betaFeedback

WorksComputer Science

Global Convergence of Detailed Balanced Mass Action Systems

Abstract

We give a self-contained proof that, for a finite detailed balanced mass-action system with positive rate constants, each positive stoichiometric compatibility class contains exactly one positive equilibrium and every positive trajectory converges to it, together with a Lean 4 formalization of the complete argument from persistence to convergence. Boundedness and persistence come from the barrier argument of OpenAI (2026) for weakly reversible systems: an arbitrary-tolerance affine approximation yields a finite-minimum barrier whose superlevel sets confine trajectories and give one logarithmic bound for all solutions from a given initial state. The Horn-Jackson dissipation identity identifies the positive equilibria as an exponential family with parameters in the orthogonal complement of the stoichiometric subspace, which gives uniqueness in each class. A bounded-speed descent argument replaces LaSalle's invariance principle; it uses neither uniqueness of solutions nor continuous dependence on initial data, and it places every cluster point in that family. The route given here is direct and elementary, and its main theorem is checked by the Lean kernel against explicit definitions. The proofs, the Lean development, and the manuscript were produced in ClawsGO Science, an agent harness that packages a complete research environment: Claude Opus 5.5 made the research decisions, GPT 6 Astra carried them out, and the Lean kernel checked every formal step.

Introduction

We consider finite mass-action reaction networks with positive constant rate coefficients and a positive detailed balanced equilibrium. The classical theory of Horn and Jackson [11 pp. 81–116], Feinberg [5 pp. 187–194], and Horn [9 pp. 172–186] establishes the equilibrium structure of complex balanced systems, including existence and uniqueness in each positive compatibility class and local asymptotic stability. Horn [10 pp. 125–137] formulated the global attractor conjecture, which asks whether every positive trajectory converges to the positive equilibrium in its compatibility class.

The entropy argument must be supplemented by an exclusion of boundary accumulation. Without this step, it permits an ω\omega-limit set contained in the positive equilibria ∪ the boundary [14 pp. 89–110]. Earlier partial results include the single linkage class case [1], low-dimensional systems [13], and strongly endotactic networks [8]. Craciun [3], Feinberg [5] gives a proof of the global attractor conjecture for complex balanced systems using toric differential inclusions.

Theorem 3.1 and § 3.1.1–§ 3.1.7 build on OpenAI’s work on boundedness and persistence of weakly reversible mass-action systems [15]; the detailed-balance argument developed here uses this theorem as input. The accompanying Lean modules for that part are adapted from the same work, with paths, namespaces, and principal declarations reorganized for this project.

Theorem 1.1 (Main theorem). Assume that the network is finite, all rate constants are positive, and the mass-action system is detailed balanced at some x∗∈R>0dx^* \in\mathbb{R}_{>0}^d. Then, for every x0∈R≥0dx_0 \in\mathbb{R}_{\geq0}^d, the following hold:

(i) The positive compatibility class (x0+S)∩R≥0d(x_0+S)\cap\mathbb{R}_{\geq0}^d contains exactly one equilibrium xˉ\bar{x}.

(ii) Equation (1) has a solution starting at x0x_0 and defined on [0,∞)[0,\infty).

(iii) Every such solution satisfies x(t)→xˉx(t)\to\bar{x} as t→∞t\to\infty.

Contributions. (1) A self-contained proof of Theorem 1.1 that joins persistence and the entropy argument into one chain. Theorem 3.1, stated for finite weakly reversible systems without detailed balance, supplies one compact container for all forward solutions from a given initial state (Corollary 3.15), and Steps 2–5 use Step 1 only through this container. The detailed-balance argument, this reduction, and the exact examples were developed for this paper.

(2) In Step 5, a bounded-speed descent argument for a non-increasing function bounded below replaces LaSalle’s invariance principle. It uses neither continuous dependence on initial conditions nor uniqueness of solutions.

(3) A machine-checked proof of the whole chain from persistence to global convergence in Lean 4 [4] with Mathlib [12]. The formal main theorem quantifies over all global forward solutions, and the exact three-species example of § 5.1 instantiates it. § 4 gives the correspondence between the paper and the formal statements and the reproducible build.

Relation to previous work. Existence, uniqueness, and the entropy structure are classical. Craciun’s proof of the global attractor theorem [3 Section 8] covers this case through toric differential inclusions; the route here is direct, uses elementary tools, and is machine-checked from persistence to convergence.

How the work was produced. The work was carried out end to end in ClawsGO Science [16], an agent harness that packages a complete research environment. Claude Opus 5.5 made the research decisions: it reduced Steps 2–5 to a single compact container for all solutions from a given initial state (Corollary 3.15), replaced LaSalle’s invariance principle by the bounded-speed descent argument of Step 5, and chose to formalize the whole chain from persistence to convergence on top of the adapted persistence modules. GPT 6 Astra carried out these decisions: it wrote the proofs and the Lean modules, adapted the 71 persistence modules, ran the builds and axiom audits, and drafted the manuscript. The Lean kernel checked every formal step, and the human authors directed the work and reviewed the results.

Proof strategy and organization. § 2 fixes the notation. § 3.1 constructs an affine barrier, proves confinement through feasible directions, and obtains a common logarithmic bound and a compact container for the trajectories. § 3.2 proves entropy dissipation, parameterizes the positive equilibria, establishes uniqueness, and treats cluster points. § 3.3 combines these steps. § 4 records the formal statements, their correspondence with the paper, and the scope of verification. § 5 gives examples and discusses persistence, complex balance, and boundary initial conditions.

Mass Action Systems and Detailed Balance

Mass action systems

Let there be d≥1d\geq1 species. A complex is a vector y∈Ndy\in\mathbb{N}^d, and C⊆Nd\mathcal{C}\subseteq\mathbb{N}^d is a finite set of complexes. A reaction network is a finite set R⊆C×C\mathcal{R}\subseteq\mathcal{C}\times\mathcal{C}. We write a reaction e=(y,y′)e=(y,y') as y→y′y\to y' and require y≠y′y\neq y'. Its source and target are s(e)=ys(e)=y and t(e)=y′t(e)=y', respectively. Each reaction has a rate constant κe>0\kappa_e>0, and we put κtot=∑e∈Rκe\kappa_{\mathrm{tot}}=\sum_{e\in\mathcal{R}}\kappa_e. For x∈Rdx\in\mathbb{R}^d, write the monomial xy=∏i=1dxiyix^y=\prod_{i=1}^d x_i^{y_i} and the inner product a⋅x=∑iaixia\cdot x=\sum_i a_i x_i. The mass-action equation is

x˙=f(x):=∑y→y′∈Rκy→y′xy(y′−y).(1)\dot{x}=f(x):=\sum_{y\to y'\in\mathcal{R}}\kappa_{y\to y'}x^y(y'-y). \tag*{(1)}

The stoichiometric subspace is S=span⁡{y′−y:y→y′∈R}S = \operatorname{span}\{y^\prime-y:y\to y^\prime\in\mathcal{R}\}. For x0∈R>0dx_0\in\mathbb{R}^d_{>0}, the set (x0+S)∩R>0d(x_0+S)\cap\mathbb{R}^d_{>0} is the positive stoichiometric compatibility class of x0x_0. The map ff is polynomial, and (1) gives f(x)∈Sf(x)\in S for every xx.

Definition 2.1 (Solutions). Let x0∈Rdx_0\in\mathbb{R}^d. (i) A map x:R→Rdx:\mathbb{R}\to\mathbb{R}^d is a global forward solution starting at x0x_0 if x(0)=x0x(0)=x_0 and, for every t≥0t\ge0, the map xx is differentiable at tt with derivative f(x(t))f(x(t)).

(ii) Let T≥0T\ge0. A map xx is a solution on [0,T][0,T] if x(0)=x0x(0)=x_0, the map xx is continuous on [0,T][0,T], and it is differentiable at every t∈(0,T)t\in(0,T) with derivative f(x(t))f(x(t)).

Definition 2.1(i) requires a two-sided derivative at t=0t=0, so xx must also be defined to the left of t=0t=0. This is the formalization convention used in Section 4. Since ff is polynomial, any classical forward solution extends a short distance backwards, so the convention does not restrict the result. The confinement argument of Lemma 3.10 and the descent argument of Step 5 apply to every solution without using uniqueness. The complete Lean proof also constructs a stronger forward-invariance property, including extension and agreement of finite solution segments; this construction uses standard ODE uniqueness for the locally Lipschitz polynomial vector field.

Reversibility and detailed balance

Definition 2.2. (i) A network is reversible if y→y′∈Ry\to y^\prime\in\mathcal{R} implies y′→y∈Ry^\prime\to y\in\mathcal{R}. In this case eˉ=(y′,y)\bar{e}=(y^\prime,y) denotes the reverse of e=(y,y′)e=(y,y^\prime), and e↦eˉe\mapsto\bar{e} is an involution on R\mathcal{R}.

(ii) A network is weakly reversible if, for every y→y′∈Ry\to y^\prime\in\mathcal{R}, there is a directed path y′=c0→c1→⋯→ck=yy^\prime=c_0\to c_1\to\cdots\to c_k=y consisting of reactions in R\mathcal{R}, with k≥1k\ge1.

(iii) A system is detailed balanced if the network is reversible and there exists x∗∈R>0dx^*\in\mathbb{R}^d_{>0} such that

κe(x∗)s(e)=κeˉ(x∗)t(e)=:Ke(e∈R).(2)\kappa_e(x^*)^{s(e)}=\kappa_{\bar{e}}(x^*)^{t(e)}=:K_e\qquad(e\in\mathcal{R}). \tag*{(2)}

By definition, Ke=Keˉ>0K_e=K_{\bar{e}}>0.

Lemma 2.3. Every reversible network is weakly reversible.

Proof. For e=(y,y′)e=(y,y^\prime), the reverse reaction eˉ=(y′,y)∈R\bar{e}=(y^\prime,y)\in\mathcal{R} itself gives a path from y′y^\prime to yy of length 1. □

Remark 2.4 (Existence of detailed balance). Assume the network is reversible. Choose a representative direction νp=yp′−yp\nu_p=y^\prime_p-y_p for each reverse pair and put ρp=ln⁡(κyp→yp′/κyp′→yp)\rho_p=\ln(\kappa_{y_p\to y^\prime_p}/\kappa_{y^\prime_p\to y_p}). Setting u=ln⁡x∗u=\ln x^* componentwise, (2) is equivalent to the linear system νp⋅u=ρp\nu_p\cdot u=\rho_p. This system is solvable if and only if every real linear relation ∑pαpνp=0\sum_p\alpha_p\nu_p=0, with αp∈R\alpha_p\in\mathbb{R}, implies ∑pαpρp=0\sum_p\alpha_p\rho_p=0. These are the Wegscheider conditions [6 pp. 1819–1827]. In particular, if the vectors νp\nu_p are linearly independent, all positive choices of rate constants admit detailed balance.

Notation

We use ∣⋅∣|\cdot| for the maximum norm on Rd\mathbb{R}^d. For h∈(0,1)h\in(0,1) and p∈Rdp\in\mathbb{R}^d, write hp:=(hp1,…,hpd)∈R>0dh^p:=(h^{p_1},\ldots,h^{p_d})\in\mathbb{R}^d_{>0}. The map p↦hpp\mapsto h^p is a bijection from Rd\mathbb{R}^d to R>0d\mathbb{R}^d_{>0} with inverse pi=ln⁡xi/ln⁡hp_i=\ln x_i/\ln h. It maps the cube {∣p∣≤1}\{|p|\le1\} to the box

Bh:=[h,h−1]d,B_h:=[h,h^{-1}]^d,

and maps the boundary {∣p∣=1}\{|p|=1\} of the cube to the topological boundary ∂Bh\partial B_h of the box. The monomial definition gives

(hp)z(hp)y=hp⋅(z−y)(y,z∈Nd).(3)\frac{(h^p)^z}{(h^p)^y}=h^{p\cdot(z-y)}\qquad(y,z\in\mathbb{N}^d). \tag*{(3)}

Proof of the Main Theorem

We prove Theorem 1.1 in five steps. Section 3.1 proves boundedness and persistence for weakly reversible systems (Theorem 3.1), without (2), and applies the result to detailed balanced systems to obtain a common compact container for all solutions from the fixed initial state (Corollary 3.15). Section 3.2 supplies Steps 2–5 of the Horn–Jackson argument: entropy dissipation, identification of positive equilibria, uniqueness in each class, and identification of cluster points. Section 3.3 finishes the proof. Steps 2–5 use Step 1 only through Corollary 3.15.

Boundedness and persistence for weakly reversible systems

Statement

Theorem 3.1 (Logarithmic confinement). Assume d≥1d \ge1, the network is finite and weakly reversible, the rate constants satisfy κe>0\kappa_e > 0, and x0∈R>0dx_0 \in\mathbb{R}_{>0}^{d}. Then:

(a) There exists a global forward solution starting at x0x_0.

