Skip to main content

ℓPart 2· Theorem H

Theorem H — Flow Stability

Prerequisites: Compactness of GF∘G^{F^\circ}, real-analyticity of E\mathcal{E}, Lojasiewicz inequality (Fact B.8).

Theorem H. If h∗h^* is a non-degenerate local minimum of EF∘\mathcal{E}_{F^\circ}, then the gradient flow h˙s=−grad E(hs)\dot h_s=-\mathrm{grad}\,\mathcal{E}(h_s) converges, and [A∞][A_\infty] is a stable fixed point.

Convergence rate: If the Lojasiewicz exponent at h∗h^* is α=12\alpha=\frac{1}{2}, convergence is exponential: d(hs,h∗⋅Gconst)≤C0 e−γsd(h_s,h^*\cdot\mathcal{G}_{\mathrm{const}})\le C_0\,e^{-\gamma s} for γ>0\gamma>0 depending on the Hessian of E\mathcal{E} at h∗h^*.

Proof. (Complete; expanded from Chapter 7.)

Step 1 (Real-analyticity). The energy EF∘:GF∘→R\mathcal{E}_{F^\circ}:G^{F^\circ}\to\mathbb{R} is real-analytic. Each term Wt(i,j) dG(gth(i,j),e)2W_t(i,j)\,d_G(g_t^h(i,j),e)^2 is analytic in hh: the group multiplication G×G→GG\times G\to G is analytic, the bi-invariant distance squared dG(⋅,e)2d_G(\cdot,e)^2 is analytic on GG (it equals ∥log⁡(⋅)∥g2\|\log(\cdot)\|^2_{\mathfrak{g}} near ee, and extends analytically by compactness and bi-invariance), and a finite sum of analytic functions is analytic.

Step 2 (Global existence). GF∘G^{F^\circ} is compact, so the negative gradient flow exists for all s≥0s\ge0.

Step 3 (Energy monotonicity). ddsE(hs)=−∥grad E(hs)∥2≤0\frac{d}{ds}\mathcal{E}(h_s)=-\|\mathrm{grad}\,\mathcal{E}(h_s)\|^2\le0. Since E≥0\mathcal{E}\ge0, E(hs)↘E∗\mathcal{E}(h_s)\searrow\mathcal{E}^*.

Step 4 (Lojasiewicz convergence). By Fact B.8, since E\mathcal{E} is real-analytic on the compact manifold GF∘G^{F^\circ}, for every critical value cc: ∥grad E(h)∥≥C∣E(h)−c∣α\|\mathrm{grad}\,\mathcal{E}(h)\|\ge C|\mathcal{E}(h)-c|^\alpha for some α∈[12,1)\alpha\in[\frac{1}{2},1) near E−1(c)\mathcal{E}^{-1}(c).

Standard consequence (Simon, 1983): ∫0∞∥h˙s∥ ds<∞\int_0^\infty\|\dot h_s\|\,ds<\infty, so hsh_s has a limit h∞h_\infty.

Step 5 (Exponential rate for α=1/2\alpha=1/2). At a non-degenerate minimum, the Hessian ∇2E∣h∗\nabla^2\mathcal{E}|_{h^*} is positive-definite on Th∗(GF∘)/Th∗(h∗⋅Gconst)T_{h^*}(G^{F^\circ})/T_{h^*}(h^*\cdot\mathcal{G}_{\mathrm{const}}) (the normal space to the gauge orbit). Denote its smallest eigenvalue by λmin⁡>0\lambda_{\min}>0. Then ∥grad E(h)∥2≥λmin⁡(E(h)−E∗)\|\mathrm{grad}\,\mathcal{E}(h)\|^2\ge\lambda_{\min}(\mathcal{E}(h)-\mathcal{E}^*) near h∗h^*, giving α=12\alpha=\frac{1}{2}.

From ddsE=−∥grad E∥2≤−λmin⁡(E−E∗)\frac{d}{ds}\mathcal{E}=-\|\mathrm{grad}\,\mathcal{E}\|^2\le-\lambda_{\min}(\mathcal{E}-\mathcal{E}^*), Gronwall gives E(hs)−E∗≤(E(h0)−E∗)e−λmin⁡s\mathcal{E}(h_s)-\mathcal{E}^*\le(\mathcal{E}(h_0)-\mathcal{E}^*)e^{-\lambda_{\min}s}.

The distance to the critical orbit: d(hs,h∗⋅Gconst)2≤C0(E(hs)−E∗)≤C0(E(h0)−E∗)e−λmin⁡sd(h_s,h^*\cdot\mathcal{G}_{\mathrm{const}})^2\le C_0(\mathcal{E}(h_s)-\mathcal{E}^*)\le C_0(\mathcal{E}(h_0)-\mathcal{E}^*)e^{-\lambda_{\min}s}. Set γ=λmin⁡/2\gamma=\lambda_{\min}/2. □\square