Global Convergence and Boundary Invariance of the Riemannian Bures Gradient Flow for d = 2 and d = 3
Gervaseus Nerik
Abstract
We extend the Riemannian Bures gradient-flow analysis of the competing-poles potential $V_{\mathrm{dyn}}[\rho]$ from the $d=2$ qubit Bloch ball to the 8-dimensional manifold $\mathcal{M}_\circ(\mathbb{C}^3)$ of full-rank qutrit density matrices. Using the coordinate-free contravariant inverse Bures metric $G_\rho^{-1}(H) = 2(\rho H + H\rho) - 4\operatorname{Tr}(\rho H)\rho$ and the normalized $SU(3)$ Hilbert–Schmidt purity distance $B_{\mathrm{HS}}(\rho) = \frac{1}{2}(3\operatorname{Tr}(\rho^2)-1)$, we prove the following analytical results without boundary clamping ($\epsilon = 0$): Operator Commutation: Every critical point on $\mathcal{M}_\circ(\mathbb{C}^d)$ commutes with the reference pure projector $\Pi_*$. Orientation Reduction: At fixed spectrum, the non-logarithmic outward eigenvalue pull is strictly increasing in $\sqrt{\rho_{00}}$, reducing the $SU(3)$ boundary-orientation problem to a two-parameter diagonal family. Forward Invariance: The one-sided derivative of $\lambda_{\min}(\rho)$ along the flow is strictly positive on the analytic collar $0 < \lambda_{\min} \le 6.84 \times 10^{-5}$, establishing manifold forward invariance (with sharp numerical threshold $\lambda_{\mathrm{collar}} = 0.00481018$). Transverse Uniqueness: No asymmetric critical points ($\lambda_2 \neq \lambda_3$) exist despite the transverse potential's concavity. Coaxial Attractor: The coaxial balance equation admits a unique root $p^* = 0.961593764$, $\lambda_\perp^* = 0.019203118$, $B_{\mathrm{HS}}(\rho^*) = 0.888100$. Global Convergence: By LaSalle's Invariance Principle, every trajectory initialized in $\mathcal{M}_\circ(\mathbb{C}^3)$ converges globally to $\rho^*$. Includes complete standalone Python verification suite (ntcr_d3_verify.py) and rigorous numerical cross-checks across 105 complex starts.