(b) There exists L≥0L \ge0 such that every global forward solution starting at x0x_0 is positive at all nonnegative times and satisfies

∣ln⁡xi(t)∣≤L(t≥0, i=1,…,d).(4)\lvert\ln x_i(t) \rvert\le L \qquad(t \ge0,\ i = 1,\ldots,d). \tag*{(4)}

The same LL works for all these solutions and all coordinates; it may depend on the network, rate constants, and x0x_0.

Here a global forward solution is understood as in Definition 2.1(i). Theorem 3.1 assumes only weak reversibility and positive constant rates; detailed balance (2) is not required.

Exponentiating (4) places every trajectory in the compact box [e−L,eL]d[e^{-L},e^L]^d inside the positive orthant. Conversely, on any nonempty compact subset of the positive orthant the finitely many continuous functions z↦∣ln⁡zi∣z \mapsto\lvert\ln z_i \rvert have a common finite maximum. Thus logarithmic confinement expresses both boundedness and separation from the boundary. We will construct one compact plateau containing all trajectories from x0x_0 and obtain LL from it.

Strategy

We construct a barrier as a finite minimum of affine functions. Let ℓj(x)=rj⋅x+cj\ell_j(x) = r_j \cdot x + c_j for j=1,…,nj = 1,\ldots,n and put

F(x)=min⁡1≤j≤nℓj(x).F(x) = \min_{1 \le j \le n} \ell_j(x).

An index jj is active at xx if ℓj(x)=F(x)\ell_j(x) = F(x), equivalently ℓj(x)≤ℓk(x)\ell_j(x) \le\ell_k(x) for every kk. Every minimizer is active when there is a tie. We construct data with the following properties.

  • There is a constant branch j0j_0, with rj0=0r_{j_0} = 0 and ℓj0≡M\ell_{j_0} \equiv M, that is active at x0x_0. Thus F≤MF \le M, and the plateau P={F=M}P = \{F = M\} contains x0x_0.

  • At every z∈Bhz \in B_h, the vector f(z)f(z) is a feasible direction of the superlevel set {y:F(y)≥F(z)}\{y : F(y) \ge F(z)\}: a sufficiently short forward segment z+sf(z)z + s f(z) stays in that set. The Lipschitz continuity of FF then makes FF non-decreasing along trajectories.

  • The constant branch is inactive on ∂Bh\partial B_h, so the plateau does not meet the boundary of the box.

Consequently, F(x(t))F(x(t)) is bounded below by MM and above by MM, and the trajectory remains on the plateau. The plateau is a compact convex polyhedron in BhB_h, on which the logarithmic coordinates are uniformly bounded, giving (4).

The feasible-direction criterion of Lemma 3.8 reduces this property to rj⋅f(z)≥0r_j \cdot f(z) \ge0 for every constraint tight at level F(z)F(z), namely every active branch. To obtain these inequalities, write z=hpz = h^p. By (3), zy=hp⋅yz^y = h^{p \cdot y}. For small hh, complexes minimizing p⋅yp \cdot y have dominant monomials in ff. If the ordering induced by rjr_j agrees with that induced by pp, the leading contributions to rj⋅fr_j \cdot f point upwards from the lowest levels (Lemma 3.2). The main difficulty is to make every active direction close enough to pp while fixing all directions before choosing $h ({xref}Proposition 3.5`).

Network lemmas

Lemma 3.2 (Cut-flux dominance). Assume that the network is weakly reversible, x∈R>0dx \in\mathbb{R}_{>0}^d, r∈Rdr \in\mathbb{R}^d, δ≥0\delta\ge0, and the following conditions hold:

(O) For all y,z∈Cy, z \in\mathcal{C}, the inequality r⋅y<r⋅zr \cdot y < r \cdot z implies xz≤δxyx^z \le\delta x^y.

(Q) For each reaction ee, we have δκtot<κe\delta\kappa_{\mathrm{tot}} < \kappa_e.

Then r⋅f(x)≥0r \cdot f(x) \ge0.

Proof. Define the height a(y)=r⋅ya(y) = r \cdot y. For c∈Rc \in\mathbb{R}, let

Uc={e:a(s(e))≤c<a(t(e))},Dc={e:a(t(e))≤c<a(s(e))},U_c = \{e : a(s(e)) \le c < a(t(e))\}, \qquad D_c = \{e : a(t(e)) \le c < a(s(e))\},

be the upward and downward reactions crossing height cc, respectively, and put

Q(c)=∑e∈Ucκexs(e)−∑e∈Dcκexs(e).Q(c) = \sum_{e \in U_c} \kappa_e x^{s(e)} - \sum_{e \in D_c} \kappa_e x^{s(e)}.

(i) We have Q(c)≥0Q(c) \ge0. This is immediate if Dc=∅D_c = \varnothing. Otherwise choose e∈Dce \in D_c. Weak reversibility gives a directed path from t(e)t(e) to s(e)s(e). Its initial height is ≤c\le c and its final height is >c> c, so some reaction gg on the path satisfies a(s(g))≤c<a(t(g))a(s(g)) \le c < a(t(g)), that is, g∈Ucg \in U_c. For any e′∈Dce' \in D_c, we have a(s(e′))>c≥a(s(g))a(s(e')) > c \ge a(s(g)), and (O) gives xs(e′)≤δxs(g)x^{s(e')} \le\delta x^{s(g)}. By (Q) and xs(g)>0x^{s(g)} > 0,

∑e′∈Dcκe′xs(e′)≤δκtotxs(g)<κgxs(g)≤∑e′′∈Ucκe′′xs(e′′).\sum_{e' \in D_c} \kappa_{e'}x^{s(e')} \le\delta\kappa_{\mathrm{tot}}x^{s(g)} < \kappa_gx^{s(g)} \le\sum_{e'' \in U_c} \kappa_{e''}x^{s(e'')}.

(ii) Sum over cuts. Arrange the distinct values in {a(y):y∈C}\{a(y) : y \in\mathcal{C}\} as a1<⋯<ama_1 < \cdots< a_m. If a reaction ee satisfies a(s(e))=aσ<aτ=a(t(e))a(s(e)) = a_\sigma< a_\tau= a(t(e)), then aτ−aσ=∑k=στ−1(ak+1−ak)a_\tau- a_\sigma= \sum_{k=\sigma}^{\tau-1}(a_{k+1} - a_k), and e∈Uake \in U_{a_k} holds exactly for k=σ,…,τ−1k = \sigma, \ldots, \tau- 1. The downward case is analogous; equal-height reactions cross no aka_k. Hence

r⋅f(x)=∑eκexs(e)(a(t(e))−a(s(e)))=∑k=1m−1(ak+1−ak)Q(ak)≥0.r \cdot f(x) = \sum_e \kappa_e x^{s(e)}\bigl(a(t(e)) - a(s(e))\bigr) = \sum_{k=1}^{m-1}(a_{k+1} - a_k)Q(a_k) \ge0.

Alternatively, (ii) can be expressed as an integral of Q(c)Q(c) with respect to cc, using ∫a(s(e))a(t(e))dc=a(t(e))−a(s(e))\int_{a(s(e))}^{a(t(e))} dc = a(t(e)) - a(s(e)), without sorting the heights. The integral and finite cut-sum formulations are equivalent. Condition (Q) is imposed reaction by reaction, so the empty reaction set needs no separate argument. One upward reaction dominates the total downward flux; reverse reactions need not be paired. This is why weak reversibility suffices.

Lemma 3.3 (Monomial ordering from an exponent gap). Let h∈(0,1]h \in(0,1], p∈Rdp \in\mathbb{R}^d, H∈RH \in\mathbb{R}, and y,z∈Ndy,z \in\mathbb{N}^d. If p⋅(z−y)≥Hp \cdot(z-y) \ge H, then (hp)z≤hH(hp)y(h^p)^z \le h^H(h^p)^y.

Proof. By (3), (hp)z=hp⋅(z−y)(hp)y(h^p)^z = h^{p \cdot(z-y)}(h^p)^y. For h≤1h \le1, the map s↦hss \mapsto h^s is non-increasing, so hp⋅(z−y)≤hHh^{p \cdot(z-y)} \le h^H.

Lemma 3.4 (Sign tolerance). Let D⊂Rd\mathcal{D} \subset\mathbb{R}^d be finite.

(a) There exists a function E:Rd→(0,12]E : \mathbb{R}^d \to\left(0,\frac{1}{2}\right] such that, for any r,p∈Rdr,p \in\mathbb{R}^d, if ∣pi−ri∣<E(r)|p_i-r_i| < E(r) for all ii, then, for every v∈Dv \in\mathcal{D} satisfying r⋅v≠0r \cdot v \ne0, the numbers p⋅vp \cdot v and r⋅vr \cdot v have the same sign and ∣p⋅v∣>∣r⋅v∣/2|p \cdot v| > |r \cdot v|/2.

(b) For every finite set A⊂Rd\mathcal{A} \subset\mathbb{R}^{d}, there exists H∈(0,1]H \in(0,1] such that H≤∣r⋅v∣/2H \leq|r \cdot v|/2 for all r∈Ar \in\mathcal{A} and v∈Dv \in\mathcal{D} with r⋅v≠0r \cdot v \neq0.

Proof. (a) Put L=1+max⁡({0}∪{∑i∣vi∣:v∈D})L = 1 + \max\left(\{0\} \cup\left\{\sum_i |v_i| : v \in\mathcal{D}\right\}\right). For fixed rr, let m(r)=min⁡{∣r⋅v∣:v∈D, r⋅v≠0}m(r) = \min\{|r \cdot v| : v \in\mathcal{D},\ r \cdot v \neq0\}, taking m(r)=1m(r) = 1 when the set is empty, and define E(r)=min⁡{12,m(r)2L}>0E(r) = \min\left\{\frac{1}{2}, \frac{m(r)}{2L}\right\} > 0. If ∣pi−ri∣<E(r)|p_i-r_i| < E(r) for every ii, and v∈Dv \in\mathcal{D} satisfies r⋅v≠0r \cdot v \neq0, then v≠0v \neq0 and

∣p⋅v−r⋅v∣≤max⁡i∣pi−ri∣∑∣vi∣<E(r)L≤12m(r)≤12∣r⋅v∣.|p \cdot v-r \cdot v| \leq\max_i |p_i-r_i| \sum|v_i| < E(r)L \leq\frac{1}{2}m(r) \leq\frac{1}{2}|r \cdot v|.

Thus ∣p⋅v∣>12∣r⋅v∣>0|p \cdot v| > \frac{1}{2}|r \cdot v| > 0, and p⋅vp \cdot v and r⋅vr \cdot v have the same sign. (b) Take the minimum of 11 and all the finitely many positive numbers ∣r⋅v∣/2|r \cdot v|/2, taking H=1H = 1 if this collection is empty. □

The definition of LL includes zero in the finite set, so it gives L=1L = 1 when D=∅\mathcal{D} = \varnothing.

The function EE is chosen pointwise; no continuity is required.

Affine approximation

Proposition 3.5 (Uniform affine approximation). For every function E:Rd→(0,∞)E : \mathbb{R}^{d} \to(0,\infty), there exist n≥1n \geq1, directions r1,…,rn∈Rdr_1,\ldots,r_n \in\mathbb{R}^{d}, functions c1,…,cn:R→Rc_1,\ldots,c_n : \mathbb{R} \to\mathbb{R}, and h0>0h_0 > 0 such that the following holds. For every h∈(0,h0)h \in(0,h_0), every p∈Rdp \in\mathbb{R}^{d} satisfying ∣p∣≤1|p| \leq1, and every index jj active at hh for the affine family {rk⋅x+ck(h)}k\{r_k \cdot x+c_k(h)\}_k, we have

∣pi−rj,i∣<E(rj)(i=1,…,d).|p_i-r_{j,i}| < E(r_j) \qquad(i=1,\ldots,d).

The directions rjr_j and intercept functions cjc_j are fixed before hh is chosen; only the values cj(h)c_j(h) vary with hh. The conclusion holds for every active index, including ties, as required by the tight-constraint criterion in Lemma 3.8. The proposition depends only on the dimension dd and tolerance EE, not on the reaction network. Section 3.1.7 proves it by induction on dimension, first constructing local label families near a fixed minimum and then joining them along a chain of intervals.

Assembly of the barrier

Proposition 3.6 (Barrier data). Assume weak reversibility, κe>0\kappa_e > 0, and x0∈R>0dx_0 \in\mathbb{R}_{>0}^{d}. There exist h∈(0,1)h \in(0,1), n≥1n \geq1, affine functions ℓj(x)=rj⋅x+cj\ell_j(x) = r_j \cdot x+c_j for j=1,…,nj=1,\ldots,n, and an index j0j_0 such that:

(B1) We have rj0=0r_{j_0}=0, and j0j_0 is active at x0x_0.

(B2) We have x0∈int⁡Bhx_0 \in\operatorname{int} B_h, that is, h<x0,i<h−1h < x_{0,i} < h^{-1} for all ii.

