OctoInformBeta

OpenAI preprint, September 2026 · §3. Moment coordinates on the whole space

Lemma 3.1: The potential φ\varphi has 0<∇2φ≤(4/ε)I0<\nabla^2\varphi\le(4/\varepsilon)I and solves log⁡det⁡∇2φ=V(∇φ)+log⁡Z−φ\log\det\nabla^2\varphi=V(\nabla\varphi)+\log Z-\varphi; its gradient is a diffeomorphism

Soft slab laws are moment measures of smooth potentials.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Fix ε>0\varepsilon>0, an even natural number qq, vectors p1,…,pmp_1,\dots,p_m of ℝn\mathbb{R}^n and real numbers h1,…,hmh_1,\dots,h_m, the widths. With ⟨p,x⟩=∑jpjxj\langle p,x\rangle=\sum_jp_jx_j the linear functional of a vector pp, for x=(x1,…,xn)x=(x_1,\dots,x_n) in ℝn\mathbb{R}^n, and |x|2=∑jxj2|x|^2=\sum_jx_j^2, the soft slab potential VV and its slab summands ψi\psi_i are V(x)=ε2|x|2+∑i=1mψi(x),ψi(x)=(⟨pi,x⟩hi)q,V(x)=\frac{\varepsilon}{2}\,|x|^2+\sum_{i=1}^m\psi_i(x),\qquad \psi_i(x)=\Big(\frac{\langle p_i,x\rangle}{h_i}\Big)^q,(1) and the soft slab law is the measure μ=Z−1e−V(x)dx,Z=∫ℝne−V(x)dx.\mu=Z^{-1}e^{-V(x)}\,dx,\qquad Z=\int_{\mathbb{R}^n}e^{-V(x)}\,dx.(2)✦

The question is whether μ\mu is the moment measure of a smooth function on the whole space: whether some C∞C^\infty function φ\varphi on ℝn\mathbb{R}^n has a gradient map ∇φ\nabla\varphi that carries the source measure ν=e−φ(z)dz\nu=e^{-\varphi(z)}\,dz, with density the weight e−φe^{-\varphi} against Lebesgue measure, onto μ\mu, written (∇φ)#ν=μ(\nabla\varphi)_\#\nu=\mu, and that is a change of coordinates of ℝn\mathbb{R}^n. The variable zz of φ\varphi is called the source coordinate, where ν\nu lives, and x=∇φ(z)x=\nabla\varphi(z) the target coordinate, where μ\mu lives.

Derivatives and Measures

The directional derivative DvfD_vf of a function ff on a real normed space along a vector vv is the function x↦Df(x)vx\mapsto Df(x)\,v, where Df(x)Df(x) is the Fréchet derivative of ff at xx.✦ The partial derivative ∂jf\partial_jf of a function ff on ℝn\mathbb{R}^n is DejfD_{e_j}f✦, for the standard basis vector eje_j with coordinate 11 at jj and 00 at every other index✦, and its gradient ∇f(x)\nabla f(x) is the vector (∂1f(x),…,∂nf(x))(\partial_1f(x),\dots,\partial_nf(x)).✦ The Hessian ∇2f(x)\nabla^2f(x) of ff at xx is the n×nn\times n matrix with entry ∂k∂jf(x)\partial_k\partial_jf(x) in row jj and column kk.✦

A Euclidean space EE is a finite-dimensional real inner product space, with norm |z||z|. Euclidean nn-space 𝔼n\mathbb{E}^n is ℝn\mathbb{R}^n with the inner product ⟨x,y⟩=∑jxjyj\langle x,y\rangle=\sum_jx_jy_j, and its standard orthonormal basis e=(e1,…,en)e=(e_1,\dots,e_n) consists of the unit coordinate vectors. The coordinate isomorphism C:𝔼n→ℝnC\colon\mathbb{E}^n\to\mathbb{R}^n is the continuous linear isomorphism that keeps every coordinate.✦

