By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

The logarithmic Brunn–Minkowski conjecture

For every n≥1n \ge 1, arbitrary origin-symmetric convex bodies K,L⊂RnK, L \subset \mathbb{R}^n (K=−KK = -K and L=−LL = -L), and every 0≤λ≤10 \le \lambda \le 1, the Wulff body of the geometric mean of their support functions satisfies ∣W[hK1−λhLλ]∣≥∣K∣1−λ∣L∣λ|\mathcal{W}[h_K^{1-\lambda} h_L^\lambda]| \ge |K|^{1-\lambda}|L|^\lambda.

For n≥2n \ge 2, N≥1N \ge 1, even integer q≥2q \ge 2, ε>0\varepsilon > 0, p1,…,pN∈Rn∖{0}p_1, \dots, p_N \in \mathbb{R}^n \setminus \{0\}, and h1,…,hN>0h_1, \dots, h_N > 0, the probability measure dμ=Z−1e−V(x)dxd\mu = Z^{-1}e^{-V(x)}dx with V(x)=ε2∣x∣2+∑i=1N(⟨pi,x⟩/hi)qV(x) = \frac{\varepsilon}{2}|x|^2 + \sum_{i=1}^N (\langle p_i, x\rangle / h_i)^q satisfies Var⁡μ(∑i=1Nbiψi)≤∑i=1Nbi2Eμψi\operatorname{Var}_\mu(\sum_{i=1}^N b_i \psi_i) \le \sum_{i=1}^N b_i^2 \mathbb{E}_\mu \psi_i for all real b1,…,bNb_1, \dots, b_N.

For n≥2n \ge 2, ε>0\varepsilon > 0, and an even V∈C∞(Rn)V \in C^\infty(\mathbb{R}^n) with D2V≥εID^2V \ge \varepsilon I, there exists an even φ∈C∞(Rn)\varphi \in C^\infty(\mathbb{R}^n) such that dν(z)=e−φ(z)dzd\nu(z) = e^{-\varphi(z)}dz is a probability measure, ∇φ\nabla\varphi is a diffeomorphism of Rn\mathbb{R}^n onto Rn\mathbb{R}^n pushing ν\nu to dμ=Z−1e−V(x)dxd\mu = Z^{-1}e^{-V(x)}dx, 0<D2φ≤4εI0 < D^2\varphi \le \frac{4}{\varepsilon}I, log⁡det⁡D2φ=V(∇φ)+log⁡Z−φ\log\det D^2\varphi = V(\nabla\varphi) + \log Z - \varphi, and φ(z)≥c∣z∣−b\varphi(z) \ge c|z| - b for some c>0c > 0 and b<∞b < \infty.

For a probability measure dμ=ρ(x)dxd\mu = \rho(x)dx on Rn\mathbb{R}^n with smooth, strictly positive, even density ρ\rho and a smooth even symmetric matrix field τ\tau with 0<τ(x)≤ΛI0 < \tau(x) \le \Lambda I, if Af=x⋅∇f−τ:D2fAf = x \cdot \nabla f - \tau : D^2f satisfies E[hAf]=E[∇h⋅τ∇f]\mathbb{E}[hAf] = \mathbb{E}[\nabla h \cdot \tau\nabla f] and the compact Hessian identity (4.7), then {A(A−1)f:f∈Cc∞(Rn),f even}\{A(A-1)f : f \in C_c^\infty(\mathbb{R}^n), f\text{ even}\} is dense in {g∈L2(μ):g even,Eg=0}\{g \in L^2(\mu) : g\text{ even}, \mathbb{E}g = 0\}.

Under the moment-coordinate setup dμ=Z−1e−Vdxd\mu = Z^{-1}e^{-V}dx, dν=e−φdzd\nu = e^{-\varphi}dz, x=∇φ(z)x = \nabla\varphi(z) with Dz2φ>0D_z^2\varphi > 0, for any real f∈Cc∞(Rxn)f \in C_c^\infty(\mathbb{R}_x^n) with τ=Dz2φ\tau = D_z^2\varphi, M=τ−1M = \tau^{-1}, H=Dx2VH = D_x^2V, B=τ(Dx2f)τB = \tau(D_x^2f)\tau, W=δμBW = \delta_\mu B, and u=δμWu = \delta_\mu W, we have Etr⁡((DxW)2)≥Etr⁡(MBHB)\mathbb{E}\operatorname{tr}((D_x W)^2) \ge \mathbb{E}\operatorname{tr}(MBHB) and Eu2≥Etr⁡(MBHB)+EWTHW\mathbb{E}u^2 \ge \mathbb{E}\operatorname{tr}(MBHB) + \mathbb{E}W^THW.