(B3) Put F=min⁡jℓjF = \min_j \ell_j. For every z∈Bhz \in B_h, there exists δz>0\delta_z > 0 such that

F(z+sf(z))≥F(z)(0≤s≤δz).F(z+sf(z)) \geq F(z) \qquad(0 \leq s \leq\delta_z).

The step size may depend on zz; the condition is imposed throughout BhB_h.

(B4) For every x∈∂Bhx \in\partial B_h, the index j0j_0 is inactive at xx.

Proof. Order of choices. Put D={e1,…,ed}∪{z−y:y,z∈C}\mathcal{D} = \{e_1,\ldots,e_d\} \cup\{z-y : y,z \in\mathcal{C}\}, where eie_i are the coordinate vectors. Choose, in order, EE from Lemma 3.4(a); the data n,rj,cj(⋅),h0n,r_j,c_j(\cdot),h_0 from Proposition 3.5 for this EE; and H∈(0,1]H \in(0,1] from Lemma 3.4(b) for A={r1,…,rn}\mathcal{A} = \{r_1,\ldots,r_n\}. Finally choose h∈(0,min⁡{h0,1})h \in(0,\min\{h_0,1\}) such that

∣ln⁡x0,iln⁡h∣<H(i=1,…,d),hHκtot<κe(e∈R).(5)\left|\frac{\ln x_{0,i}}{\ln h}\right| < H \quad(i=1,\ldots,d), \qquad h^{H\kappa_{\mathrm{tot}}} < \kappa_e \quad(e \in\mathcal{R}). \tag*{(5)}

This is possible: as h↓0h \downarrow0, we have ∣ln⁡h∣→∞|\ln h| \to\infty and hH→0h^H \to0. Put pi0=ln⁡x0,i/ln⁡hp_i^0 = \ln x_{0,i}/\ln h. Then x0=hp0x_0 = h^{p^0} and ∣pi0∣<H≤1|p_i^0| < H \leq1. Fix the intercepts cj:=cj(h)c_j := c_j(h). Since the directions are fixed before hh is chosen, the gap HH is independent of hh, allowing both inequalities in (3.2) to hold.

Key estimate. Suppose ∣p∣≤1|p|\le1, the index jj is active at hph^{p}, and v∈Dv\in D satisfies rj⋅v≠0r_j\cdot v\ne0. Proposition 3.5 gives ∣pi−rj,i∣<E(rj)|p_i-r_{j,i}|<E(r_j), and Lemma 3.4 yields

sign⁡(p⋅v)=sign⁡(rj⋅v),∣p⋅v∣>∣rj⋅v∣/2≥H.(6)\operatorname{sign}(p\cdot v)=\operatorname{sign}(r_j\cdot v),\qquad|p\cdot v|>|r_j\cdot v|/2\ge H. \tag*{(6)}

For (B3), let x∈Bhx\in B_h and write x=hpx=h^p with ∣p∣≤1|p|\le1. Let jj be active at xx. Apply Lemma 3.2 with δ=hH\delta=h^H. Condition (Q) is the second inequality in (3.2). For (O), if y,z∈Cy,z\in C and rj⋅y<rj⋅zr_j\cdot y<r_j\cdot z, take v=z−y∈Dv=z-y\in D. Then rj⋅v>0r_j\cdot v>0, so (3.3) gives p⋅(z−y)>0p\cdot(z-y)>0 and p⋅(z−y)≥Hp\cdot(z-y)\ge H. Lemma 3.3 gives xz≤hHxyx^z\le h^H x^y. Therefore rj⋅f(x)≥0r_j\cdot f(x)\ge0.

Apply the finite-constraint criterion of Lemma 3.8 with level m=F(x)m=F(x), base point xx, and direction f(x)f(x). Its tight constraints are precisely the active branches just considered. It gives a positive step on which every ℓj\ell_j stays at least F(x)F(x), proving (B3). The criterion is an elementary affine calculation independent of the barrier construction.

For (B1), a finite non-empty family always has an active index. Choose j0j_0 active at x0x_0. If rj0,i≠0r_{j_0,i}\ne0, take v=eiv=e_i. (6), with p=p0p=p^0, gives ∣pi0∣≥H|p_i^0|\ge H, contradicting ∣pi0∣<H|p_i^0|<H. Thus rj0=0r_{j_0}=0.

For (B2), the inequalities ∣pi0∣<1|p_i^0|<1 and 0<h<10<h<1 give h=h1<hpi0<h−1h=h^1<h^{p_i^0}<h^{-1}.

For (B4), let x∈∂Bhx\in\partial B_h and write x=hpx=h^p. Then ∣p∣≤1|p|\le1, and ∣pi∣=1|p_i|=1 for some coordinate. If j0j_0 were active at xx, Proposition 3.5 would give ∣pi−0∣<E(0)≤12|p_i-0|<E(0)\le\frac12, a contradiction. □\square

The plateau and confinement

Take the data from Proposition 3.6 and put M=cj0M=c_{j_0}, F=min⁡jℓjF=\min_j\ell_j, and

P={x∈Rd:ℓj(x)≥M, j=1,…,n}.P=\left\{x\in\mathbb{R}^d:\ell_j(x)\ge M,\ j=1,\ldots,n\right\}.

Lemma 3.7 (Plateau geometry). (a) We have F≤MF\le M everywhere, and x∈Px\in P if and only if F(x)=MF(x)=M, equivalently j0j_0 is active at xx. (b) The set PP is an intersection of finitely many closed half-spaces, hence a closed convex polyhedron, and x0∈Px_0\in P. (c) We have P⊂int⁡BhP\subset\operatorname{int}B_h; in particular, PP is compact and contained in R>0d\mathbb{R}^d_{>0}.

Proof. (a) Since rj0=0r_{j_0}=0, we have ℓj0≡M\ell_{j_0}\equiv M and hence F≤MF\le M. The equality F(x)=MF(x)=M is equivalent to ℓj(x)≥M=ℓj0(x)\ell_j(x)\ge M=\ell_{j_0}(x) for every branch. (b) This follows from the definition and (B1). (c) Parts (a) and (B4) imply P∩∂Bh=∅P\cap\partial B_h=\varnothing. Thus P=(P∩int⁡Bh)∪(P∖Bh)P=(P\cap\operatorname{int}B_h)\cup(P\setminus B_h) is a disjoint union of relatively open sets. The set PP is convex and therefore connected. By (B2), the first set contains x0x_0 and is non-empty, so the second is empty. Thus PP is a closed subset of the compact box BhB_h and is compact. □\square

Lemma 3.8 (Feasible directions under finite affine constraints). Let JJ be finite, let ℓj(y)=rj⋅y+cj\ell_j(y)=r_j\cdot y+c_j for j∈Jj\in J, and fix a level m∈Rm\in\mathbb{R}. Put Qm={y:ℓj(y)≥m for all j∈J}Q_m=\{y:\ell_j(y)\ge m\text{ for all }j\in J\}. For z∈Qmz\in Q_m, a vector vv satisfies

z+sv∈Qm(0≤s≤δ)z+sv\in Q_m\qquad(0\le s\le\delta)

for some δ>0\delta>0 if and only if rj⋅v≥0r_j\cdot v\ge0 for every tight constraint jj, that is, every jj with ℓj(z)=m\ell_j(z)=m.

Proof. Write aj=ℓj(z)−m≥0a_j=\ell_j(z)-m\ge0 and bj=rj⋅vb_j=r_j\cdot v. Then

ℓj(z+sv)−m=aj+sbj.\ell_j(z+sv)-m=a_j+sb_j.

Suppose the tight constraints have nonnegative slopes. For aj=0a_j=0, set δj=1\delta_j=1; the displayed expression is nonnegative for all s≥0s\ge0. For aj>0a_j>0, set δj=aj/(∣bj∣+1)>0\delta_j=a_j/(|b_j|+1)>0. If 0≤s≤δj0\le s\le\delta_j, then

aj+sbj≥aj−s∣bj∣≥s≥0.a_j+sb_j\ge a_j-s|b_j|\ge s\ge0.

Thus δ=min⁡({1}∪{δj:j∈J})>0\delta=\min(\{1\}\cup\{\delta_j:j\in J\})>0 works for every constraint, including when JJ is empty. Conversely, a feasible segment and a tight constraint give 0≤aj+δbj=δbj0\le a_j+\delta b_j=\delta b_j, hence bj≥0b_j\ge0. □\square

For the barrier family, taking m=F(z)m = F(z) identifies the tight constraints with all active branches at zz. At a plateau point, m=Mm = M and Qm=PQ_m = P. The finite minimum defining δ\delta accounts both for ties and for the slack of the other constraints.

Lemma 3.9 (From feasible directions to monotonicity). Let F=min⁡1≤j≤nℓjF = \min_{1 \le j \le n} \ell_j, where n≥1n \ge1 and the ℓj\ell_j are affine. Suppose x:[a,b]→Rdx : [a,b] \to\mathbb{R}^d is continuous and differentiable on (a,b)(a,b). If, for each t∈(a,b)t \in(a,b), the direction x˙(t)\dot{x}(t) is feasible in {y:F(y)≥F(x(t))}\{y : F(y) \ge F(x(t))\} at x(t)x(t), then F∘xF \circ x is non-decreasing on [a,b][a,b].

Proof. The maximum norm gives

∣F(y)−F(z)∣≤C∣y−z∣,C=max⁡1≤j≤n∑i∣rj,i∣.\lvert F(y) - F(z) \rvert\le C\lvert y-z\rvert,\qquad C = \max_{1 \le j \le n} \sum_i \lvert r_{j,i}\rvert.

Indeed, use a minimizing branch at zz to bound F(y)−F(z)F(y)-F(z), and then exchange y,zy,z. Put g=F∘xg = F \circ x; it is continuous on [a,b][a,b], because xx is continuous and FF is Lipschitz. For fixed t∈(a,b)t \in(a,b), differentiability gives x(t+s)=x(t)+sx˙(t)+et(s)x(t+s) = x(t) + s\dot{x}(t) + e_t(s), with ∣et(s)∣/s→0\lvert e_t(s)\rvert/s \to0 as s↓0s \downarrow0. For every sufficiently small positive ss, feasibility and the Lipschitz estimate imply

g(t+s)≥F(x(t)+sx˙(t))−C∣et(s)∣≥g(t)−C∣et(s)∣.g(t+s) \ge F(x(t) + s\dot{x}(t)) - C\lvert e_t(s)\rvert\ge g(t) - C\lvert e_t(s)\rvert.

Consequently lim inf⁡s↓0(g(t+s)−g(t))/s≥0\liminf_{s \downarrow0}(g(t+s)-g(t))/s \ge0. The feasible step may vary with tt; only the limit at this fixed tt is used.

To pass from this local inequality to monotonicity, fix η>0\eta> 0 and put gη(t)=g(t)+ηtg_\eta(t) = g(t) + \eta t. If a<t1<t2<ba < t_1 < t_2 < b and gη(t1)>gη(t2)g_\eta(t_1) > g_\eta(t_2), the nonempty closed set {s∈[t1,t2]:gη(s)≥gη(t1)}\{s \in[t_1,t_2] : g_\eta(s) \ge g_\eta(t_1)\} has a largest point s∗<t2s^* < t_2. At s∗s^* the lower right Dini derivative is at least η>0\eta> 0, so gη(s)>gη(s∗)g_\eta(s) > g_\eta(s^*) for all sufficiently close s>s∗s > s^*. This contradicts maximality. Let η↓0\eta\downarrow0 and use continuity at the endpoints. □

Lemma 3.10 (Confinement). Let z∈Pz \in P and T≥0T \ge0, and let xx be a solution starting at zz on [0,T][0,T] in the sense of Definition 2.1(ii). Then x([0,T])⊂Px([0,T]) \subset P.

Proof. The case T=0T = 0 is immediate. For T>0T > 0, put τ=sup⁡{s∈[0,T]:x([0,s])⊂Bh}\tau= \sup\{s \in[0,T] : x([0,s]) \subset B_h\}. By z∈P⊂int⁡Bhz \in P \subset\operatorname{int} B_h and continuity, τ>0\tau> 0. Since BhB_h is closed, x([0,τ])⊂Bhx([0,\tau]) \subset B_h. For every t∈(0,τ)t \in(0,\tau), condition (B3) at x(t)x(t) says that x˙(t)=f(x(t))\dot{x}(t) = f(x(t)) is feasible in the current superlevel set. Lemma 3.9 therefore gives F(x(t))≥F(z)=MF(x(t)) \ge F(z) = M on [0,τ][0,\tau]. Together with F≤MF \le M, this gives x(t)∈Px(t) \in P throughout that interval. If τ<T\tau< T, then x(τ)∈P⊂int⁡Bhx(\tau) \in P \subset\operatorname{int} B_h and continuity keeps xx in BhB_h for a short interval after τ\tau, contradicting the definition of τ\tau. Hence τ=T\tau= T. □

Confinement holds for every solution, with no uniqueness, global existence, or positivity assumed in advance. Positivity follows from Bh⊂R>0dB_h \subset\mathbb{R}^d_{>0}.

