OctoInformBeta

OpenAI preprint, September 2026 · §3. From affine labels to persistent trajectories

§3.1 (setting): Uniform affine approximation supplies the construction for weakly reversible networks with positive rates and a positive initial state

Affine barriers from finitely many comparisons.

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

The construction concerns affine barriers around a positive initial concentration state, with inward conditions determined by the mass-action vector field on a reciprocal box.

A reaction network NN on dd species consists of a finite complex set C⊆ℕdC\subseteq\mathbb N^d and a finite reaction set ℛ⊆{(y,z)∈C2:y≠z}\mathcal R\subseteq\{(y,z)\in C^2:y\ne z\}.✦ A network reaction is a pair (y,z)∈ℛ(y,z)\in\mathcal R, written y→zy\to z.✦ A reaction network NN is weakly reversible if every reaction y→zy\to z of NN has a directed path of one or more reactions from zz back to yy.✦ A state x∈ℝdx\in\mathbb R^d is positive if xi>0x_i>0 for every species ii.✦

For positive constant reaction rates κy→z\kappa_{y\to z}, the mass-action monomial xyx^y of a complex y∈Cy\in C is a scalar-valued function of the state xx, and the mass-action vector field ff of NN is a vector-valued function of xx, as defined by (1). xy=∏i=1dxiyi,f(x)=∑y→z∈ℛκy→zxy(z−y).x^y=\prod_{i=1}^d x_i^{y_i}, \qquad f(x)=\sum_{y\to z\in\mathcal R}\kappa_{y\to z}x^y(z-y).(1)

The dot product of r,x∈ℝdr,x\in\mathbb R^d is r⋅x=∑i=1drixir\cdot x=\sum_{i=1}^d r_ix_i.✦ The coordinate direction ei∈ℝde_i\in\mathbb R^d has component (ei)k=δik(e_i)_k=\delta_{ik}✦, where δik\delta_{ik} is one when i=ki=k and zero otherwise, and the complex difference z−y∈ℝdz-y\in\mathbb R^d has component (z−y)i=zi−yi(z-y)_i=z_i-y_i.✦

Definition 1. The comparison directions of NN form the finite set DND_N in (2). DN={ei:1≤i≤d}∪{z−y:y,z∈C}.D_N=\{e_i:1\le i\le d\}\cup\{z-y:y,z\in C\}.(2)✦

For an exponent vector p∈ℝdp\in\mathbb R^d, the power-coordinate state Xh(p)X_h(p) and the reciprocal box BhB_h are defined by (3). Norms on exponent vectors are supremum norms. Xh(p)=(hpi)i=1d(h>0),Bh=[h,h−1]d(0<h<1).X_h(p)=(h^{p_i})_{i=1}^d\quad(h>0), \qquad B_h=[h,h^{-1}]^d\quad(0<h<1).(3)

An affine family with labels j∈{1,…,n}j\in\{1,\ldots,n\}, n≥1n\ge1, consists of slopes rj∈ℝdr_j\in\mathbb R^d and real offsets cj(h)c_j(h), with branch value rj⋅x+cj(h)r_j\cdot x+c_j(h). For an affine family in ℝd\mathbb R^d indexed by a label set Λ\Lambda of arbitrary size, an active affine label j∈Λj\in\Lambda at a state x∈ℝdx\in\mathbb R^d is a label whose branch value is no greater than that of every label in Λ\Lambda, including ties; for the finite family above, the set 𝒜h(x)\mathcal A_h(x) of active labels is defined by (4). j∈𝒜h(x)⇔rj⋅x+cj(h)≤rk⋅x+ck(h)for every k∈{1,…,n}.j\in\mathcal A_h(x) \quad\Longleftrightarrow\quad r_j\cdot x+c_j(h)\le r_k\cdot x+c_k(h) \quad\text{for every }k\in\{1,\ldots,n\}.(4)✦

Uniform affine approximation in dimension dd is the property that every function E:ℝd→(0,∞)E:\mathbb R^d\to(0,\infty) admits a nonempty finite affine family and a threshold h0>0h_0>0, with slopes rjr_j and offset functions cjc_j fixed independently of hh, satisfying (5). ∥p−rj∥∞<E(rj)(0<h<h0,p∈[−1,1]d,j∈𝒜h(Xh(p))).\|p-r_j\|_\infty<E(r_j) \quad \left( 0<h<h_0,\quad p\in[-1,1]^d,\quad j\in\mathcal A_h(X_h(p)) \right).(5)✦

For a reaction network NN, reaction rates κ=(κy→z)y→z∈ℛ\kappa=(\kappa_{y\to z})_{y\to z\in\mathcal R} and an initial state x0∈ℝdx_0\in\mathbb R^d, the affine barrier data form the set ℬ(N,κ,x0)\mathcal B(N,\kappa,x_0) of tuples consisting of a parameter 0<h<10<h<1, a nonempty finite affine family at hh, and a distinguished label j*j_*, satisfying (6). rj*=0,j*∈𝒜h(x0),h<x0,i<h−1(1≤i≤d),rj⋅f(Xh(p))≥0(p∈[−1,1]d,j∈𝒜h(Xh(p))),j*∉𝒜h(Xh(p))(p∈[−1,1]d,∃i:|pi|=1).\begin{aligned} r_{j_*}&=0,\qquad j_*\in\mathcal A_h(x_0), &h&<x_{0,i}<h^{-1}\quad(1\le i\le d),\\ r_j\cdot f(X_h(p))&\ge0 &&\bigl(p\in[-1,1]^d,\ j\in\mathcal A_h(X_h(p))\bigr),\\ j_*&\notin\mathcal A_h(X_h(p)) &&\bigl(p\in[-1,1]^d,\ \exists i:\ |p_i|=1\bigr). \end{aligned}(6)✦