For n≥1n \ge 1, full-dimensional origin-symmetric convex bodies K,L⊂RnK, L \subset \mathbb{R}^n, 0<p<10 < p < 1, and 0≤λ≤10 \le \lambda \le 1, ∣W[((1−λ)hKp+λhLp)1/p]∣p/n≥(1−λ)∣K∣p/n+λ∣L∣p/n|\mathcal{W}[((1-\lambda)h_K^p + \lambda h_L^p)^{1/p}]|^{p/n} \ge (1-\lambda)|K|^{p/n} + \lambda|L|^{p/n}.

For n≥1n \ge 1, any even log-concave Radon measure μ\mu on Rn\mathbb{R}^n, and any origin-symmetric convex bodies K,L⊂RnK, L \subset \mathbb{R}^n, μ(W[hK1−λhLλ])≥μ(K)1−λμ(L)λ\mu(\mathcal{W}[h_K^{1-\lambda}h_L^\lambda]) \ge \mu(K)^{1-\lambda}\mu(L)^\lambda for every 0≤λ≤10 \le \lambda \le 1, and consequently t↦μ(etK)t \mapsto \mu(e^t K) is log-concave on R\mathbb{R}.

Methods

Smooth approximation of symmetric polytopes by exponential weights e−ε2∣x∣2−∑i(⟨pi,x⟩/hi)qe^{-\frac{\varepsilon}{2}|x|^2 - \sum_i (\langle p_i, x\rangle / h_i)^q} with q≥2q \ge 2 even and ε>0\varepsilon > 0, followed by q→∞q \to \infty, ε↓0\varepsilon \downarrow 0, and outer approximation by nested finite spanning direction sets.

Source excerpts

“For finitely many spanning unit vectors pi and positive numbers hi, the constraints |<pi, x>| ≤ hi define a symmetric polytope. Replace its indicator by the smooth weight exp(−ε/2 |x|^2 − ∑_i (<pi, x>/hi)^q), ε > 0, q ≥ 2 even.” · Show in paper

“Choose nested finite sets Dm ⊂ Sn−1S^{n−1} that contain the coordinate vectors and whose union is dense in the sphere.” · Show in paper

Construction of global moment coordinates x=∇φ(z)x = \nabla\varphi(z) with 0<D2φ≤(4/ε)I0 < D^2\varphi \le (4/\varepsilon)I by exhausting Rn\mathbb{R}^n with balls BRB_R, applying Berman–Berndtsson's compact moment theorem and Klartag's Hessian bound, and passing to a smooth limit via Evans–Krylov and Schauder interior estimates.

Source excerpts

“The moment theorem of Berman–Berndtsson [2, Theorem 1.1], applied to the body \overline{BR} and the positive smooth density e−Ve^{-V}/Z_R, gives a smooth convex potential φR, unique up to translation, for which ∇φR : Rn → BR is a diffeomorphism and det D^2 φR = Z_R exp{V(∇φR) − φR}.” · Show in paper

“The nonhomogeneous Evans–Krylov estimate [33, Corollary 2.3] applied to F(D^2 φR) = fR gives uniform interior C2,αC^{2,α} bounds for some α > 0.” · Show in paper

Proving L2(μ)L^2(\mu)-density of even double-divergence test functions A(A−1)fA(A-1)f in the centered even subspace using interior elliptic regularity for A(A−1)g=0A(A-1)g = 0 and L2L^2 cutoff estimates.

Source excerpts

“Interior elliptic regularity consequently gives g ∈ C^∞ [14, Theorem 14.2]. Only positivity on compact sets is needed here. Splitting the equation in L^2. Choose 0 ≤ ηr ≤ 1 smooth, equal to one on Br, zero outside B2r, with |∇ηr| ≤ C/r and |D^2 ηr| ≤ C/r^2, where r ≥ 1.” · Show in paper

Pointwise tensor comparison of Etr⁡((DxW)2)\mathbb{E}\operatorname{tr}((D_x W)^2) and Etr⁡(MBHB)\mathbb{E}\operatorname{tr}(MBHB) in moment coordinates by integrating by parts and evaluating the difference in dual coordinates normalized so that τ=I\tau = I at a fixed point.

Source excerpts

“The common first term cancels. We now evaluate the three remaining contracted terms in coordinates where τ = I at the point under consideration. This normalization uses a constant dual change of coordinates.” · Show in paper

Support-subspace reduction for even log-concave Radon measures using restriction to a large ball, Borell's characterization on the affine hull EE of the support, and Saroglou's transfer theorem in dimension dim⁡E\dim E.

Source excerpts

“Let E be the affine hull of supp ν. Evenness gives E = −E, so E is a linear subspace. ... Borell’s characterization [4, Theorem 1.1, with s = 0], applied to ν on E, represents it as an affine image of a log-concave density in a dimension k ≤ d.” · Show in paper