Lemma 3.11 (Global existence from compact confinement). Let K⊂RdK \subset\mathbb{R}^d be compact. Suppose that, for every z∈Kz \in K and T≥0T \ge0, all solutions starting at zz on [0,T][0,T] remain in KK. Then a global forward solution exists from every z∈Kz \in K.

Proof. The polynomial map ff is bounded and Lipschitz on a compact neighborhood of KK. The Picard–Lindelöf theorem gives a time τ>0\tau> 0 independent of z∈Kz \in K such that a C1C^1 solution exists on [−τ,τ][-\tau,\tau] from every z∈Kz \in K. Choose such a solution x(0)x^{(0)} from zz. By confinement, x(0)(τ)∈Kx^{(0)}(\tau) \in K. Choose another local solution starting at x(0)(τ)x^{(0)}(\tau) and translate its forward segment to [τ,2τ][\tau,2\tau], extending the forward solution to [0,2τ][0,2\tau]. Continue inductively. Adjacent pieces agree in value at their junction, and both derivatives equal ff at that value, so the concatenation is differentiable there. This gives a curve on [−τ,∞)[-\tau,\infty) satisfying the equation at every t≥0t \ge0. Extend it arbitrarily to (−∞,−τ)(-\infty,-\tau). □

Proof of Theorem 3.1. Take the data of Proposition 3.6 and the plateau PP. Lemmas 3.10 and 3.11 give a global forward solution from x0∈Px_0 \in P, proving (a). For any global forward solution from x0x_0, apply Lemma 3.10 on each finite interval to obtain x(t)∈Px(t) \in P for all t≥0t \ge0. By Lemma 3.7, PP is nonempty, compact, and contained in the positive orthant. Each continuous function z↦∣ln⁡zi∣z \mapsto\lvert\ln z_i\rvert therefore attains a finite maximum on PP. Set

L=max⁡1≤i≤dmax⁡z∈P∣ln⁡zi∣≥0.L = \max_{1 \le i \le d}\max_{z \in P} \lvert\ln z_i\rvert\ge0.

The plateau was chosen from the network and x0x_0 before choosing a solution. Thus this one constant bounds every trajectory, proving (b).

Proof of Proposition 3.5

We first prove a sequential version on a general cube [a,b]m[a,b]^m (Lemma 3.12), and then obtain the uniform conclusion of Proposition 3.5 by compactness. We denote the dimension by mm because the proof proceeds by induction on dimension.

Notation. For an integer m≥1m \ge1, u∈Rmu \in\mathbb{R}^m, and 0<h<10 < h < 1, continue to write hu=(hu1,…,hum)h^u = (h^{u_1}, \ldots, h^{u_m}). A label is a pair λ=(rλ,cλ)\lambda= (r_\lambda,c_\lambda), where the slope rλ∈Rmr_\lambda\in\mathbb{R}^m is independent of hh and the intercept cλ(h)c_\lambda(h) is a real function defined for all sufficiently small h>0h > 0. Given a finite non-empty label family Λ\Lambda, a fixed small h>0h > 0, and x∈Rmx \in\mathbb{R}^m, we call λ∈Λ\lambda\in\Lambda active at xx if

rλ⋅x+cλ(h)=min⁡η∈Λ(rη⋅x+cη(h)).r_\lambda\cdot x + c_\lambda(h) = \min_{\eta\in\Lambda}\left(r_\eta\cdot x + c_\eta(h)\right).

All minimizers are active, including ties. When finitely many label families are combined, restrict their intercepts to a common domain 0<h<h∗0 < h < h_*. In every sequential statement below, discard finitely many initial terms so that all parameters lie in this common domain. Write μ(u)=min⁡iui\mu(u) = \min_i u_i and 1=(1,…,1)\mathbf{1} = (1,\ldots,1). The largest coordinate of huh^u is hμ(u)h^{\mu(u)}.

Lemma 3.12 (Affine approximation with arbitrary tolerance). Let m≥1m \ge1, a<ba < b, and let E:[a,b]m→(0,∞)E : [a,b]^m \to(0,\infty) be any function. There exists a finite non-empty label family Λ\Lambda with slopes rλ∈[a,b]mr_\lambda\in[a,b]^m and intercepts satisfying cλ(h)=o(ha)c_\lambda(h) = o(h^a) as h→0+h \to0^+, with the following property. For every hν→0+h_\nu\to0^+ and every sequence pν∈Rmp^\nu\in\mathbb{R}^m converging to p∈[a,b]mp \in[a,b]^m, if a fixed λ∈Λ\lambda\in\Lambda is active at hνpνh_\nu^{p^\nu} for every ν\nu, then

∣p−rλ∣<E(rλ).\lvert p-r_\lambda\rvert< E(r_\lambda).

No regularity is assumed for EE, and the approximating sequence pνp^\nu need not lie in [a,b]m[a,b]^m. This is necessary for induction: after projection, the sequence may approach a face of a smaller cube from either side. The conclusion holds for every label that remains active along the sequence, including tied labels.

Proof. We use strong induction on mm, invoking the induction hypothesis only in positive dimensions below mm. First fix a value tt of the minimum exponent μ(p)\mu(p) and construct a label family. Negative intercepts compensate for increasing slope components in coordinates where hph^p is small. Compactness extends the guarantee to nearby values of μ\mu around tt. A finite interval construction then selects among these families. When m=1m = 1, Stage 1 uses only a singleton family, so the remaining stages also establish the induction base.

Stage 1. Construction at a fixed minimum.

Fix t∈[a,b]t \in[a,b]. We construct a finite family Λt\Lambda_t whose slopes all satisfy μ(r)=t\mu(r) = t and which contains the label (t1,0)(t\mathbf{1},0). Its active labels satisfy the conclusion whenever the limit p∈[a,b]mp \in[a,b]^m has μ(p)=t\mu(p) = t and activity is taken within Λt\Lambda_t.

If m=1m = 1 or t=bt = b, take Λt={(t1,0)}\Lambda_t = \{(t\mathbf{1},0)\}. In either case, the only point of the cube with μ(p)=t\mu(p) = t is 1\mathbf{1}, so the conclusion is immediate. Henceforth assume m>1m > 1 and t<bt < b.

A label has type kk if exactly kk slope components are strictly greater than $t and all others equal tt. Type 0 consists only of (t1,0)(t\mathbf{1},0). Construct the remaining types in the order k=1,…,m−1k = 1,\ldots,m-1. After constructing type k−1k-1, choose βk\beta_k such that

t<βk<b,βk<βk−1 (k>1),βk−t<E(r)  for every type-(k−1) slope r.t < \beta_k < b,\qquad\beta_k < \beta_{k-1}\ (k > 1),\qquad\beta_k-t < E(r)\ \text{ for every type-}(k-1)\text{ slope }r.

There are only finitely many slopes of type k−1k-1, each with a positive tolerance, so such a βk\beta_k exists. Only these finitely many tolerance values are compared; no positive lower bound for EE on the entire cube is needed.

For each kk-element subset U⊂{1,…,m}U \subset\{1,\ldots,m\} and each σ∈[βk,b]U\sigma\in[\beta_k,b]^U, let r(σ)∈[a,b]mr(\sigma) \in[a,b]^m equal σ\sigma on UU and tt outside UU. Apply the induction hypothesis in dimension k<mk<m to the cube [βk,b]U[\beta_k,b]^U and tolerance σ↦E(r(σ))\sigma\mapsto E(r(\sigma)), ordering the coordinates of UU naturally. For every resulting label (σ,c′)(\sigma,c') with c′(h)=o(hβk)c'(h)=o\left(h^{\beta_k}\right), include the label

r=r(σ),c(h)=−hβk+c′(h)r=r(\sigma),\qquad c(h)=-h^{\beta_k}+c'(h)

in type kk. Taking all coordinate subsets and all types gives the finite non-empty family Λt\Lambda_t. Since βk>t\beta_k>t, every intercept is o(ht)o\left(h^t\right). The negative term −hβk-h^{\beta_k} favors increasing the slope in coordinates of xx that are small relative to hth^t. As βk\beta_k decreases, this preference becomes stronger at each successive type.

Stage 2. Verification at a fixed minimum. Assume m>1m>1, t<bt<b, hν→0+h_\nu\to0^+, pν→p∈[a,b]mp^\nu\to p\in[a,b]^m, and μ(p)=t\mu(p)=t. Fix a type kk label (r,c)(r,c) that remains active in Λt\Lambda_t along the sequence, and let UU be its raised coordinate set. All asymptotic comparisons below are along this sequence; write h=hνh=h_\nu. Whenever a limiting exponent lies strictly on one side of a threshold, convergence gives a fixed positive exponent gap for sufficiently large ν\nu. We use only these gaps and do not replace hpiνh^{p_i^\nu} by hpih^{p_i}, which convergence of exponents alone would not justify.

If k>0k>0, comparison with (t1,0)(t\mathbf{1},0) and activity give

∑i∈U(ri−t)hpiν≤hβk−c′(h)=(1+o(1))hβk.\sum_{i\in U}(r_i-t)h^{p_i^\nu}\le h^{\beta_k}-c'(h)=(1+o(1))h^{\beta_k}.

Every coefficient on the left is ≥βk−t>0\ge\beta_k-t>0. If pi<βkp_i<\beta_k for some i∈Ui\in U, a fixed δ>0\delta>0 satisfies piν≤βk−δp_i^\nu\le\beta_k-\delta for all sufficiently large ν\nu. Dividing by hβkh^{\beta_k} then gives a term tending to infinity on the left, a contradiction. Therefore

pi≥βk(i∈U).p_i\ge\beta_k\qquad(i\in U).

Among labels constructed using the same UU, the contribution t∑i∉Uxit\sum_{i\notin U}x_i from coordinates outside UU and the intercept term −hβk-h^{\beta_k} are common. Thus (σ,c′)(\sigma,c') is active at hpUνh^{p_U^\nu} within the family supplied by induction. Its limiting exponent vector pUp_U lies in [βk,b]U[\beta_k,b]^U. The induction hypothesis gives

max⁡i∈U∣pi−ri∣<E(r).\max_{i\in U}|p_i-r_i|<E(r).

Here it is essential to allow approximating sequences outside the cube: pνp^\nu need not lie in [βk,b]U[\beta_k,b]^U, although its limit does. In particular, a limiting coordinate equal to βk\beta_k or bb requires no strict inequality.

We next control coordinates outside UU, taking U=∅U=\varnothing for type 0. Suppose k<m−1k<m-1. We claim that

pi≤βk+1(i∉U).p_i\le\beta_{k+1}\qquad(i\notin U).

Otherwise there is some j∉Uj\notin U with pj>βk+1p_j>\beta_{k+1}, so every limiting exponent on U∪{j}U\cup\{j\} is strictly greater than βk+1\beta_{k+1}. When k>0k>0, use pi≥βk>βk+1p_i\ge\beta_k>\beta_{k+1} for i∈Ui\in U; when k=0k=0, no such check is needed. Choose any type k+1k+1 label whose raised coordinates are U∪{j}U\cup\{j\}. Its affine value minus the value of (t1,0)(t\mathbf{1},0) is

−hβk+1+o(hβk+1),-h^{\beta_{k+1}}+o\left(h^{\beta_{k+1}}\right),

since hpiν=o(hβk+1)h^{p_i^\nu}=o\left(h^{\beta_{k+1}}\right) in every raised coordinate. For the purported active label, the corresponding difference is zero when k=0k=0 and is bounded below by −hβk+o(hβk)-h^{\beta_k}+o\left(h^{\beta_k}\right) when k>0k>0. In the latter case, hβk=o(hβk+1)h^{\beta_k}=o\left(h^{\beta_{k+1}}\right). Thus, for sufficiently large ν\nu, the competitor has strictly smaller value, contradicting activity. This proves the claim. Outside UU we have pi≥tp_i \ge t and ri=tr_i = t, so

∣pi−ri∣≤βk+1−t<E(r)(i∉U),|p_i-r_i| \le\beta_{k+1}-t < E(r) \qquad(i \notin U),

where the strict inequality is precisely the condition imposed on βk+1\beta_{k+1} when constructing type k+1k+1.

If k=m−1k=m-1, only one coordinate lies outside UU. Each exponent on UU is ≥βk>t\ge\beta_k > t, and μ(p)=t\mu(p)=t, so the remaining limiting exponent equals tt and has zero error. Combining the cases gives ∣p−r∣≤E(r)|p-r| \le E(r) for every type, including types 0 and m−1m-1. Strictness is not needed at the threshold pi=βkp_i=\beta_k.

Stage 3. Extension to nearby minima.

