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

Midpoint convexity from two recursive potentials

For N=NΣN = N_\Sigma on XΣX_\Sigma or N=NJN = N_J on XJX_J, the averaged asymptotic midpoint modulus satisfies δ^N(t)≥t3/128\widehat{\delta}_N(t) \ge t^3/128 for 0<t<10 < t < 1.

For X0X_0, a finite initial set HH, a vector xx supported in HH, and z∈FH∖{0}z \in F_H \setminus \{0\}, one has N0(x+z)+N0(x−z)2≥N0(x)+N0(z)38(2N0(x)+N0(z))2\frac{N_0(x+z)+N_0(x-z)}{2} \ge N_0(x) + \frac{N_0(z)^3}{8(2N_0(x)+N_0(z))^2}.

Each of XΣ,X0,XJX_\Sigma, X_0, X_J admits no equivalent asymptotically uniformly convex (AUC) norm.

The formulas NΣ,N0,NJN_\Sigma, N_0, N_J define infinite-dimensional Banach spaces XΣ,X0,XJX_\Sigma, X_0, X_J after completion, coordinate evaluations and fields extend continuously, ΠH\Pi_H is a contractive projection with finite-codimensional kernel FHF_H in which finitely supported elements are dense, and both X0X_0 and XJX_J are reflexive.

In XJX_J, for every ε>0\varepsilon > 0 there is η(ε)∈(0,1)\eta(\varepsilon) \in (0, 1) such that for x,z1,z2,⋯∈XJx, z_1, z_2, \dots \in X_J, if NJ(x+zj),NJ(x−zj)≤1N_J(x+z_j), N_J(x-z_j) \le 1 (j≥1j \ge 1) and NJ(zi−zj)≥εN_J(z_i - z_j) \ge \varepsilon (i≠ji \ne j), then NJ(x)≤1−η(ε)N_J(x) \le 1 - \eta(\varepsilon); for 0<ε≤20 < \varepsilon \le 2, one can take η(ε)=min⁡{1/2,γ(2,ε/16)}\eta(\varepsilon) = \min\{1/2, \gamma(2, \varepsilon/16)\}.

The variable-exponent space Xv=(⨁h=1∞Eh)ℓ2X_{\mathrm{v}} = (\bigoplus_{h=1}^\infty E_h)_{\ell_2} with ph=1+1/hp_h = 1 + 1/h is complete and reflexive, its dual is separable, vectors with finite total coordinate support are dense, every finite coordinate projection is bounded and has finite rank, and XvX_{\mathrm{v}} admits no equivalent AUC norm.

For x∈Xvx \in X_{\mathrm{v}} with finite total support, finite ancestral sets AhA_h containing the support of x(h)x^{(h)} in each nonzero component, and y∈Xvy \in X_{\mathrm{v}} vanishing on every AhA_h, with S=∥x+y∥v2+∥x−y∥v2S = \|x+y\|_{\mathrm{v}}^2 + \|x-y\|_{\mathrm{v}}^2, one has S−2∥x∥v2≥c6S−2∥y∥v6S - 2\|x\|_{\mathrm{v}}^2 \ge c_6 S^{-2} \|y\|_{\mathrm{v}}^6 where c6=2exp⁡(−(e−1))4−6c_6 = 2\exp(-(\mathrm{e}-1))4^{-6} (and the right side is zero when S=0S=0).

For x∈Xvx \in X_{\mathrm{v}} with finite total support, finite ancestral sets AhA_h containing the support of x(h)x^{(h)} in each nonzero component, and y∈Xvy \in X_{\mathrm{v}} vanishing on these sets, if R>0R > 0 and ∥x+y∥v,∥x−y∥v≤R\|x+y\|_{\mathrm{v}}, \|x-y\|_{\mathrm{v}} \le R, then R2≥∥x∥v2+c322R∥y∥v3R^2 \ge \|x\|_{\mathrm{v}}^2 + \frac{c_3}{2\sqrt{2}R}\|y\|_{\mathrm{v}}^3 where c3=(6e)−3c_3 = (6\mathrm{e})^{-3}; consequently, δ^∥⋅∥v(t)≥c3242t3\widehat{\delta}_{\|\cdot\|_{\mathrm{v}}}(t) \ge \frac{c_3}{24\sqrt{2}} t^3 for 0<t<10 < t < 1.

