OctoInformBeta

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

§3.4 (setting): Forward invariance requires confinement and global extension of finite solutions

Affine barriers give compact positive plateaux.

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

A mass-action trajectory evolves in concentration space, where affine functions assign scalar values to states. A reaction network N=(C,R)N=(C,R) on dd species consists of a finite complex set C⊆ℕdC\subseteq\mathbb N^d and a finite reaction set R⊆{(y,z)∈C×C:y≠z}R\subseteq\{(y,z)\in C\times C:y\ne z\}.✦ A network reaction y→zy\to z is an ordered pair in RR.✦ A positive state x∈ℝdx\in\mathbb R^d has xi>0x_i>0 for every species index ii; write ℝ>0d\mathbb R^d_{>0} for the set of positive states.✦

Fix positive constant reaction rates κy→z\kappa_{y\to z}. For real vectors r,xr,x, the dot product, the mass-action monomial associated with a complex yy, and the mass-action vector field ff are given by (1): r⋅x=∑i=1drixi,xy=∏i=1dxiyi,f(x)=∑y→z∈Rκy→zxy(z−y).\begin{aligned} r\cdot x&=\sum_{i=1}^d r_i x_i, &x^y&=\prod_{i=1}^d x_i^{y_i},\\ f(x)&=\sum_{y\to z\in R}\kappa_{y\to z}x^y(z-y). \end{aligned}(1)

For 0<h<10<h<1, an affine family indexed by a set Λ\Lambda is a collection of affine functions on concentration space with slopes rj∈ℝdr_j\in\mathbb R^d and offsets cj(h)∈ℝc_j(h)\in\mathbb R, whose values are given in (2). An active affine label j∈Λj\in\Lambda at a state xx is a label belonging to 𝒜(x)\mathcal A(x) in (2); every tied minimizing branch is included.✦ ℓj(x)=rj⋅x+cj(h),𝒜(x)={j∈Λ:ℓj(x)≤ℓk(x) for every k∈Λ}.\ell_j(x)=r_j\cdot x+c_j(h),\qquad \mathcal A(x)=\{j\in\Lambda:\ell_j(x)\le\ell_k(x) \text{ for every }k\in\Lambda\}.(2)

The reciprocal box BhB_h is the subset of concentration space given in (3), and the power coordinates Xh(p)X_h(p), for p∈ℝdp\in\mathbb R^d, are the vector given there. Bh=[h,h−1]d,Xh(p)=(hpi)i=1d.B_h=[h,h^{-1}]^d,\qquad X_h(p)=(h^{p_i})_{i=1}^d.(3)

For an initial state x0x_0 in concentration space, affine barrier data for (N,κ,x0)(N,\kappa,x_0) are a finite nonempty affine family, indexed by Λ={1,…,n}\Lambda=\{1,\ldots,n\} with n≥1n\ge1, a parameter h∈(0,1)h\in(0,1), and a distinguished branch j0∈Λj_0\in\Lambda satisfying (4).✦ The last two conditions are the inward condition and boundary exclusion, respectively. rj0=0,j0∈𝒜(x0),h<x0,i<h−1(1≤i≤d),rj⋅f(Xh(p))≥0(p∈[−1,1]d,j∈𝒜(Xh(p))),j0∉𝒜(Xh(p))(p∈[−1,1]d,|pi|=1 for some i).\begin{aligned} r_{j_0}&=0,\qquad j_0\in\mathcal A(x_0),\qquad 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(X_h(p))\bigr),\\ j_0&\notin\mathcal A(X_h(p)) &&\bigl(p\in[-1,1]^d,\ |p_i|=1\text{ for some }i\bigr). \end{aligned}(4)