Fix tt and the finite family Λt\Lambda_t. For each label λ=(r,c)\lambda=(r,c), let Aλ⊂[a,b]mA_\lambda\subset[a,b]^m be the set of limiting points realized by sequences hν→0+h_\nu\to0^+ and pν∈Rmp^\nu\in\mathbb{R}^m such that λ\lambda is active in Λt\Lambda_t at hνpνh_\nu^{p^\nu} and pνp^\nu converges to the given point. The set AλA_\lambda is closed. Indeed, if zn∈Aλz_n \in A_\lambda and zn→zz_n \to z, choose a point unu_n and a parameter hnh_n from a realizing sequence for znz_n such that

0<hn<1/n,∣un−zn∣<1/n,0<h_n<1/n,\qquad|u_n-z_n|<1/n,

and λ\lambda is active at hnunh_n^{u_n}. Then un→zu_n \to z, so z∈Aλz \in A_\lambda. This diagonal selection uses actual active points and does not require continuity of c(h)c(h).

Consequently, the set

Bt=⋃λ=(r,c)∈Λt{p∈Aλ:∣p−r∣≥E(r)}B_t=\bigcup_{\lambda=(r,c)\in\Lambda_t}\{p\in A_\lambda:|p-r|\ge E(r)\}

is compact. Every E(r)E(r) here is a fixed real number, so continuity of EE is unnecessary. Stages 1 and 2 show that no point of BtB_t satisfies μ(p)=t\mu(p)=t. Continuity of μ\mu therefore gives ρt>0\rho_t>0 such that no point of BtB_t satisfies

μ(p)∈It:=(t−ρt,t+ρt).\mu(p)\in I_t:=\left(t-\rho_t,t+\rho_t\right).

Indeed, if BtB_t is non-empty, choose ρt\rho_t below the positive minimum of ∣μ(p)−t∣|\mu(p)-t| on BtB_t. If BtB_t is empty, any positive radius works. Thus the limiting active-label conclusion holds whenever p∈[a,b]mp\in[a,b]^m and μ(p)∈It\mu(p)\in I_t.

If necessary, decrease ρt\rho_t further so that every local intercept satisfies

c(h)=o(ht+ρt).(7)c(h)=o\left(h^{t+\rho_t}\right). \tag*{(7)}

For a non-singleton family, it suffices to impose t+ρt<min⁡1≤k≤m−1βkt+\rho_t<\min_{1\le k\le m-1}\beta_k; singleton intercepts vanish. This stronger decay uses the constructed form c(h)=−hβk+o(hβk)c(h)=-h^{\beta_k}+o(h^{\beta_k}) and βk>t\beta_k>t and does not follow from c(h)=o(ht)c(h)=o(h^t) alone. Decreasing the radius preserves the preceding guarantee. We now have a correct local family near every possible value of μ(p)\mu(p); it remains to combine finitely many families so that the appropriate one is selected.

Stage 4. Finite interval construction and gluing.

Each local family gives the required approximation near its own minimum level. Add a common intercept shift to each family so that every label active in the combined family comes from an appropriate neighborhood. Overlap between adjacent neighborhoods allows equality at the transition cuts.

Choose a finite subcover of [a,b][a,b] from {It}t∈[a,b]\{I_t\}_{t\in[a,b]} and discard any selected interval contained in another, retaining one of any identical pair. Order the remaining centers as a≤t1<⋯<tN≤ba\le t_1<\cdots<t_N\le b and write Itj=(Lj,Rj)I_{t_j}=(L_j,R_j). Non-containment forces the left and right endpoints to have the same order. Since tj=(Lj+Rj)/2t_j=(L_j+R_j)/2, this is also the order of the centers. Hence

L1<⋯<LN,R1<⋯<RN.L_1<\cdots<L_N,\qquad R_1<\cdots<R_N.

Adjacent intervals overlap. Otherwise Rj≤Lj+1R_j \le L_{j+1}, leaving points in [Rj,Lj+1]⊂(tj,tj+1)[R_j,L_{j+1}] \subset(t_j,t_{j+1}) uncovered by every selected interval. The endpoint ordering, coverage, and the fact that all centers lie in [a,b][a,b] imply that the first interval contains aa and the last contains bb.

For 1≤j<N1 \le j < N, choose

qj∈(max⁡{tj,Lj+1},min⁡{tj+1,Rj}).q_j \in\left(\max\{t_j,L_{j+1}\},\min\{t_{j+1},R_j\}\right).

This open interval is non-empty: the adjacent intervals overlap, each contains its center, and tj<tj+1t_j < t_{j+1}. Thus qjq_j belongs to both neighboring intervals and lies strictly between their centers. Put q0=aq_0=a and qN=bq_N=b. Then

tj<qj<tj+1(1≤j<N),[qj−1,qj]⊂Itj(1≤j≤N).(8)t_j < q_j < t_{j+1}\quad(1 \le j < N), \qquad[q_{j-1},q_j] \subset I_{t_j}\quad(1 \le j \le N). \tag*{(8)}

When N=1N=1, no cut is needed and the single interval contains all of [a,b][a,b].

Take the union of the selected families and, at level tjt_j, replace each intercept c(h)c(h) by

c~(h)=c(h)−∑ℓ=1j−1hqℓ(9)\widetilde{c}(h)=c(h)-\sum_{\ell=1}^{j-1}h^{q_\ell} \tag*{(9)}

The sum is empty for j=1j=1. Every new intercept is o(ha)o(h^a): use (3.4) and tj+ρtj>at_j+\rho_{t_j}>a, while each shift exponent satisfies qℓ>tℓ≥aq_\ell>t_\ell\ge a. Within one local family, a common shift does not change activity. Between adjacent families, the extra term −hqj-h^{q_j} selects the family on the appropriate side of the cut.

Suppose a fixed label (r,c)(r,c) from level tjt_j remains active in the combined family along hν→0+h_\nu\to0^+ and pν→p∈[a,b]mp^\nu\to p \in[a,b]^m. Let cc be its original local intercept and continue to write h=hνh=h_\nu. We claim that

qj−1≤μ(p)≤qj.q_{j-1}\le\mu(p)\le q_j.

If j<Nj<N and μ(p)>qj\mu(p)>q_j, each hpiνh^{p_i^\nu} is o(hqj)o(h^{q_j}). By (3.4) and qj∈Itjq_j\in I_{t_j}, the local intercept is also o(hqj)o(h^{q_j}). Subtracting the value of the shifted label (tj+11,0)(t_{j+1}\mathbf{1},0) gives

(r−tj+11)⋅hpν+c(h)+hqj=hqj+o(hqj)>0(r-t_{j+1}\mathbf{1})\cdot h^{p^\nu}+c(h)+h^{q_j}=h^{q_j}+o(h^{q_j})>0

for sufficiently large ν\nu, contradicting activity. If j>1j>1 and μ(p)<qj−1\mu(p)<q_{j-1}, subtract instead the shifted label (tj−11,0)(t_{j-1}\mathbf{1},0). The difference is

(r−tj−11)⋅hpν+c(h)−hqj−1.(r-t_{j-1}\mathbf{1})\cdot h^{p^\nu}+c(h)-h^{q_{j-1}}.

Every coefficient in the inner product is ≥tj−tj−1>0\ge t_j-t_{j-1}>0. There is a coordinate ii with pi<qj−1p_i<q_{j-1}, so hpiν/hqj−1→∞h^{p_i^\nu}/h^{q_{j-1}}\to\infty. Also qj−1<tj<tj+ρtjq_{j-1}<t_j<t_j+\rho_{t_j} gives c(h)=o(hqj−1)c(h)=o(h^{q_{j-1}}). Dividing the difference by hqj−1h^{q_{j-1}} therefore gives a quantity tending to +∞+\infty, again contradicting activity. At j=1j=1 or j=Nj=N, the outer bounds follow from μ(p)∈[a,b]\mu(p)\in[a,b].

By (3.5), the claim implies μ(p)∈Itj\mu(p)\in I_{t_j}, including equality at either cut. Labels at the same level receive the same shift, so activity in the combined family implies activity in the original local family. The Stage 3 guarantee for ItjI_{t_j} gives ∣p−r∣<E(r)|p-r|<E(r). Each contradiction uses a competitor with strictly smaller value, so the argument also applies to tied active labels. This completes the induction. □

Corollary 3.13 (Uniform active-label bound). Let m,a,b,Em,a,b,E be as in Lemma 3.12 and fix its label family Λ\Lambda. There exists h0>0h_0>0 such that, for every 0<h<h00<h<h_0 and every p∈[a,b]mp\in[a,b]^m, each label active at hph^p satisfies ∣p−r∣<E(r)|p-r|<E(r). The family Λ\Lambda is fixed before h0h_0 is chosen.

Proof. Otherwise, for every ν\nu there exist 0<hν<1/ν0<h_\nu<1/\nu, a point pν∈[a,b]mp^\nu\in[a,b]^m, and a label λν\lambda_\nu active at hpνh^{p^\nu} such that ∣pν−rλν∣≥E(rλν)|p^\nu-r_{\lambda_\nu}|\ge E(r_{\lambda_\nu}). Since Λ\Lambda is finite and [a,b]m[a,b]^m is compact, a subsequence has λν=λ\lambda_\nu=\lambda fixed and pν→p∈[a,b]mp^\nu\to p\in[a,b]^m. On this subsequence E(rλ)E(r_\lambda) is constant. Passing to the limit gives ∣p−rλ∣≥E(rλ)|p-r\lambda|\ge E(r\lambda), contrary to the strict inequality in Lemma 3.12. No regularity of EE is used. □

Proof of Proposition 3.5. Take m=dm=d, a=−1a=-1, and b=1b=1. Apply Lemma 3.12 and Corollary 3.13 to the restriction of EE to [−1,1]d[-1,1]^d, obtaining a finite non-empty family Λ\Lambda and h0>0h_0>0. Decrease h0h_0 if needed so that all intercepts are defined on (0,h0)(0,h_0), and extend them arbitrarily to the remaining values of hh, for example by 00; these values do not affect the conclusion. Enumerate Λ\Lambda as {(rj,cj)}j=1n\{(r_j,c_j)\}_{j=1}^n with n≥1n\ge1. The condition ∣p∣≤1|p|\le1 is equivalent to p∈[−1,1]dp\in[-1,1]^d, and the assertion that jj is active at hph^p for {rk⋅x+ck(h)}k\{r_k\cdot x+c_k(h)\}_k is exactly the activity defined above. Corollary 3.13 therefore gives ∣pi−rj,i∣<E(rj)|p_i-r_{j,i}|<E(r_j) for i=1,…,di=1,\ldots,d. □

Proposition 3.5 enters Step 1 in two ways. The directions rjr_j are fixed before hh, so the gap HH in Lemma 3.4(b) is independent of hh. The conclusion holds for every active index, so every tight constraint passes the feasible-direction test in Lemma 3.8. The tolerance EE may be discontinuous, which allows Proposition 3.6 to impose sign conditions near finitely many hyperplanes simultaneously.

A common compact container in the positive orthant

We return to detailed balanced systems and fix x0∈R>0dx_0\in\mathbb{R}^d_{>0}.

Lemma 3.14 (Invariance of the compatibility class). Let xx be a global forward solution starting at x0x_0. Then x(t)−x0∈Sx(t)-x_0\in S for every t≥0t\ge0.

Proof. Let φ\varphi be any linear functional satisfying φ∣S=0\varphi|_S=0. Since f(x)∈Sf(x)\in S, we have ddsφ(x(s))=φ(f(x(s)))=0\frac{d}{ds}\varphi(x(s))=\varphi(f(x(s)))=0 for s≥0s\ge0. The mean value theorem gives φ(x(t))=φ(x0)\varphi(x(t))=\varphi(x_0), that is, φ(x(t)−x0)=0\varphi(x(t)-x_0)=0. In finite dimensions, a vector annihilated by every functional in the annihilator of SS belongs to SS. □

This proof uses the mean value theorem and does not require integration.

Corollary 3.15 (Step 1). Assume detailed balance and x0∈R>0dx_0\in\mathbb{R}^d_{>0}. There exists L≥0L\ge0 such that the set

K=[e−L,eL]d∩(x0+S)K=[e^{-L},e^L]^d\cap(x_0+S)

has the following properties. (a) A global forward solution exists from x0x_0. (b) Every global forward solution from x0x_0 satisfies x(t)∈Kx(t)\in K for t≥0t\ge0. (c) The set KK is compact and K⊂R>0dK\subset\mathbb{R}^d_{>0}. (d) There exists B>0B>0 such that ∣f(x)∣≤B|f(x)|\le B for every x∈Kx\in K.

Proof. Lemma 2.3 gives weak reversibility, so Theorem 3.1 supplies LL and (a). Exponentiating (4) gives e−L≤xi(t)≤eLe^{-L}\le x_i(t)\le e^L for every solution and nonnegative time. Lemma 3.14 places these states in x0+Sx_0+S, proving (b). The box is compact and x0+Sx_0+S is closed, so their intersection is compact; every coordinate on it is at least e−L>0e^{-L}>0. This proves (c). Continuity of ff on KK gives a finite speed bound; increasing it to a positive number proves (d). □

The detailed-balanced convergence argument uses Theorem 3.1 only through Corollary 3.15. For the fixed initial state x0x_0, the set KK is a common compact container for all its forward solutions. The invariant plateau PP constructed earlier provides the stronger trapping property. Compact containment is sufficient for the subsequent convergence proof.

