiner can land.
By the previous section, \(\operatorname{End}(V) \cong V^* \otimes V \cong V \otimes V\),
the last step using the
self-duality of the irreducible \(V\).
The
Clebsch-Gordan decomposition
of \(V \otimes V\) writes it as a direct sum of irreducibles in which the
three-dimensional irreducible \(\mathbb{C}^3\) occurs at most once — it occurs
exactly once when \(V\) is nontrivial, and not at all when \(V\) is the trivial
representation. This single-occurrence is the multiplicity-freeness established on the
previous page, read off the decomposition.
Suppose first that \(V\) is trivial. Then \(\operatorname{End}(V)\) contains no copy of
\(\mathbb{C}^3\), so the intertwiner \(\phi_{\mathbf{A}} : \mathbb{C}^3 \to
\operatorname{End}(V)\) maps an irreducible into a representation containing no copy of
it; by the
first part of Schur's lemma
it must be zero. The same holds for \(\phi_{\mathbf{B}}\), so \(\mathbf{A} = \mathbf{B}
= 0\), contradicting that \(\mathbf{A}\) is nonzero. Hence \(V\) is nontrivial, and
\(\operatorname{End}(V)\) contains exactly one copy of \(\mathbb{C}^3\); call it \(U\).
Now \(\phi_{\mathbf{A}}\) maps the irreducible \(\mathbb{C}^3\) into
\(\operatorname{End}(V)\). By Schur's lemma its image is either zero or an irreducible
copy of \(\mathbb{C}^3\); since the only such copy is \(U\), the map \(\phi_{\mathbf{A}}\)
lands in \(U\), and likewise \(\phi_{\mathbf{B}}\). Because \(\mathbf{A}\) is nonzero,
\(\phi_{\mathbf{A}}\) is a nonzero intertwiner \(\mathbb{C}^3 \to U\) between
isomorphic irreducibles. Applying
the part of Schur's lemma comparing two intertwiners
to the pair \(\phi_{\mathbf{B}}, \phi_{\mathbf{A}} : \mathbb{C}^3 \to U\), there is a
constant \(c \in \mathbb{C}\) with \(\phi_{\mathbf{B}} = c\, \phi_{\mathbf{A}}\).
Evaluating on \(\mathbf{v} \in \mathbb{R}^3\) gives \(\mathbf{v} \cdot \mathbf{B} =
c\,(\mathbf{v} \cdot \mathbf{A})\) for every direction, which is precisely
\(\mathbf{B} = c\, \mathbf{A}\). \(\blacksquare\)
One Number for the Whole Operator
The force of the theorem is that the proportionality constant \(c\) is the only datum
not fixed by symmetry. Two vector operators built from entirely different physical
quantities — say a position and a momentum, restricted to an irreducible subspace —
must be scalar multiples of each other there, with the multiple depending on the
subspace but on nothing else. The geometry of rotation has used up every other degree
of freedom. The next section extends this from a single irreducible to operators
between two different irreducible pieces, where the same mechanism reappears as a rule
about matrix entries.
Matrix Elements and Equivariance
The irreducible case assumed the whole space \(V\) was irreducible. In applications the
interesting operators act on a large, reducible space and we care about how they connect
two irreducible pieces of it. The general theorem handles this, and the conclusion is the
one that organizes computation: between any two fixed irreducible subspaces, the entries of
every vector operator are the same up to a single constant.
From Endomorphisms to Homomorphisms
The bridge from the previous section generalizes from operators on one space to
homomorphisms between two. If \(W_1\) and \(W_2\) are finite-dimensional representations of
\(SO(3)\), the space \(\operatorname{Hom}(W_2, W_1)\) of linear maps \(W_2 \to W_1\)
carries the conjugation-type action \(C \mapsto \Pi(R)\, C\, \Pi(R)^{-1}\), and the same
construction as before identifies it with a tensor product.
Proposition (Homomorphisms as a Tensor Product)
Let \(W_1\) and \(W_2\) be finite-dimensional representations of \(SO(3)\). Then there
is an isomorphism of representations
\[
\operatorname{Hom}(W_2, W_1) \;\cong\; W_2^* \otimes W_1.
\]
If in addition \(W_2\) is irreducible, hence
self-dual, this
becomes \(\operatorname{Hom}(W_2, W_1) \cong W_2 \otimes W_1\).
The proof repeats that of the endomorphism case verbatim, with the rule
\(\Psi(\varphi \otimes w_1)(w_2) = \varphi(w_2)\, w_1\) for \(\varphi \in W_2^*\) and
\(w_1 \in W_1\); taking \(W_1 = W_2 = V\) recovers
the endomorphism isomorphism.
