Wiki card — concept node for
exp-families-stability.
Let \((P_t)_{t\ge 0}\) be the heat semigroup on \(\mathbb{R}^d\): \((P_t f)(x) = \int f(x - y)\, \gamma_{\sqrt{t}}(y)\, dy\). Basic properties:
Intertwining with Gaussian convolution. \(P_t \equiv *\gamma_{\sqrt{t}}\) (up to the \(\tfrac12\) convention in the generator). For the problem statement, the stability condition (4) is \[ p_\theta * \gamma_\sigma \;\stackrel{\mathrm{up\ to\ norm.}}{=}\; p_{\tilde\theta} \qquad\Longleftrightarrow\qquad P_{\sigma^2}\, p_\theta \;\propto\; p_{\tilde\theta}. \]
Key identity (Dynkin's formula on \(f = \log p\)). With \(p_t := P_t p\) and \(f_t := -\log p_t\), \[ \boxed{\;\partial_t f_t \;=\; \tfrac12 \Delta f_t \;-\; \tfrac12 \|\nabla f_t\|^2\;} \] is the commutator between \(P_t\) and \(-\log\).
On \(\mathbb{R}\), define Hermite polynomials \(H_n(x) = (-1)^n e^{x^2/2} \partial_x^n e^{-x^2/2}\): \(H_0 = 1\), \(H_1 = x\), \(H_2 = x^2 - 1\), \(H_3 = x^3 - 3x\), \(\ldots\).
Claim (for standard heat \(P_t\) with \(\partial_t = \tfrac12\Delta\)): \[ P_t H_n(x) \;=\; (1 + t)^{n/2}\, H_n\!\left(\frac{x}{\sqrt{1 + t}}\right). \]
(This is the "heat-rescaling" action of \(P_t\) on Hermite — standard.)
Check on \(H_2(x) = x^2 - 1\). Compute \(P_t(x^2) = x^2 + t\) (since \(\int (x-y)^2 \gamma_{\sqrt{t}}(dy) = x^2 + t\)), so \(P_t H_2 = x^2 + t - 1 = (x^2 - (1 - t))\). The claimed formula: $(1+t) H_2(x/\sqrt{1+t}) = (1+t)(x^2/(1+t) -
Read-off. \(\mathrm{Span}\{H_0, H_1, H_2\} = \mathrm{Span}\{1, x, x^2\}\) is invariant under \(P_t\): the quadratic class is stable under heat flow — this is exactly Q2's conclusion, stated at the level of individual basis elements.
Contrast with Hermite on OU. Under the OU semigroup \(P_t^{\mathrm{OU}}\), Hermite polynomials are eigenfunctions: \(P_t^{\mathrm{OU}} H_n = e^{-nt} H_n\). The subspace \(\bigoplus_{n \le 2} H_n\) is again invariant — the stable class coincides between heat and OU at the quadratic-linear level, confirming Wheeler's invariant reformulation.
Failure at degree \(\ge 3\). For \(\varphi(x) = x^3\): \(P_t(x^3) = x^3 + 3 t x\), so \(\mathrm{Span}\{x^3\}\) alone is not \(P_t\)-invariant (one needs \(x\) too, and then \(x^2\) via \(\nabla\varphi\cdot\nabla\varphi = 9x^4 \notin\) any finite polynomial span). Finite-dimensional closure breaks: the Q1-obstruction at \(\deg \ge 3\) is explicit.