The Horn Jackson argument

Throughout this subsection and Section 3.3, assume detailed balance, with x∗x^* and KeK_e as in (2). For x∈R>0dx\in\mathbb{R}^d_{>0} and y∈Ndy\in\mathbb{N}^d, define

Ly(x)=∑i=1dyi(ln⁡xi−ln⁡xi∗)=ln⁡xy(x∗)yL_y(x)=\sum_{i=1}^{d}y_i(\ln x_i-\ln x_i^*)=\ln\frac{x^y}{(x^*)^y}

Then xy=(x∗)yeLy(x)x^y=(x^*)^y e^{L_y(x)}.

Entropy monotonicity

Define

V(x)=∑i=1d(xi(ln⁡xi−ln⁡xi∗)−xi+xi∗),D(x)=∑i=1d(ln⁡xi−ln⁡xi∗)fi(x)(x∈R>0d).(10)V(x) = \sum_{i=1}^{d} (x_i(\ln x_i - \ln x_i^*) - x_i + x_i^*), \qquad D(x) = \sum_{i=1}^{d}(\ln x_i - \ln x_i^*)f_i(x) \qquad(x \in\mathbb{R}_{>0}^{d}). \tag*{(10)}

The quantity D=∇V⋅fD = \nabla V \cdot f is the directional derivative of VV along the vector field.

Lemma 3.16 (Entropy properties). (a) We have V≥0V \ge0 on R>0d\mathbb{R}_{>0}^{d}. (b) If xx is differentiable at tt with derivative f(x(t))f(x(t)) and x(t)∈R>0dx(t) \in\mathbb{R}_{>0}^{d}, then s↦V(x(s))s \mapsto V(x(s)) is differentiable at tt with derivative D(x(t))D(x(t)). (c) The function DD is continuous on R>0d\mathbb{R}_{>0}^{d}.

Proof. (a) For a,b>0a,b>0, the inequality ln⁡u≤u−1\ln u \le u-1 with u=b/au=b/a gives ln⁡b−ln⁡a≤b/a−1\ln b-\ln a \le b/a-1. Multiplication by aa gives a(ln⁡a−ln⁡b)−a+b≥0a(\ln a-\ln b)-a+b \ge0; sum componentwise. (b) Apply the chain rule componentwise: ddt[xi(ln⁡xi−ln⁡xi∗)−xi]=x˙i(ln⁡xi−ln⁡xi∗)+x˙i−x˙i\frac{d}{dt}[x_i(\ln x_i-\ln x_i^*)-x_i]=\dot{x}_i(\ln x_i-\ln x_i^*)+\dot{x}_i-\dot{x}_i. (c) The function ln⁡\ln is continuous on the positive reals, and ff is polynomial. □

Lemma 3.17 (Dissipation identity). For x∈R>0dx \in\mathbb{R}_{>0}^{d}, we have

2D(x)=∑e∈RKe(Lt(e)−Ls(e))(eLs(e)−eLt(e))≤0,(11)2D(x)=\sum_{e\in\mathcal{R}}K_e\left(L_{t(e)}-L_{s(e)}\right)\left(e^{L_{s(e)}}-e^{L_{t(e)}}\right)\le0, \tag*{(11)}

where Ly=Ly(x)L_y=L_y(x). Every summand is ≤0\le0 and equals 00 if and only if Ls(e)=Lt(e)L_{s(e)}=L_{t(e)}.

Proof. Interchanging the sums gives D(x)=∑eκexs(e)∑i(t(e)i−s(e)i)(ln⁡xi−ln⁡xi∗)=∑eκexs(e)(Lt(e)−Ls(e))D(x)=\sum_e\kappa_e x^{s(e)}\sum_i(t(e)_i-s(e)_i)(\ln x_i-\ln x_i^*)=\sum_e\kappa_e x^{s(e)}\left(L_{t(e)}-L_{s(e)}\right). By (2),

κexs(e)=KeeLs(e),κeˉxt(e)=κeˉ(x∗)t(e)eLt(e)=KeeLt(e).\kappa_e x^{s(e)}=K_e e^{L_{s(e)}}, \qquad\kappa_{\bar e}x^{t(e)}=\kappa_{\bar e}(x^*)^{t(e)}e^{L_{t(e)}}=K_e e^{L_{t(e)}}.

The map e↦eˉe\mapsto\bar e is an involution on R\mathcal{R}, so ∑eG(e)=∑eG(eˉ)\sum_eG(e)=\sum_eG(\bar e) for any GG. Hence 2D=∑e(G(e)+G(eˉ))2D=\sum_e(G(e)+G(\bar e)), where G(e)=κexs(e)(Lt(e)−Ls(e))G(e)=\kappa_e x^{s(e)}\left(L_{t(e)}-L_{s(e)}\right). Substitution gives

G(e)+G(eˉ)=KeeLs(e)(Lt(e)−Ls(e))+KeeLt(e)(Ls(e)−Lt(e)).G(e)+G(\bar e)=K_e e^{L_{s(e)}}\left(L_{t(e)}-L_{s(e)}\right)+K_e e^{L_{t(e)}}\left(L_{s(e)}-L_{t(e)}\right).

which is (3.8). The function exp⁡\exp is strictly increasing, so (b−a)(ea−eb)≤0(b-a)(e^a-e^b)\le0, with equality if and only if a=ba=b. Also Ke>0K_e>0. □

Lemmas 3.16 and 3.17 show that, along every positive solution, t↦V(x(t))t\mapsto V(x(t)) is differentiable with derivative D(x(t))≤0D(x(t))\le0. It is therefore non-increasing and bounded below by 0.

Zero dissipation and positive equilibria

Lemma 3.18 (Exponential parameterization of positive equilibria). Assume positive rates and detailed balance at x∗x^*. Define Φ(u)=(xi∗eui)i=1d\Phi(u)=(x_i^*e^{u_i})_{i=1}^d. Then

{x∈R>0d:f(x)=0}={x∈R>0d:D(x)=0}=Φ(S⊥).\{x\in\mathbb{R}_{>0}^{d}:f(x)=0\}=\{x\in\mathbb{R}_{>0}^{d}:D(x)=0\}=\Phi(S^\perp).

The map Φ:S⊥→{x>0:f(x)=0}\Phi:S^\perp\to\{x>0:f(x)=0\} is bijective, with inverse x↦ln⁡x−ln⁡x∗x\mapsto\ln x-\ln x^*, where logarithms are componentwise.

Proof. First let u∈S⊥u\in S^\perp and x=Φ(u)x=\Phi(u). Since every reaction vector belongs to SS,

Lt(e)(x)−Ls(e)(x)=u⋅(t(e)−s(e))=0.L_{t(e)}(x)-L_{s(e)}(x)=u\cdot(t(e)-s(e))=0.

The flux identities of Lemma 3.17 give κexs(e)=KeeLs(e)=KeeLt(e)=κeˉxt(e)\kappa_e x^{s(e)}=K_e e^{L_{s(e)}}=K_e e^{L_{t(e)}}=\kappa_{\bar e}x^{t(e)}. Reindexing (1) by the reverse-reaction involution therefore gives

2f(x)=∑e(κexs(e)−κeˉxt(e))(t(e)−s(e))=0.2f(x)=\sum_e\left(\kappa_e x^{s(e)}-\kappa_{\bar e}x^{t(e)}\right)(t(e)-s(e))=0.

Thus Φ(S⊥)\Phi(S^\perp) consists of positive equilibria, and every positive equilibrium satisfies D(x)=∇V(x)⋅f(x)=0D(x)=\nabla V(x)\cdot f(x)=0.

Conversely, let x>0x>0 satisfy D(x)=0D(x)=0, and put u=ln⁡x−ln⁡x∗u=\ln x-\ln x^*. In (3.8) every summand is nonpositive. Since Ke>0K_e>0, a nonzero logarithmic gap would make the sum strictly negative. Hence u⋅(t(e)−s(e))=Lt(e)−Ls(e)=0u\cdot(t(e)-s(e))=L_{t(e)}-L_{s(e)}=0 for every reaction. The reaction vectors span SS, so u∈S⊥u\in S^\perp, and exponentiation gives x=Φ(u)x=\Phi(u). This proves equality of the three sets. Finally, x=Φ(v)x=\Phi(v) forces vi=ln⁡xi−ln⁡xi∗v_i=\ln x_i-\ln x_i^* for every ii, proving uniqueness of the parameter. □

The parameterization describes all positive equilibria, including those outside the compatibility class of x∗x^*. If there are no reactions, S={0}S=\{0\} and Φ(Rd)=R>0d\Phi(\mathbb{R}^d)=\mathbb{R}^d_{>0}, as required.

Uniqueness in each compatibility class

Lemma 3.19 (Uniqueness in each compatibility class). If a,b∈R>0da,b\in\mathbb{R}^d_{>0}, f(a)=f(b)=0f(a)=f(b)=0, and a−b∈Sa-b\in S, then a=ba=b.

Proof. By Lemma 3.18, write a=Φ(u)a=\Phi(u) and b=Φ(v)b=\Phi(v) with u,v∈S⊥u,v\in S^\perp. Since u−v∈S⊥u-v\in S^\perp and a−b∈Sa-b\in S,

0=(u−v)⋅(a−b)=∑ixi∗(ui−vi)(eui−evi).0=(u-v)\cdot(a-b)=\sum_i x_i^*(u_i-v_i)(e^{u_i}-e^{v_i}).

Every summand is nonnegative, and strict monotonicity of exp⁡\exp, together with xi∗>0x_i^*>0, makes it zero only when ui=viu_i=v_i. Therefore u=vu=v and a=ba=b. □

Cluster points and convergence

A point zz is a cluster point of x(t)x(t) as t→∞t\to\infty if there exists tn→∞t_n\to\infty such that x(tn)→zx(t_n)\to z.

Lemma 3.20 (Descent argument). Let xx be a global forward solution with x(t)∈Kx(t)\in K for all t≥0t\geq0, where K⊂R>0dK\subset\mathbb{R}^d_{>0} is compact. Fix B>0B>0 such that ∣f(w)∣≤B\lvert f(w)\rvert\leq B on KK. If z∈Kz\in K is a cluster point, then D(z)=0D(z)=0.

Proof. Put W(t)=V(x(t))W(t)=V(x(t)). Step 2 and the mean value theorem show that WW is non-increasing on [0,∞)[0,\infty) and satisfies W≥0W\geq0. Let c=inf⁡t≥0W(t)c=\inf_{t\geq0}W(t).

Suppose D(z)≠0D(z)\ne0. Lemma 3.17 gives D(z)<0D(z)<0. Put η=−D(z)/2>0\eta=-D(z)/2>0. Continuity of DD at zz gives r>0r>0 such that ∣w−z∣<r\lvert w-z\rvert<r implies D(w)<−ηD(w)<-\eta; decrease the radius so that this neighborhood lies in the positive orthant. Put δ=r/(2B)\delta=r/(2B). Choose T≥0T\geq0 with W(T)<c+ηδW(T)<c+\eta\delta, and then choose t≥Tt\geq T with ∣x(t)−z∣<r/2\lvert x(t)-z\rvert<r/2, using that zz is a cluster point. For s∈[t,t+δ]s\in[t,t+\delta], the mean value inequality gives ∣x(s)−x(t)∣≤B(s−t)≤Bδ=r/2\lvert x(s)-x(t)\rvert\leq B(s-t)\leq B\delta=r/2, hence ∣x(s)−z∣<r\lvert x(s)-z\rvert<r and ddsW(s)=D(x(s))<−η\frac{d}{ds}W(s)=D(x(s))<-\eta. A further application of the mean value theorem gives

W(t+δ)≤W(t)−ηδ≤W(T)−ηδ<c,W(t+\delta)\leq W(t)-\eta\delta\leq W(T)-\eta\delta<c,

contradicting the definition of cc as the infimum. □

A standard alternative uses LaSalle’s invariance principle: the ω\omega-limit set is forward invariant, VV is constant on it, and therefore D=0D=0. That formulation uses continuous dependence on initial conditions. Lemma 3.20 uses only a speed bound.

Completion of the proof

Theorem 3.21 (Main theorem). Theorem 1.1 holds. In particular, the equilibrium xˉ\bar{x} in part (i) has the property that every zz in (x0+S)∩R>0d(x_0+S)\cap\mathbb{R}^d_{>0} satisfying f(z)=0f(z)=0 equals xˉ\bar{x}.

Proof. Take L,K,BL,K,B from Corollary 3.15 and choose a global forward solution x(1)x^{(1)}. Since it remains in the compact set KK, it has a cluster point xˉ∈K\bar{x}\in K. Lemma 3.20 gives D(xˉ)=0D(\bar{x})=0, so Lemma 3.18 places xˉ\bar{x} in Φ(S⊥)\Phi(S^\perp), the set of positive equilibria. Also xˉ∈K⊂(x0+S)∩R>0d\bar{x}\in K\subset(x_0+S)\cap\mathbb{R}^d_{>0}. If zz is another equilibrium in this class, then z−xˉ∈Sz-\bar{x}\in S, and Lemma 3.19 gives z=xˉz=\bar{x}. This proves (i); part (ii) is Corollary 3.15(a).