The gradient ∇φ(z)\nabla\varphi(z) of a differentiable function φ\varphi on EE at zz is the vector with ⟨∇φ(z),v⟩=Dvφ(z)\langle\nabla\varphi(z),v\rangle=D_v\varphi(z) for every vv, and the second derivative D2φ(z)D^2\varphi(z) of φ\varphi at zz is the derivative of ∇φ\nabla\varphi at zz, a linear map of EE into itself. For an orthonormal basis b=(b1,…,bn)b=(b_1,\dots,b_n) of EE, the Hessian matrix [D2φ(z)]b[D^2\varphi(z)]_b is the matrix of D2φ(z)D^2\varphi(z) in bb, with entry ⟨bj,D2φ(z)bk⟩\langle b_j,D^2\varphi(z)\,b_k\rangle in row jj and column kk.✦

For real n×nn\times n matrices, A>0A>0 means that AA is symmetric and positive definite, and A≤BA\le B that B−AB-A is symmetric and positive semidefinite; II is the n×nn\times n identity matrix. The symmetric second difference of φ\varphi at zz with step hh is Δhφ(z)=φ(z+h)+φ(z−h)−2φ(z).\Delta_h\varphi(z)=\varphi(z+h)+\varphi(z-h)-2\varphi(z).(3)✦

The normalized measure μU\mu_U of a function UU on a Euclidean space EE is the measure ZU−1e−U(x)dxZ_U^{-1}e^{-U(x)}\,dx, whose density is the weight e−Ue^{-U} divided by its integral ZU=∫Ee−U(x)dxZ_U=\int_Ee^{-U(x)}\,dx.✦

The Claim

Theorem 1. Let ε>0\varepsilon>0 and qq even, and let VV, ZZ and μ\mu be as in (1) and (2). On each of the spaces ℝn\mathbb{R}^n and 𝔼n\mathbb{E}^n there is a C∞C^\infty function φ\varphi whose gradient is a homeomorphism of the space onto itself with C∞C^\infty inverse and which satisfies, on ℝn\mathbb{R}^n, at every point zz, 0<∇2φ(z)≤4εI,log⁡det⁡∇2φ(z)=V(∇φ(z))+log⁡Z−φ(z),(∇φ)#(e−φ(z)dz)=μ,\begin{gathered} 0<\nabla^2\varphi(z)\le\frac{4}{\varepsilon}\,I,\qquad \log\det\nabla^2\varphi(z)=V(\nabla\varphi(z))+\log Z-\varphi(z),\\ (\nabla\varphi)_\#\big(e^{-\varphi(z)}\,dz\big)=\mu, \end{gathered}(4)✦ and, on 𝔼n\mathbb{E}^n, at every point zz, 0<[D2φ(z)]e≤4εI,log⁡det⁡[D2φ(z)]e=V(C∇φ(z))+log⁡ZV∘C−φ(z),(∇φ)#(e−φ(z)dz)=μV∘C.\begin{gathered} 0<[D^2\varphi(z)]_e\le\frac{4}{\varepsilon}\,I,\qquad \log\det\,[D^2\varphi(z)]_e=V(C\nabla\varphi(z))+\log Z_{V\circ C}-\varphi(z),\\ (\nabla\varphi)_\#\big(e^{-\varphi(z)}\,dz\big)=\mu_{V\circ C}. \end{gathered}(5)✦ Moreover, let EE be a Euclidean space and bb an orthonormal basis of EE.

  1. If dim⁡E≥1\dim E\ge1, ε>0\varepsilon>0, and UU is a continuous even function on EE with e−Ue^{-U} integrable and 12(U(x)+U(y))−U(12(x+y))≥ε8|x−y|2\tfrac12\big(U(x)+U(y)\big)-U\big(\tfrac12(x+y)\big)\ge\tfrac{\varepsilon}{8}\,|x-y|^2 for all xx and yy, then there are a convex even C1C^1 function φ\varphi on EE whose gradient maps EE onto EE, a constant c>0c>0 and a real number dd such that, for all zz, ww and hh, Δhφ(z)≤4ε|h|2,|∇φ(z)−∇φ(w)|≤16ε|z−w|,φ(z)≥c|z|−d,∫Ee−φ(z)dz=1,(∇φ)#(e−φ(z)dz)=μU.\begin{gathered} \Delta_h\varphi(z)\le\frac{4}{\varepsilon}\,|h|^2,\qquad |\nabla\varphi(z)-\nabla\varphi(w)|\le\frac{16}{\varepsilon}\,|z-w|,\qquad \varphi(z)\ge c|z|-d,\\ \int_Ee^{-\varphi(z)}\,dz=1,\qquad (\nabla\varphi)_\#\big(e^{-\varphi(z)}\,dz\big)=\mu_U. \end{gathered}(6)✦

