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.

All papers

Claims

Findings

Uniform Bi-Holder Transport from Weak MTW

On a fixed smooth connected compact boundaryless Riemannian manifold (M,g)(M, g) of dimension n≥2n \ge 2 satisfying the weak Ma–Trudinger–Wang condition, for any 0<λ≤Λ<∞0 < \lambda \le \Lambda < \infty with a nonempty density class, there exist α∈(0,1]\alpha \in (0, 1] and C<∞C < \infty depending only on (M,g),λ,Λ(M, g), \lambda, \Lambda such that for every pair of probability densities ρ0,ρ1\rho_0, \rho_1 satisfying λ≤ρi≤Λ\lambda \le \rho_i \le \Lambda vol\mathrm{vol}-almost everywhere, the almost-everywhere unique optimal transport map for c(x,y)=12d(x,y)2c(x, y) = \frac{1}{2}d(x, y)^2 has a homeomorphic representative T~:M→M\widetilde{T}: M \to M with d(T~(x),T~(x′))≤Cd(x,x′)αd(\widetilde{T}(x), \widetilde{T}(x')) \le C d(x, x')^\alpha and d(T~−1(y),T~−1(y′))≤Cd(y,y′)αd(\widetilde{T}^{-1}(y), \widetilde{T}^{-1}(y')) \le C d(y, y')^\alpha for all x,x′,y,y′∈Mx, x', y, y' \in M.

If the section-oscillation supremum F(r)F(r) satisfies F(r)≤C0rβF(r) \le C_0 r^\beta for 0<r<r00 < r < r_0 with β>0\beta > 0, C0<∞C_0 < \infty, and r0>0r_0 > 0, then all optimal maps in the fixed density class have homeomorphic representatives with a common Hölder estimate in both directions for any exponent no larger than 12min⁡{β,1}\frac{1}{2}\min\{\beta, 1\}.

The section-oscillation supremum function F(r)F(r) of Lemma 4.2 has a positive power upper bound at zero.

There are constants cr>0c_{\mathrm{r}} > 0 and Cr<∞C_{\mathrm{r}} < \infty depending only on (M,g)(M, g) such that for x∈Mx \in M, v∈Gxv \in G_x, 0<δ≤1/20 < \delta \le 1/2, and 1≤j≤n1 \le j \le n, the singular values satisfy cr(sj(x,v)+δ)≤sj(x,(1−δ)v)≤Cr(sj(x,v)+δ)c_{\mathrm{r}}(s_j(x, v) + \delta) \le s_j(x, (1-\delta)v) \le C_{\mathrm{r}}(s_j(x, v) + \delta), and Nx,δ(v)=−δ(Ax((1−δ)v)−Id)N_{x,\delta}(v) = -\delta(A_x((1-\delta)v) - \mathrm{Id}) is within CrδC_{\mathrm{r}}\delta of a positive definite form Qx,v,δQ_{x,v,\delta} whose eigenvalues are comparable to δ/(sj(x,v)+δ)\delta / (s_j(x, v) + \delta).

Assuming weak MTW, for fixed a0>0a_0 > 0 and B<∞B < \infty, there is cs>0c_{\mathrm{s}} > 0 depending only on (M,g),a0,B(M, g), a_0, B such that whenever p+tq∈Gxp + tq \in G_x for −a0≤t≤1-a_0 \le t \le 1, p⋅q>0p \cdot q > 0, and ∣q∣2/(p⋅q)≤B|q|^2 / (p \cdot q) \le B, one has sj(x,p)≥cssj(x,p+q)s_j(x, p) \ge c_{\mathrm{s}} s_j(x, p + q) for every 1≤j≤n1 \le j \le n, and consequently σx(p)≥csnσx(p+q)\sigma_x(p) \ge c_{\mathrm{s}}^n \sigma_x(p + q).