For (iii), let xx be any global forward solution starting at x0x_0. It remains in KK. Since KK is closed, every cluster point lies in KK, so Lemma 3.20 applies. The same argument shows that every cluster point is an equilibrium in (x0+S)∩R>0d(x_0+S)\cap\mathbb{R}^d_{>0} and therefore equals xˉ\bar{x}. A trajectory in a compact set with a unique cluster point converges to that point. Hence x(t)→xˉx(t)\to\bar{x}. □

Remark 3.22. (1) The proof also yields existence of a positive equilibrium in every positive class. This classical conclusion can alternatively be obtained by Birch-type convex analysis [7, 11]. (2) Theorem 3.1 assumes d≥1d\ge1. For d=0d=0, the space R0\mathbb{R}^0 has one point, so the conclusion of the main theorem is immediate.

Corollary 3.23 (Asymptotic stability within the class). Put Π=(xˉ+S)∩R>0d\Pi=(\bar{x}+S)\cap\mathbb{R}^d_{>0}. Then xˉ\bar{x} is the strict global minimizer of V∣ΠV|_{\Pi} and is Lyapunov stable in Π\Pi. Together with Theorem 3.21, this implies that xˉ\bar{x} is globally asymptotically stable in Π\Pi.

Proof. If d=0d=0, the class is a singleton and both conclusions are immediate. Assume d≥1d\ge1, and use the maximum norm throughout this proof. The function VV is strictly convex on R>0d\mathbb{R}^d_{>0}, with Hessian diag⁡(1/xi)\operatorname{diag}(1/x_i). Lemma 3.18 gives ∇V(xˉ)=ln⁡xˉ−ln⁡x∗∈S⊥\nabla V(\bar{x})=\ln\bar{x}-\ln x^*\in S^\perp. Thus, for x∈Π∖{xˉ}x\in\Pi\setminus\{\bar{x}\}, strict convexity gives V(x)>V(xˉ)+∇V(xˉ)⋅(x−xˉ)=V(xˉ)V(x)>V(\bar{x})+\nabla V(\bar{x})\cdot(x-\bar{x})=V(\bar{x}). For stability, fix any η>0\eta>0 and choose 0<ρ<min⁡{η,min⁡ixˉi}0<\rho<\min\{\eta,\min_i\bar{x}_i\}. The closed ball B‾(xˉ,ρ)\overline{B}(\bar{x},\rho) lies in the positive orthant. Its sphere intersected with Π\Pi is compact, since it equals the intersection of that compact sphere with the closed affine space xˉ+S\bar{x}+S. If this intersection is empty, take δ=ρ/2\delta=\rho/2; continuity prevents a trajectory in Π\Pi starting within distance δ\delta from exiting this ball. Otherwise put m=min⁡{V(x)−V(xˉ):x∈Π, ∥x−xˉ∥=ρ}>0m=\min\{V(x)-V(\bar{x}):x\in\Pi,\ \lVert x-\bar{x}\rVert=\rho\}>0. This minimum exists and is positive by compactness and strict minimality. By continuity of VV, choose 0<δ<ρ0<\delta<\rho such that ∥x−xˉ∥<δ\lVert x-\bar{x}\rVert<\delta implies V(x)−V(xˉ)<mV(x)-V(\bar{x})<m. For any such initial state in Π\Pi, Theorem 3.21 gives a global forward solution. Corollary 3.15(b), applied to this initial state, places every global forward solution xx in a compact subset of the positive orthant, so V∘xV\circ x is non-increasing, and Lemma 3.14 keeps xx in Π\Pi. If ∥x(t)−xˉ∥≥ρ\lVert x(t)-\bar{x}\rVert\ge\rho for some t≥0t\ge0, then ∥x(0)−xˉ∥<δ<ρ\lVert x(0)-\bar{x}\rVert<\delta<\rho and continuity give t1>0t_1>0 with ∥x(t1)−xˉ∥=ρ\lVert x(t_1)-\bar{x}\rVert=\rho and x(t1)∈Πx(t_1)\in\Pi. Hence V(x(t1))−V(xˉ)≥m>V(x(0))−V(xˉ)V(x(t_1))-V(\bar{x})\ge m>V(x(0))-V(\bar{x}), contradicting monotonicity. Consequently every such solution satisfies ∥x(t)−xˉ∥<ρ<η\lVert x(t)-\bar{x}\rVert<\rho<\eta for all t≥0t\ge0. This proves Lyapunov stability with the required quantifiers. □

Lean Formalization

Scope

Theorem 1.1 is formalized in Lean 4 with Mathlib, together with every numbered step of Section 3 from Theorem 3.1 to Theorem 3.21 and the exact example of Section 5.1. The development has about 9,300 lines. About 7,900 of them form the persistence part, adapted from OpenAI’s formalization [15], which supplies Theorem 3.1; the rest was written for this paper and contains the entropy argument, the convergence proof, and the examples. Section 4.2 gives the formal main theorem and the definitions it depends on, Section 4.3 the correspondence with the paper and the points where the formal proof departs from the text, and Section 4.4 the verification and the build.

Formal statement of the main theorem

structure DBCondition (N : CRN d)
    (κ : RxnIdx N → ℝ) (c : Fin d → ℝ) : Prop where
  hRev : HasReverses N
  hPos : InOpenOrthant c
  hEq : ∀ e : RxnIdx N,
    κ e * xPow c e.val.1 = κ (oppRxn hRev e) * xPow c e.val.2

theorem db_unique_equilibrium_attracts
(N:CRN d) (κ:RxnIdx N→R) (hκ:∀e, 0<κ e){c:Fin d→R} (hDB:DBCondition N κ c)(x0:Fin d→R) (hx0:InOpenOrthant x0):∃c⋆:Fin d→R,InOpenOrthant c⋆∧rateField N κ c⋆=0∧c⋆−x0∈reactionSpan N∧(∀z, InOpenOrthant z→rateField N κ z=0→z−x0∈reactionSpan N→z=c⋆)∧(∃x, SolvesForward N κ x0 x)∧∀x, SolvesForward N κ x0 x→Tendsto x atTop (N c⋆)\begin{gathered} (N : \mathrm{CRN}\ d)\ (\kappa: \mathrm{RxnIdx}\ N \to\mathbb{R})\ (h_\kappa: \forall e,\ 0 < \kappa\ e)\\ \{c : \mathrm{Fin}\ d \to\mathbb{R}\}\ (h_{\mathrm{DB}} : \mathrm{DBCondition}\ N\ \kappa\ c)\\ (x_0 : \mathrm{Fin}\ d \to\mathbb{R})\ (h_{x_0} : \mathrm{InOpenOrthant}\ x_0) :\\ \exists c^\star: \mathrm{Fin}\ d \to\mathbb{R},\\ \quad\mathrm{InOpenOrthant}\ c^\star\land\mathrm{rateField}\ N\ \kappa\ c^\star= 0 \land\\ \quad c^\star- x_0 \in\mathrm{reactionSpan}\ N \land\\ \quad(\forall z,\ \mathrm{InOpenOrthant}\ z \to\mathrm{rateField}\ N\ \kappa\ z = 0 \to\\ \qquad z - x_0 \in\mathrm{reactionSpan}\ N \to z = c^\star) \land\\ \quad(\exists x,\ \mathrm{SolvesForward}\ N\ \kappa\ x_0\ x) \land\\ \quad\forall x,\ \mathrm{SolvesForward}\ N\ \kappa\ x_0\ x \to\\ \qquad\mathrm{Tendsto}\ x\ \mathrm{atTop}\ (N\ c^\star) \end{gathered}

The statement is taken from GAC/Paper.lean, with the proof omitted. The Lean kernel checks that the proof establishes exactly this statement. Whether the statement expresses Theorem 1.1 then depends only on the definitions it uses; together with the statement they occupy fewer than sixty lines on top of Mathlib. The names in the listing are transparent abbreviations of these definitions, which read as follows.

  • N : CRN d is a network as in § 2.1: a finite set of complexes in Nd\mathbb{N}^d and a finite set of reactions, each an ordered pair of distinct complexes of the network. RxnIdx N is the type of its reactions, so κ\kappa assigns a rate constant to each reaction, and hκh_\kappa requires all of them to be positive.

  • xPow c y is the monomial cy=∏iciyic^y = \prod_i c_i^{y_i}, and rateField N κ is the vector field ff of (1), with the reaction vectors y′−yy' - y computed in Rd\mathbb{R}^d.

  • InOpenOrthant x means x∈R>0dx \in\mathbb{R}_{>0}^d, and reactionSpan N is the stoichiometric subspace SS.

  • HasReverses N is reversibility (Definition 2.2(i)), oppRxn maps ee to eˉ\bar{e}, and DBCondition N κ c is detailed balance (2), with cc in the role of x∗x^*.

  • SolvesForward N κ x0 x is Definition 2.1(i): x:R→Rdx : \mathbb{R} \to\mathbb{R}^d, x(0)=x0x(0) = x_0, and xx has derivative f(x(t))f(x(t)) at every t≥0t \ge0. It assumes neither positivity nor convergence.

With these readings, the first three conjuncts of the conclusion say that c∗c^* is a positive equilibrium in x0+Sx_0 + S, and the fourth that it is the only one; this is Theorem 1.1(i). The last two conjuncts are Theorem 1.1(ii) and (iii). The final conjunct quantifies over all global forward solutions, so it does not rely on uniqueness of solutions.

Two instances check that the hypotheses can be met and that the definitions take the expected values. For the network of § 5.1, the development proves detailed balance at the stated rates, the conservation law xA+2xB+3xCx_A + 2x_B + 3x_C, and convergence to (12,12,12)\left(\frac{1}{2}, \frac{1}{2}, \frac{1}{2}\right) from every positive initial state in the class xA+2xB+3xC=3x_A + 2x_B + 3x_C = 3, including the five initial states of Figure 1 written as exact rationals. A second instance treats A⇌BA \rightleftharpoons B.

Trajectories and entropy along trajectories

Figure 1. Trajectories for A+B⇌CA+B\rightleftharpoons C and 2A⇌B2A\rightleftharpoons B in the class xA+2xB+3xC=3x_A+2x_B+3x_C=3. (a) Projection onto the (xA,xB)(x_A,x_B) plane; dots mark initial states and the star marks xˉ=(12,12,12)\bar{x}=\left(\frac{1}{2},\frac{1}{2},\frac{1}{2}\right). (b) A logarithmic plot of V(x(t))V(x(t)); the entropy panel is plotted down to 10−1210^{-12}.

Correspondence with the paper

Table 1 lists, for each numbered result, the Lean declarations that state it; names are relative to the namespaces GAC.Paper, GAC.Persistence, and GAC.Reformulated, and to GAC.PaperExample for the example rows. File locations, and the declarations behind individual proof steps, are recorded in docs/PAPER_MAPPING.md in the repository. The formal proof follows Section 3, with the following departures.

PaperLean declarations
1.1 / 3.21db_unique_equilibrium_attracts
2.3strongConn_of_hasReverses
3.1wr_trajectory_log_bound<br>bounded_persistent_of_weaklyReversible
3.1 / 3.15compact_positive_iff_log_bound<br>solution_family_log_confinement
3.2Cuts.cut_flux_nonneg_of_ordered_sources<br>massAction_points_inward_of_order
3.3monomial_bound_of_exponent_gap
3.4exists_sign_margin_function<br>exists_uniform_sign_gap
3.5uniform_affine_approximation<br>Affine.uniform_data_of_full_certificates
3.6exists_minBarrierSpec<br>Barrier.exists_barrier_data
3.7Barrier.BarrierData.plateau_subset_interior_box<br>.isCompact_plateau
3.8feasible_direction_iff_tight
3.6 / 3.8Plateau special cases:<br>plateau_active_iff_tight<br>plateau_feasible_direction_iff_active<br>plateau_massAction_feasible
3.9–3.10finite_minimum_monotone_of_active_slopes<br>Confinement.forward_solution_trapped
3.11global_solution_of_compact_trap<br>global_massAction_of_compact_trap
3.12Affine.exists_full_certificate
3.13Affine.FullCertificate.uniform_activity_bound
3.14trajectory_in_coset
3.15GAC.exists_compact_class_container
3.16entropyFn_ge_zero<br>entropyFn_deriv<br>entropyRate_continuous
PaperLean declarations
3.17GAC.dissip_eq_sum
entropyRate_double
entropyRate_le_zero
3.18logGap_vanish_of_rate_zero
rateField_zero_of_logGap
entropyRate_eq_zero_iff
positive_equilibria_eq_exp_normal
exponential_parameter_unique
3.19equilibria_unique_of_exp_parameterization
3.20entropyRate_zero_at_accumulation
5.1net_detailed_balanced
total_conserved
tends_to_equilibrium_of_total_eq_three
Figure 1figure_initial_states_positive
figure_initial_states_total
figure_initial_states_converge

Table 1. Numbered results and the Lean declarations that state them.

