Wiki card — concept node for
exp-families-stability.
Let \(u \in C^{1,2}(\mathbb{R}_+ \times M)\) with \(u > 0\) solve the linear parabolic equation \(\partial_t u = L u\) for a diffusion generator \(L\) (in the BGL sense with carré-du-champ \(\Gamma\)). Define \(f := -\log u\). Then
\[ \boxed{\;\partial_t f \;=\; L f \;-\; \Gamma(f, f)\;} \tag{CH} \]
Euclidean specialisation. For \(M = \mathbb{R}^d\) and \(L = \tfrac12\Delta\), \(\Gamma(f,f) = \tfrac12\|\nabla f\|^2\), so (CH) becomes the viscous Hamilton–Jacobi equation
\[ \partial_t f \;=\; \tfrac12 \Delta f \;-\; \tfrac12 \|\nabla f\|^2. \]
Equivalently, the Cole–Hopf operator \(\mathcal{L}_{\mathrm{CH}} f := \Gamma(f,f) - Lf\) satisfies \(\partial_t f = -\mathcal{L}_{\mathrm{CH}} f\).
Historical note. Hopf 1950 and Cole 1951 independently used \(u = e^{-f/\nu}\) to linearise Burgers' equation \(\partial_t f + f \partial_x f = \nu \partial_{xx}f\) into the heat equation \(\partial_t u = \nu \partial_{xx} u\). The identity transposes verbatim to any diffusion \((M, g, L, \Gamma)\) by the chain rule.
docs/problem.md.
Starting from stability \(p_{\theta_t} =
p_\theta *
\gamma_{\sqrt{2t}}\) and writing \(f_t
:= -\log p_{\theta_t} = \theta_t^\top
\varphi + \log Z_{\theta_t}\), (CH) gives exactly equation (6)
after identifying terms in \(\varphi\).invariant-reformulation.md
§3).Take \(M = \mathbb{R}\), \(L = \tfrac12\partial_{xx}\), and \(u(t,x) = (2\pi t)^{-1/2} e^{-x^2/(2t)}\) (heat kernel centred at \(0\)). Then \(f(t,x) = -\log u = \tfrac12 \log(2\pi t) + x^2/(2t)\).
Check (CH) directly:
Then \(Lf - \Gamma(f,f) = \tfrac{1}{2t} - \tfrac{x^2}{2t^2} = \partial_t f\). ✓
The nonlinear HJ equation on \(f\) is exactly equivalent to the linear heat equation on \(u = e^{-f}\) — no approximation. This is the lift that underlies the whole program: the quadratic family \(\varphi(x) = x^2\) is stable under \(*\gamma_\sigma\) because \(x^2\) (up to a linear and a constant) is closed under \(f \mapsto \Gamma(f,f) - Lf\) evaluated on its span.