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 Permanence in Weakly Reversible Mass-Action Systems

For any finite weakly reversible reaction network with fixed positive rate constants and any positive stoichiometric class PP, there exists a nonempty compact convex forward-invariant set KP⊂PK_P \subset P such that every initial point x0∈Px^0 \in P gives a unique global positive solution that enters KPK_P in finite time T(x0)≥0T(x^0) \ge 0 and remains there, yielding uniform positive lower and finite upper bounds εP≤xi(t)≤εP−1\varepsilon_P \le x_i(t) \le \varepsilon_P^{-1} (1≤i≤d1 \le i \le d, t≥T(x0)t \ge T(x^0)) depending only on the network, the fixed rates, and PP.

For 0<h≤hP0 < h \le h_P, the plateau Kh={x∈Rd:Fh(x)=Fh(c)}K_h = \{x \in \mathbb{R}^d : F_h(x) = F_h(c)\} is nonempty, compact, and convex, and satisfies Kh⊂int⁡BhK_h \subset \operatorname{int} B_{\sqrt{h}}.

For fixed 0<h≤hP0 < h \le h_P, every x∈Bhx \in B_h and active label λ∈Ah(x)\lambda \in A_h(x) satisfies rλ⋅f(x)≥0r_\lambda \cdot f(x) \ge 0, and there exists δh>0\delta_h > 0 such that rλ⋅f(x)≥δhr_\lambda \cdot f(x) \ge \delta_h whenever rλ∉S⊥r_\lambda \notin S^\perp, which holds for all active labels at x∈Bh∩(P∖Kh)x \in B_h \cap (P \setminus K_h).

For 0<h≤hP0 < h \le h_P, t↦Fh(x(t))t \mapsto F_h(x(t)) is nondecreasing along solution segments in BhB_h and increases at rate at least δh\delta_h on Bh∩(P∖Kh)B_h \cap (P \setminus K_h); every solution starting in KhK_h is global and remains in KhK_h; and any global solution in PP that stays in BhB_h from some time onward enters KhK_h in finite time and remains there.

Methods

Constructing a finite minimum of affine functions Fh(x)=min⁡λ∈Λ(rλ⋅x+bλ(h))F_h(x) = \min_{\lambda \in \Lambda} (r_\lambda \cdot x + b_\lambda(h)) whose active slopes approximate exponent vectors within an arbitrary discontinuous tolerance E:[a,b]m→(0,∞)E : [a, b]^m \to (0, \infty) uniformly for small hh, via induction on dimension and shifted-offset combination of local families.

Source excerpts

“There is a finite nonempty family Λ of labels with rλ ∈ [a, b]m, bλ(h) = o(ha) (h → 0+), such that the following implication holds. If hν → 0+, pν ∈ Rm converges to p ∈ [a, b]m, and a fixed label λ ∈ Λ is active at Xhν (pν) for every ν, then ∥p − rλ∥∞ < E(rλ).” · Show in paper

Cut-flux decomposition and monomial dominance for weakly reversible networks: ordering complexes by rλ⋅yr_\lambda \cdot y, separating consecutive levels by cuts, and using weak reversibility to dominate downward fluxes by upward fluxes.

Source excerpts

“List the distinct values of r · y, for y ∈ C, as a1 < · · · < am. For 1 ≤ j < m, define the net upward flux across the cut at aj by Qj (x) = ∑ u→u′∈R : r·u≤aj<r·u′ ku→u′xu − ∑ u→u′∈R : r·u′≤aj<r·u ku→u′xu.” · Show in paper

Scale-stepping iteration via repeated square roots hj+1=min⁡{hj,hP}h_{j+1} = \min\{\sqrt{h_j}, h_P\} using the inclusion Kh⊂int⁡BhK_h \subset \operatorname{int} B_{\sqrt{h}} to pass from an initial-point-dependent trapping set Kh0K_{h_0} to a classwise uniform absorbing set KP=KhP∩(c+S)K_P = K_{h_P} \cap (c + S).

Source excerpts

“Starting at h0, define hj+1 = min{ √ hj , hP }. (4.2) We claim that for every j there is a finite time Tj such that x(t) ∈ Khj for all t ≥ Tj .” · Show in paper