The Claim

Theorem 1. For every finite direction set D⊆ℝdD\subseteq\mathbb R^d, there is a sign tolerance E:ℝd→ℝE:\mathbb R^d\to\mathbb R satisfying (7). 0<E(r)≤12(r∈ℝd),(r⋅v)(p⋅v)>0,12|r⋅v|<|p⋅v|(r,p∈ℝd,∥p−r∥∞<E(r),v∈D,r⋅v≠0).\begin{gathered} 0<E(r)\le\tfrac12\qquad(r\in\mathbb R^d),\\ (r\cdot v)(p\cdot v)>0,\qquad \tfrac12|r\cdot v|<|p\cdot v|\\ \bigl(r,p\in\mathbb R^d,\ \|p-r\|_\infty<E(r),\ v\in D,\ r\cdot v\ne0\bigr). \end{gathered}(7)✦

For finite slope and direction sets S,D⊆ℝdS,D\subseteq\mathbb R^d, a uniform comparison gap HH exists as in (8). 0<H≤1,H≤12|r⋅v|(r∈S,v∈D,r⋅v≠0).0<H\le1,\qquad H\le\tfrac12|r\cdot v| \quad(r\in S,\ v\in D,\ r\cdot v\ne0).(8)✦

Given a positive initial state x0∈ℝdx_0\in\mathbb R^d, any finite family of positive rates (κe)e∈I(\kappa_e)_{e\in I}, and H,h0>0H,h_0>0, a parameter hh and initial exponents pp can be chosen with (9). 0<h<h0,h<1,p∈(−H,H)d,Xh(p)=x0,hH∑e∈Iκe<κe(e∈I).\begin{gathered} 0<h<h_0,\qquad h<1,\qquad p\in(-H,H)^d,\\ X_h(p)=x_0,\qquad h^H\sum_{e\in I}\kappa_e<\kappa_e\quad(e\in I). \end{gathered}(9)✦

For 0<h≤10<h\le1, H∈ℝH\in\mathbb R, p∈ℝdp\in\mathbb R^d and y,z∈ℕdy,z\in\mathbb N^d with H≤p⋅(z−y)H\le p\cdot(z-y), the monomial comparison is (10). Xh(p)z≤hHXh(p)y.X_h(p)^z\le h^H X_h(p)^y.(10)✦

Let NN be weakly reversible, with positive rates κ\kappa and positive initial state x0x_0. Under uniform affine approximation in dimension dd, affine barrier data exist as stated in (11). ℬ(N,κ,x0)≠⌀.\mathcal B(N,\kappa,x_0)\ne\varnothing.(11)✦

In the paper, the result reads: for the fixed network, positive rates and positive x0x_0 of Theorem 1.1, take D={ei:1≤i≤d}∪{z−y:y,z∈C,y≠z}D=\{e_i:1\le i\le d\}\cup\{z-y:y,z\in C,\ y\ne z\}, set E(r)=min⁡({1/2}∪{|r⋅v|/(2∥v∥1):v∈D,r⋅v≠0})E(r)=\min\bigl(\{1/2\}\cup \{|r\cdot v|/(2\|v\|_1):v\in D,\ r\cdot v\ne0\}\bigr), where ∥v∥1=∑i|vi|\|v\|_1=\sum_i|v_i|, choose a fixed 0<H≤10<H\le1 from the finitely many nonzero comparisons after fixing the affine family, and only then choose 0<h<h00<h<h_0 with |log⁡x0,i/log⁡h|<H|\log x_{0,i}/\log h|<H for every ii and, when there are reactions, hH∑y→z∈ℛκy→z<min⁡y→z∈ℛκy→zh^H\sum_{y\to z\in\mathcal R}\kappa_{y\to z} <\min_{y\to z\in\mathcal R}\kappa_{y\to z}, so that an active strict slope comparison r⋅y<r⋅zr\cdot y<r\cdot z implies xz≤hHxyx^z\le h^H x^y throughout Bh=[h,h−1]dB_h=[h,h^{-1}]^d.

What Is Checked

The comparison statements concern finite direction and slope sets; the barrier existence statement concerns a weakly reversible finite reaction network with positive rates and a positive initial state, under uniform affine approximation in its species dimension. Initial and rate separation concerns a positive initial state, a finite family of positive rates, and positive bounds HH and h0h_0; the monomial comparison concerns 0<h≤10<h\le1, a real number HH, a real exponent vector pp, and complexes y,z∈ℕdy,z\in\mathbb N^d satisfying H≤p⋅(z−y)H\le p\cdot(z-y).