Proposition. Scalar descent and transport of Gaussian observables [fdip-000Y]

On the enriched four-cycle groupoid, the trace ratio, with the cycle values, is a scalar on isomorphism classes, while covariance and effective precision are transported operators on their specified field spaces. For \(x=(a,b,m,R)\) write \(C_x=Q_x^{-1}\). If \(g:x\to y\), then \[\begin {aligned} C_y&=P_gC_xP_g^T,\\ (P_g)_*\mu _{Q_x}&=\mu _{Q_y},\\ \langle F\rangle _{Q_y}&=\langle F\circ P_g\rangle _{Q_x}. \end {aligned}\] The last identity holds for Borel \(F\in L^1(\mu _{Q_y})\); integrability is equivalent to that of \(F\circ P_g\) in the source law. These statements use the inverse-precision covariance and Gaussian transport, including their normalization.

Let \(D=V\setminus R\). In retained/eliminated coordinates write \[\begin {aligned} Q_x&=\begin {pmatrix}A&H\\H^T&C_D\end {pmatrix},\\ S_R(x)&=A-HC_D^{-1}H^T. \end {aligned}\] This Schur complement is the positive precision of the marginal. The restrictions of \(g\) give permutation isometries \(P_R:\mathbb R^R\to \mathbb R^{gR}\) and \(P_D:\mathbb R^D\to \mathbb R^{gD}\). Then \[\begin {aligned} S_{gR}(y)&=P_RS_R(x)P_R^T,\\ S_{gR}(y)^{-1}&=P_RS_R(x)^{-1}P_R^T,\\ \pi _{gR}P_g&=P_R\pi _R. \end {aligned}\] Thus the marginal laws are compatible on the changing retained spaces.

For the nonzero positive semidefinite \(L_x\), define \[\mathcal R(x)=\frac {\operatorname {tr}(L_x^2)} {\bigl (\operatorname {tr}L_x\bigr )^2}.\] Then \(\mathcal R(gx)=\mathcal R(x)\), and \[\mathcal R(a,b)=\frac {a^2+ab+b^2}{2(a+b)^2}.\] This scalar descends to the coarse orbit set but does not classify the enriched objects.

Now forget mass, state and partition and admit positive scaling of the Laplacian, as in the broader comparison: \[L'=cP_gLP_g^T,\qquad c>0.\] The numerator of \(\mathcal R\) gains \(c^2\), while its denominator gains the same factor. Thus the ratio is invariant under this different relation. A weight-swapping relabeling followed by \(c=1/t\) sends \(L(1,t)\) to \(L(1,1/t)\). Reciprocal parameters are identified in this reduced problem only; no scale arrow has been added to the original Gaussian groupoid. At fixed mass, \(cL+m^2I\ne cQ\) for \(c\ne 1\). If one instead rescales the entire precision to \(cQ\), then \[\begin {aligned} (cQ)^{-1}&=c^{-1}Q^{-1},\\ cA-(cH)(cC_D)^{-1}(cH)^T&=cS_R. \end {aligned}\] Raw covariance and effective precision therefore are not scale invariants. Any normalized operator ratio would need its own definition and scaling proof.

Mathlib's Matrix.trace_units_conj and Matrix.trace_smul support the trace conjugacy and scaling identities. Matrix.reindexLinearEquiv and Matrix.reindexRingEquiv supply coordinate infrastructure. TauCeti's TauCeti.Probability.map_affine_multivariateGaussian allows positive semidefinite covariance and an arbitrary rectangular linear map. Set its mean and translation to zero, its covariance to \(Q_x^{-1}\), and its linear map to \(P_g\) or \(\pi _R\). The precision-density identification and the Schur and groupoid computations above remain separate obligations.