  2. Let φ\varphi be a convex C1C^1 function on EE whose gradient is Lipschitz and maps EE onto EE, and UU a function on EE with e−Ue^{-U} integrable and (∇φ)#(e−φ(z)dz)=μU(\nabla\varphi)_\#\big(e^{-\varphi(z)}\,dz\big)=\mu_U. If UU is convex and C1C^1, then φ∈C2(E).\varphi\in C^2(E).(7)✦ If φ\varphi is C2C^2 and UU is convex and C∞C^\infty, then φ∈C∞(E).\varphi\in C^\infty(E).(8)✦ If φ\varphi is C2C^2 and UU is continuous, then at every point zz [D2φ(z)]b>0,log⁡det⁡D2φ(z)=U(∇φ(z))+log⁡ZU−φ(z).[D^2\varphi(z)]_b>0,\qquad \log\det D^2\varphi(z)=U(\nabla\varphi(z))+\log Z_U-\varphi(z).(9)✦

  3. If φ\varphi is a C2C^2 function on EE with [D2φ(z)]b>0[D^2\varphi(z)]_b>0 at every zz, and Λ\Lambda is a real number with Δhφ(z)≤Λ|h|2\Delta_h\varphi(z)\le\Lambda|h|^2 for all hh and zz, then at every point zz [D2φ(z)]b≤ΛI.[D^2\varphi(z)]_b\le\Lambda I.(10)✦

  4. If GG is a C∞C^\infty map of EE onto EE with ⟨v,DG(z)v⟩>0\langle v,DG(z)\,v\rangle>0 for every zz and every nonzero vv, then Gis a homeomorphism of E onto E,G−1∈C∞(E,E).G\ \text{is a homeomorphism of } E \text{ onto } E,\qquad G^{-1}\in C^\infty(E,E).(11)✦

In the paper, the result reads: for n≥2n\ge2, ε>0\varepsilon>0 and an even V∈C∞(ℝn)V\in C^\infty(\mathbb{R}^n) with D2V≥εID^2V\ge\varepsilon I, Z=∫e−VZ=\int e^{-V} and μ=Z−1e−V(x)dx\mu=Z^{-1}e^{-V(x)}\,dx, there is an even φ∈C∞(ℝn)\varphi\in C^\infty(\mathbb{R}^n) such that ν=e−φ(z)dz\nu=e^{-\varphi(z)}\,dz is a probability measure, ∇φ\nabla\varphi is a diffeomorphism of ℝn\mathbb{R}^n onto ℝn\mathbb{R}^n with (∇φ)#ν=μ(\nabla\varphi)_\#\nu=\mu, 0<D2φ≤(4/ε)I0<D^2\varphi\le(4/\varepsilon)I and log⁡det⁡D2φ=V(∇φ)+log⁡Z−φ\log\det D^2\varphi=V(\nabla\varphi)+\log Z-\varphi, and moreover φ(z)≥c|z|−d\varphi(z)\ge c|z|-d for some c>0c>0 and some real dd.

What Is Checked

The claim is stated over the soft slab potential and law (1), (2), the normalized measures μU\mu_U, the source measures e−φ(z)dze^{-\varphi(z)}\,dz, the gradient and Hessian built from partial derivatives on ℝn\mathbb{R}^n, the Hessian matrix [D2φ]b[D^2\varphi]_b in an orthonormal basis, the second difference (3) and the coordinate isomorphism CC.