For any KK and corresponding outer template, at all sufficiently small scales in Lemma 4.2, there exist a parameter b∈[b−,b+]b \in [b_-, b_+], a pole xx, center and active logs p,pi∈Gxp, p_i \in G_x, weights mi>0m_i > 0 summing to one, and symmetric matrices H,L,V,ViH, L, V, V_i satisfying the one-dimensional span, outward selection, matrix inequalities (5.23)–(5.24), nonconjugacy of the modified and true limiting branches, and the determinant bounds det⁡V≥cσx(p)2σy(r/l)\det V \ge c \frac{\sigma_x(p)^2}{\sigma_y(r/l)} and det⁡Vi≤Cσx(pi)2σyi(ri/li)\det V_i \le C \frac{\sigma_x(p_i)^2}{\sigma_{y_i}(r_i/l_i)} with c,C,m∗c, C, m_* independent of KK and the outer template.

If symmetric matrices H,L,V,ViH, L, V, V_i and vectors p,pi,ep, p_i, e satisfy (5.21)–(5.24), then det⁡Vi≥cKdet⁡V\det V_i \ge c K \det V, where c>0c > 0 depends only on the fixed center-template constants, the dimension nn, and m∗m_*, and is independent of KK.

Methods

Factoring the Jacobi matrix along a minimizing geodesic at a fixed interior splitting time h=1/4h = 1/4 to obtain uniform radial transversality estimates and divided-action Hessian bounds near conjugate endpoints.

Source excerpts

“For each such vector consider γ(s) = expx (sv) on [0, 1]. Fix the splitting time h = 1/4 and allow the final time T to vary in [1/2, 1].” · Show in paper

“Jacobi propagation through the joining point gives the exact factorization J(T) = Ch(T − h)J(h) + Jh(T − h)J ′ (h) = Jh(T − h)D(T)J(h).” · Show in paper

Conjugating radial-shortened cost Hessians into a uniformly bounded Loewner-convex matrix family under weak MTW and choosing a separating logarithmic radial scale on a countable dense subset to prove uniform singular-value comparison along an affine segment.

Source excerpts

“For 0 < δ ≤ 1/2, put Nδ(t) = −δ (Ax((1 − δ)p(t)) − Id) , Fδ(t) = L(t) ∗Nδ(t)L(t).” · Show in paper

“We now choose one radial scale that separates every fixed member of this set. Take log δk in the middle half of the interval [log sj,k(0), log sj,k(1)].” · Show in paper

Constructing two smooth scalar template modifications v+bBc(s)v + b B_c(s) and v+bBo(s)v + b B_o(s) at separated logarithmic-plateau gap scales and maximizing the penalized regularized envelope Φt(b,z)=Qtgb(z)+Q1−twc,b(z)−π(b)\Phi_t(b, z) = Q_t g_b(z) + Q_{1-t} w_{c,b}(z) - \pi(b) to obtain an interior maximum and universal parameter inequality.

Source excerpts

“At the normalized level s = (v − a)/D, we will use the two modifications v + bBc(s) and v + bBo(s). The first is the center datum; we compare it with the double cost transform of the second.” · Show in paper

“Φt(b, z) = Qtgb(z) + Q1−twc,b(z) − π(b), (b, z) ∈ [b−, b+] × M, has a positive global maximum with b in the interior of its interval.” · Show in paper

Transferring lower Taylor jets from a prefix regularization to active outer endpoints via an exact pole-matrix identity, sampling Jensen points off null sets, and ordering limits so the outer contact-volume bound controls the center Hessian before applying the original Jacobian equation.

Source excerpts

“In the fixed-pole coordinates z(q) = expx (tq), define L = D2 q{ct(x, z(q)) − F(z(q))} q=p . (5.5)” · Show in paper

“Step 2: null sets are removed before Jensen sampling.” · Show in paper

Using an oblique projection P=Id−ep∗/qP = \mathrm{Id} - e p^* / q onto p⊥p^\perp along ee and evaluating determinants in the nonorthogonal basis [Q,e][Q, e] to extract a determinant gain proportional to the outer curvature parameter KK.

Source excerpts

“Define P = Id − ep∗ /q. This is the projection onto p ⊥ along e. Since LE = 0, P ∗LP = L.” · Show in paper