Canonical: https://knowlpedia.com/ergodic-theory/linear-gaussian-dynamics/

# Linear Gaussian-preserving dynamics

The linear maps preserving standard finite-dimensional Gaussian measure are orthogonal, and none is ergodic.

For \(d\geq1\), a real linear map \(T_Q(x)=Qx\) preserves standard Gaussian probability \(\gamma_d\) if and only if \(Q\in O(d)\). None of these transformations is [ergodic](https://knowlpedia.com/ergodic-theory/ergodic-transformation/) on the full Gaussian probability space. An affine map \(x\mapsto Qx+b\) preserves \(\gamma_d\) exactly when \(b=0\) and \(Q\in O(d)\).

## Proof

Preserving mean and covariance forces \(b=0\) and \(QQ^{\mathsf T}=I_d\). Conversely, an orthogonal map preserves the norm, the Gaussian density, and Lebesgue measure. Every such map fixes the ball \(\{\|x\|<1\}\), whose Gaussian probability lies strictly between zero and one. Hence ergodicity fails.

## Koopman interpretation

For linear observables \(f_v(x)=v^{\mathsf T}x\), forward composition gives \(U_Qf_v=f_{Q^{\mathsf T}v}\). This finite-dimensional subrepresentation can have no fixed vectors while the full Koopman representation has one: \(r(x)=\|x\|^2-d\) is a nonzero mean-zero fixed vector. The bounded function \(e^{-\|x\|^2}\) also witnesses a nontrivial fixed-point algebra.

Thus studying only coordinate observables can miss an invariant nonlinear quantity.