Forward Solutions

For u∈ℝdu\in\mathbb R^d and T≥0T\ge0, a forward solution from uu on [0,T][0,T] is a map x:ℝ→ℝdx:\mathbb R\to\mathbb R^d belonging to the set 𝒮T(u)\mathcal S_T(u) defined in (5). 𝒮T(u)={x:ℝ→ℝd:x(0)=u,x|[0,T] is continuous,x′(t)=f(x(t))for 0<t<T}.\mathcal S_T(u)=\left\{x:\mathbb R\to\mathbb R^d: \begin{array}{l} x(0)=u,\quad x|_{[0,T]}\text{ is continuous},\\ x'(t)=f(x(t))\quad\text{for }0<t<T \end{array}\right\}.(5)✦ A global forward solution from uu is a map x:ℝ→ℝdx:\mathbb R\to\mathbb R^d belonging to 𝒮∞(u)\mathcal S_\infty(u) in (6). Derivatives in both definitions are ordinary two-sided derivatives, including the derivative at zero in (6). 𝒮∞(u)={x:ℝ→ℝd:x(0)=u,x′(t)=f(x(t)) for every t≥0}.\mathcal S_\infty(u)=\{x:\mathbb R\to\mathbb R^d: x(0)=u,\quad x'(t)=f(x(t))\text{ for every }t\ge0\}.(6)✦

The Claim

Definition 1. A finite polyhedron P⊆ℝdP\subseteq\mathbb R^d is a set of the form (7), for some integer m≥0m\ge0, vectors a1,…,am∈ℝda_1,\ldots,a_m\in\mathbb R^d, and constants b1,…,bm∈ℝb_1,\ldots,b_m\in\mathbb R. P=⋂j=1m{x∈ℝd:bj≤aj⋅x}.P=\bigcap_{j=1}^m\{x\in\mathbb R^d:b_j\le a_j\cdot x\}.(7)✦

The barrier plateau KK of affine barrier data with distinguished constant branch j0j_0 is the set in (8), with M=cj0(h)M=c_{j_0}(h). K={x∈ℝd:M≤rj⋅x+cj(h) for every j∈Λ}.K=\{x\in\mathbb R^d:M\le r_j\cdot x+c_j(h) \text{ for every }j\in\Lambda\}.(8)✦

A set K⊆ℝdK\subseteq\mathbb R^d is forward invariant for the mass-action system if it satisfies the existence, confinement, and extension conditions in (9). 𝒮∞(u)≠⌀(u∈K),x([0,T])⊆K,∃xg∈𝒮∞(u):xg|[0,T]=x|[0,T](u∈K,T≥0,x∈𝒮T(u)).\begin{aligned} &\mathcal S_\infty(u)\ne\varnothing && (u\in K),\\ &x([0,T])\subseteq K,\qquad \exists x_g\in\mathcal S_\infty(u): x_g|_{[0,T]}=x|_{[0,T]} && \bigl(u\in K,\ T\ge0,\ x\in\mathcal S_T(u)\bigr). \end{aligned}(9)✦

Theorem 1. Let affine barrier data for (N,κ,x0)(N,\kappa,x_0) be given, and let KK be their plateau. Its geometric properties are (10). K is compact and convex and is a finite polyhedron,x0∈K,K⊆ℝ>0d.\begin{gathered} K\text{ is compact and convex and is a finite polyhedron},\\ x_0\in K,\qquad K\subseteq\mathbb R^d_{>0}. \end{gathered}(10)✦

In the paper, the result reads: every branch active at the prescribed x0x_0 has zero slope; choosing one such constant branch of value MM, the affine minimum F(x)=min⁡j∈Λ(rj⋅x+cj(h))F(x)=\min_{j\in\Lambda}(r_j\cdot x+c_j(h)) satisfies F≤MF\le M everywhere, its maximum set K={x:F(x)=M}=⋂j∈Λ{x:rj⋅x+cj(h)≥M}K=\{x:F(x)=M\}=\bigcap_{j\in\Lambda}\{x:r_j\cdot x+c_j(h)\ge M\} contains x0x_0, misses ∂Bh\partial B_h, and by convexity lies in int⁡Bh\operatorname{int}B_h, and forward invariance means that every solution from every point of KK exists for all forward time and remains in KK.

What Is Checked

The geometric statement is over affine barrier data for a finite reaction network with fixed rates and initial state, and forward invariance is stated over the finite and global forward solutions of its mass-action vector field.