The General Theorem
Theorem (Wigner-Eckart)
Let \(V\) be an inner product space, possibly infinite-dimensional, carrying a
representation \(\Pi\) of \(SO(3)\) that preserves the inner product. Let \(W_1\) and
\(W_2\) be finite-dimensional irreducible invariant subspaces of \(V\), and let
\(\mathbf{A}\) and \(\mathbf{B}\) be vector operators on \(V\) for which
\(\langle w, A_j w' \rangle\) is nonzero for some \(w \in W_1\), \(w' \in W_2\), and
\(j \in \{1,2,3\}\). Then there is a constant \(c \in \mathbb{C}\) such that
\[
\langle w, B_j w' \rangle = c\, \langle w, A_j w' \rangle
\]
for all \(w \in W_1\), all \(w' \in W_2\), and all \(j = 1, 2, 3\).
Proof:
The operators \(A_j\) need not carry \(W_2\) into \(W_1\), but taking the inner product
of \(A_j w'\) against an element of \(W_1\) records only the \(W_1\)-component of
\(A_j w'\). Let \(P_1 : V \to W_1\) be the orthogonal projection onto \(W_1\); it
exists because \(W_1\) is finite-dimensional, and since \(W_1\) is invariant and
\(\Pi\) preserves the inner product, the complement \(W_1^\perp\) is invariant too, so
\(P_1\) commutes with every \(\Pi(R)\). Define
\[
\phi_{\mathbf{A}} : \mathbb{R}^3 \longrightarrow \operatorname{Hom}(W_2, W_1),
\qquad \phi_{\mathbf{A}}(\mathbf{v})(w') = P_1\big((\mathbf{v} \cdot \mathbf{A})\, w'\big),
\]
and likewise \(\phi_{\mathbf{B}}\). Because \(\mathbf{A}\) is a vector operator and
\(P_1\) commutes with the action, \(\phi_{\mathbf{A}}\) is an
intertwining map
from \(\mathbb{C}^3\) into \(\operatorname{Hom}(W_2, W_1)\).
Now \(\operatorname{Hom}(W_2, W_1) \cong W_2 \otimes W_1\) by the proposition above, and
the
Clebsch-Gordan decomposition
of \(W_2 \otimes W_1\) contains the three-dimensional irreducible \(\mathbb{C}^3\) at
most once. If it does not occur, both \(\phi_{\mathbf{A}}\) and \(\phi_{\mathbf{B}}\)
vanish by the
first part of Schur's lemma,
forcing all the inner products \(\langle w, A_j w'\rangle\) to be zero, against the
hypothesis. So \(\mathbb{C}^3\) occurs exactly once, and both \(\phi_{\mathbf{A}}\) and
\(\phi_{\mathbf{B}}\) map into that single copy. Since \(\phi_{\mathbf{A}}\) is nonzero,
comparison of two intertwiners
gives a constant \(c\) with \(\phi_{\mathbf{B}} = c\, \phi_{\mathbf{A}}\).
Finally, the orthogonal projection \(P_1\) is self-adjoint and acts as the identity on
\(W_1\), so for \(w \in W_1\),
\[
\langle w,\, (\mathbf{v} \cdot \mathbf{A})\, w' \rangle
= \langle w,\, P_1 (\mathbf{v} \cdot \mathbf{A})\, w' \rangle
= \langle w,\, \phi_{\mathbf{A}}(\mathbf{v})\, w' \rangle,
\]
and the same with \(\mathbf{A}\) replaced by \(\mathbf{B}\). Therefore
\(\langle w, (\mathbf{v} \cdot \mathbf{B}) w' \rangle = c\, \langle w, (\mathbf{v} \cdot
\mathbf{A}) w' \rangle\) for every \(\mathbf{v}\); taking \(\mathbf{v} = e_j\) yields the
claim. \(\blacksquare\)
Reduced Matrix Elements and Weight Sharing
The constant \(c\) is what physics calls the reduced matrix element. Read as a
statement about computation, the theorem says that the full array of numbers
\(\langle w, A_j w' \rangle\), as \(w\), \(w'\), and \(j\) vary over a pair of
irreducible subspaces, is determined by one scalar once the geometry-fixed part is
known. The dependence on the indices is universal — the same for every vector operator
between that pair of irreducibles — and only the overall scale carries information
about the specific operator.
This is the principle a rotation-equivariant network exploits when it couples features.
A layer mapping type-\(\ell_2\) features to type-\(\ell_1\) features through a vector
quantity has, a priori, many weights; equivariance forces those weights to follow the
fixed index pattern, leaving a single free parameter for each pair of feature types.
The reduction from a full weight matrix to one learnable scalar per irreducible pair is
the sample-efficiency advantage of equivariant models stated in exact terms, and it is
Schur's lemma — not an architectural choice — that does the reducing.
The setting also accommodates the infinite-dimensional spaces of analysis. When \(V\) is a
space of square-integrable functions on \(\mathbb{R}^3\) under the rotation action, its
decomposition into finite-dimensional irreducible subspaces is the harmonic content of the
sphere, and the matrix entries of any vector operator between two such pieces again reduce
to a single constant. Pursuing that decomposition leads to the theory of harmonic analysis
on the rotation group, where the irreducibles classified here reappear as the building
blocks of every square-integrable function.