Irreducible Representation

An irreducible representation is a basic transformation block that cannot be decomposed into smaller independent representations.

For a group action represented by $D(g)$, a feature transforms as:

$$h^{\prime }=D(g)h$$

Irreducible representations organize channels by transformation behavior, such as scalar or vector channels.

For 3D rotations, irreducible representations of $SO(3)$ are indexed by degree $\ell$:

$$h_{\ell }^{\prime }=D^{\ell }(R)h_{\ell },R\in SO(3)$$

The channel group $h_{\ell }$ has dimension $2\ell +1$. Common cases are:

Degree	Feature Type	Transform Behavior
$\ell =0$	scalar	invariant under rotation
$\ell =1$	vector-like feature	rotates like a 3D direction
$\ell \ge 2$	higher-order tensor feature	carries angular patterns beyond vectors

An equivariant model keeps these feature types separate unless an operation is allowed by the group representation.

Mixing Rule

If a layer combines two irreducible features, the output degrees are constrained by angular momentum coupling:

$${\ell }_{\mathrm{out}}\in \{|{\ell }_1-{\ell }_2|,\dots ,{\ell }_1+{\ell }_2\}$$

This is why tensor products and Clebsch-Gordan coefficients appear in many $SO(3)$-equivariant architectures. The purpose is not mathematical decoration; it prevents arbitrary channel mixing that would break equivariance.

Invariant Readout

If the final target is a scalar such as energy, affinity, class probability, or ranking score, the final readout usually needs $\ell =0$ output:

$$\hat{y}=\rho (h_{\ell =0})$$

If the target is a coordinate update, force, direction, or vector field, the output should transform equivariantly, often as $\ell =1$:

$${\hat{v}}^{\prime }=R\hat{v}$$

Why It Matters

• Equivariant networks use irreducible representations to organize feature channels by transformation behavior.
• Scalars, vectors, and higher-order features can be tracked explicitly instead of mixed arbitrarily.
• This vocabulary clarifies how SO(3) and E(3) models handle symmetry.
• It makes the output contract explicit: scalar outputs should be invariant, while vector or coordinate outputs should transform predictably.

Checks

• Which feature types are present: scalars, vectors, or higher-order channels?
• Are operations allowed to mix representations only in symmetry-respecting ways?
• Does the architecture need high-order features, or are scalar/vector features enough?
• Are tensor products and nonlinearities the computational bottleneck?

Related

• Equivariance
• SO(3)
• E(3)
• Spherical harmonics