Methods

Two-potential tree recursion: defining a norm on a rooted tree via two nonnegative fields Ps(v),Qs(v)P_s(v), Q_s(v) computed upwards from the support such that Ps(v)−Qs(v)=vsP_s(v) - Q_s(v) = v_s and each field dominates the Euclidean (or ℓph\ell_{p_h}) norm of its children, choosing the coordinatewise least feasible pair so that at least one child inequality is tight at every node.

Source excerpts

“As(v) = (∑c∈ch(s) Pc(v)^2)^1/2, Bs(v) = (∑c∈ch(s) Qc(v)^2)^1/2, Ps(v) = max{As(v), vs + Bs(v)}, Qs(v) = max{Bs(v), −vs + As(v)}.” · Show in paper

“Ps(v) = max{as, vs + bs}, Qs(v) = max{as − vs, bs}.” · Show in paper

Two-field energy propagation: bounding the sum of squared boundary defects ∑b∈Jdb2\sum_{b \in J} d_b^2 at a first-exit boundary by [do(2Po+do)(2Qo+do)]2/3[d_o(2P_o+d_o)(2Q_o+d_o)]^{2/3} at the root using Hölder's inequality and the tight child inequality.

Source excerpts

“Put Es = [ds(2Ps + ds)(2Qs + ds)]^2/3. We show that Es ≥ ∑c Ec.” · Show in paper

Bounded-tree AUC obstruction: excluding equivalent asymptotically uniformly convex norms by constructing uniformly bounded trees of vectors whose child increments are weakly null and have norms uniformly bounded away from zero.

Source excerpts

“Suppose that for every h ≥ 1 there are vectors u^(h)_s , indexed by s ∈ N^≤h, such that 0 < m ≤ ∥u^(h)_s∥ ≤ M < ∞ uniformly. Suppose also that for |s| < h the increments d^(h)_s,k = u^(h)_{s⌢k} − u^(h)_s are weakly null as k → ∞ and have norms at least c > 0, uniformly in h, s, k. If N is a norm and 0 < α ≤ β < ∞ satisfy α∥ · ∥ ≤ N ≤ β∥ · ∥, then δN (αc / (2βM)) = 0.” · Show in paper

Stopping-time and predictable martingale transform: propagating normalized defects along a branch by stopping when a normalized base field, a product of field ratios, or an endpoint defect crosses a threshold, and controlling the small-product event via a predictable martingale transform and Taylor's inequality.

Source excerpts

“At an inspected node s, stop when ℓs < λ, Zs < ζ, or max_σ d^σ_s > h.” · Show in paper

“The sum M is a predictable martingale transform [4].” · Show in paper

Logarithmic propagation with bounded supermartingales: propagating boundary mass on a finite tree by distributing probability masses along the tree, forming two bounded supermartingales, and applying a logarithmic transform to control a product of ratios.

Source excerpts

“Taking logarithms turns that product into increments of two bounded supermartingales.” · Show in paper

Variable-exponent ℓ1+1/h\ell_{1+1/h} aggregation: replacing Euclidean child aggregation on height-hh components with ℓ1+1/h\ell_{1+1/h} aggregation to keep phh≤ep_h^h \le \mathrm{e} bounded independently of hh, combined with either normalized iteration or accumulated layer energy.

Source excerpts

“We now change the child aggregation to an ℓp norm, choosing p = 1 + 1/h on a tree of height h. This choice keeps the powers p^h ≤ e that arise in iteration bounded independently of h.” · Show in paper