Table 1.

  • Lemmas 3.9–3.10. Lean derives confinement from the monotonicity of a finite minimum along curves on which every active branch has nonnegative slope. The Lipschitz estimate and the Dini argument of Lemma 3.9 have no separate Lean counterpart; the conclusion is the same.

  • Proposition 3.6 (B3). Lean proves that every active branch has nonnegative slope along ff at each point of BhB_h. The feasible-segment form follows from Lemma 3.8, as in the text; Lean states it on the plateau.

  • Theorem 3.1. Lean states positivity of solutions separately from the logarithmic bound, because under the Mathlib convention log⁡0=0\log0 = 0 the bound alone does not exclude zero coordinates. The forward-invariance construction in its proof uses ODE uniqueness, as noted after Definition 2.1.

  • Corollary 3.15. The Lean container is [ε,ε−1]d∩(x0+S)[\varepsilon,\varepsilon^{-1}]^d \cap(x_0 + S) for some 0<ε<10 < \varepsilon< 1, rather than [e−L,eL]d∩(x0+S)[e^{-L},e^L]^d \cap(x_0 + S), with the same conclusions. The case d=0d = 0 is treated separately.

  • Lemma 3.18. The equality of the three sets is assembled from several declarations rather than stated as one.

The following are not formalized: Corollary 3.23; the Wegscheider criterion of Remark 2.4 and the claim in § 5.1 that the example is detailed balanced for all positive rates; the discussions in § 5.2–§ 5.5; and the numerical computations behind Figure 1.

Verification and reproducibility

The project is available at https://github.com/BitMiracle-AI/detailed-balance-gac (tag v1.0). It pins Lean 4.34.1 and Mathlib commit d13f23b723b8a846827a245b89c10fc7d3f11612. In the lean/ directory, lake exe cache get followed by lake build GAC compiles the project from source in a few minutes. Then lake env lean Audit.lean and lake env lean GAC/Reformulated/Audit.lean print the axioms used by the main theorem, the example theorems, and the principal intermediate results. Each depends only on propext, Classical.choice, and Quot.sound, and none on sorryAx. A scan of the sources finds no sorry, admit, additional axioms, unsafe declarations, native_decide, or implemented_by. The repository README describes the module layout and gives the complete commands.

Examples and Discussion

Numerical example

Consider the three-species network A+B⇌CA + B \rightleftharpoons C and 2A⇌B2A \rightleftharpoons B, with rate constants κ1=2,κ2=1\kappa_1 = 2,\kappa_2 = 1 for the first pair and κ3=1,κ4=12\kappa_3 = 1,\kappa_4 = \frac{1}{2} for the second. The reaction vectors (−1,−1,1)(-1,-1,1) and (−2,1,0)(-2,1,0) are linearly independent, so Remark 2.4 gives detailed balance for any positive rate constants. The space S⊥S^\perp is spanned by (1,2,3)(1,2,3), giving the conserved quantity xA+2xB+3xCx_A + 2x_B + 3x_C. Equilibria satisfy xB=(κ3/κ4)xA2x_B = (\kappa_3/\kappa_4)x_A^2 and xC=(κ1/κ2)xAxBx_C = (\kappa_1/\kappa_2)x_Ax_B. In the class xA+2xB+3xC=3x_A + 2x_B + 3x_C = 3, the equilibrium is x=(12,12,12)x = \left(\frac{1}{2},\frac{1}{2},\frac{1}{2}\right). Substitution expresses the conserved quantity as a strictly increasing function of xAx_A, confirming uniqueness.

Figure 1 shows five positive trajectories computed with LSODA using relative tolerance 10−1010^{-10} and absolute tolerance 10−1210^{-12} on 0≤t≤80 \le t \le8. The initial states are (2.997,0.001,0.001/3)(2.997, 0.001, 0.001/3), (0.001,1.499,0.001/3)(0.001, 1.499, 0.001/3), (0.001,0.22,2.559/3)(0.001, 0.22, 2.559/3), (1,1/2,1/3)(1, 1/2, 1/3), and (0.3,0.3,0.7)(0.3, 0.3, 0.7). All lie in the stated class. The trajectories approach the equilibrium, and the entropy evaluated relative to that equilibrium decreases. The computation illustrates the theorem.

Why persistence is needed

Without Step 1, the entropy argument permits an ω\omega-limit set contained in the positive equilibria ∪ the boundary of the positive orthant [14 pp. 89–110]. Boundary equilibria may exist, VV is not differentiable there, and the magnitude of DD can become arbitrarily small near the boundary.

Example 5.1. Consider 2A⇌A+B2A\rightleftharpoons A+B with forward rate κ1\kappa_1 and reverse rate κ2\kappa_2. There is only one reaction vector, so detailed balance holds; take b∗/a∗=κ1/κ2b^*/a^*=\kappa_1/\kappa_2. The conserved quantity is a+b=Ta+b=T and the equation is a˙=a(κ2T−(κ1+κ2)a)\dot{a}=a\left(\kappa_2T-(\kappa_1+\kappa_2)a\right). Besides the positive equilibrium (κ2Tκ1+κ2,κ1Tκ1+κ2)\left(\frac{\kappa_2T}{\kappa_1+\kappa_2},\frac{\kappa_1T}{\kappa_1+\kappa_2}\right), the closure of the class contains the boundary equilibrium (0,T)(0,T). By (11), as a→0a\to0 we have D=O(a∣ln⁡a∣)→0D=O(a\lvert\ln a\rvert)\to0. Thus dissipation can approach zero at a boundary equilibrium as well as at a positive equilibrium. In this example, the logistic equation, or the specific shape of VV on the class, excludes convergence to (0,T)(0,T) from positive initial states. Theorem 3.1 supplies the general compact-interior containment needed for the cluster-point argument.

The role of persistence under detailed balance

The function VV extends continuously to the closed orthant under the convention 0ln⁡0=00\ln0=0, and its sublevel sets there are compact. Along positive trajectories, V(x(t))≤V(x0)V(x(t))\leq V(x_0), so VV alone gives boundedness in the detailed balanced case. What Step 5 needs from Step 1 is lim inf⁡t→∞xi(t)>0\liminf_{t\to\infty}x_i(t)>0 for i=1,…,di=1,\ldots,d. We use the full statement of Theorem 3.1 because it also provides global existence and matches the formal development.

Complex balanced systems

For a complex balanced system, there is a point x∗∈R>0dx^*\in\mathbb{R}^d_{>0} at which total inflow equals total outflow at every complex. Replace the detailed-balance dissipation and equilibrium-identification steps as follows. Put wy=xy/(x∗)yw_y=x^y/(x^*)^y. The inequality wln⁡(w′/w)≤w′−ww\ln(w'/w)\leq w'-w and the complex balance equations give

D(x)=∑eκe(x∗)s(e)ws(e)ln⁡wt(e)ws(e)≤∑eκe(x∗)s(e)(wt(e)−ws(e))=0.D(x)=\sum_e\kappa_e(x^*)^{s(e)}w_{s(e)}\ln\frac{w_{t(e)}}{w_{s(e)}}\leq\sum_e\kappa_e(x^*)^{s(e)}\left(w_{t(e)}-w_{s(e)}\right)=0.

Equality holds exactly when ws(e)=wt(e)w_{s(e)}=w_{t(e)} for every reaction. Thus wyw_y is constant on each weakly connected component. Set ae=κe(x∗)s(e)>0a_e=\kappa_e(x^*)^{s(e)}>0. At zero dissipation, regrouping the vector field by complexes gives

f(x)=∑y∈Cy(∑t(e)=yaews(e)−∑s(e)=yaewy)=∑y∈Cywy(∑t(e)=yae−∑s(e)=yae)=0.f(x)=\sum_{y\in\mathcal{C}}y\left(\sum_{t(e)=y}a_ew_{s(e)}-\sum_{s(e)=y}a_ew_y\right)=\sum_{y\in\mathcal{C}}yw_y\left(\sum_{t(e)=y}a_e-\sum_{s(e)=y}a_e\right)=0.

The first equality groups incoming and outgoing contributions at each complex. On an incoming edge, the equality condition gives ws(e)=wyw_{s(e)} = w_y; complex balance makes each final bracket zero. Conversely, f(x)=0f(x) = 0 implies D(x)=0D(x) = 0. At every positive equilibrium, ln⁡(x/x∗)⋅(t(e)−s(e))=0\ln(x/x^*) \cdot(t(e) - s(e)) = 0 on each reaction, so the logarithmic uniqueness argument of Lemma 3.19 applies. Complex balance implies weak reversibility [5, pp. 187–194]; [9, pp. 172–186], allowing Theorem 3.1 and the unchanged cluster-point argument to apply. This recovers the convergence conclusion covered by Craciun [3, Section 8].

Boundary initial states

The results concern strictly positive initial states. If a nonempty set of initially absent species is a siphon, meaning that every reaction producing a species in the set has a species of the set in its source complex [2], uniqueness of solutions of the polynomial system keeps those species absent, and the trajectory stays on the boundary for as long as the solution exists. In Example 5.1, {A}\{A\} is a siphon. The global attractor conjecture likewise concerns positive initial states.

Logical dependencies

Section 3 presents the mathematical proof, with the geometric input isolated in Theorem 3.1. Section 4 records the formalization, its correspondence with the paper, and the reproducible build. The complete Lean proof uses standard ODE uniqueness in its strong-invariance construction, whereas the direct descent argument of Step 5 has no such dependency. Steps 2–5 apply to any global positive trajectory contained in a compact subset of its compatibility class. Suitable siphon criteria [2] or low-dimensional results [13] may establish the needed persistence for particular networks under their respective assumptions.

Development in ClawsGO Science

The harness closes the loop between proposing and checking. Each formal step written by GPT 6 Astra went directly to the Lean kernel, and the full build and axiom audits of Section 4.4 certify the result, including the adaptation of the 71 upstream modules under new paths, namespaces, and declaration names. For the handwritten arguments and the correspondence between the paper and the code, the harness ran three independent review sessions in parallel, covering Sections 2–3.1, Sections 3.2–5, and the correspondence with Lean; every finding is incorporated in the present version.

Code and data availability

The Lean 4 project of Section 4 is publicly available at https://github.com/BitMiracle-AI/detailed-balance-gac under the Apache License 2.0; the version described in this paper is tagged v1.0, and Section 4.4 gives the commands that reproduce the check. The project and the numerical recomputation of Section 5.1, cross-checked with a second solver, were produced in ClawsGO Science [16].

Acknowledgements and use of AI

We acknowledge OpenAI’s work on boundedness and persistence of weakly reversible mass-action systems [15], from which Theorem 3.1 and Sections 3.1.1–3.1.7 derive. Seventy-one Lean modules of that work, at commit 1581d4cbdd69a0faf34b2089614fe2a814316a6a, are adapted into GAC/Persistence/, with renamed paths, namespaces and principal declarations and with modification notices. The Apache-2.0 license is retained in GAC/Persistence/LICENSE; the old/new paths and separate upstream/current hashes are recorded in lean/docs/SOURCE_MAP.md and its machine-readable companion.

The proofs, the Lean development, and the manuscript were produced in ClawsGO Science [16], with Claude Opus 5.5 making the research decisions and GPT 6 Astra executing them. Together they developed the detailed-balance argument and its connection with Theorem 3.1, the logarithmic bounds, feasible directions, and exponential parameterization in the main text, wrote and audited the Lean 4 development (the GAC modules and the manuscript and coordinate/geometric interfaces), constructed the exact examples and the numerical illustration, and drafted and translated the manuscript. The human authors directed the work, checked the arguments and the correspondence between the Lean statements and the mathematical definitions, and take full responsibility for the content. No AI system is an author of this paper.

References

  1. [1]D. F. Anderson, A proof of the global attractor conjecture in the single linkage class case, SIAM J. Appl. Math. 71 (2011), no. 4, 1487–1508. https://doi.org/10.1137/11082631X.
  2. [2]D. Angeli, P. De Leenheer, and E. D. Sontag, A Petri net approach to the study of persistence in chemical reaction networks, Math. Biosci. 210 (2007), no. 2, 598–618. https://doi.org/10.1016/j.mbs.2007.07.003.
  3. [3]G. Craciun, Toric differential inclusions and a proof of the global attractor conjecture, preprint, arXiv:1501.02860v3 (2026). https://arxiv.org/abs/1501.02860v3.
  4. [4]L. de Moura and S. Ullrich, The Lean 4 theorem prover and programming language, in: A. Platzer and G. Sutcliffe (eds.), Automated Deduction – CADE 28, Lecture Notes in Comput. Sci., vol. 12699, Springer, Cham, 2021, 625–635. https://doi.org/10.1007/978-3-030-79876-5_37.
  5. [5]M. Feinberg, Complex balancing in general kinetic systems, Arch. Rational Mech. Anal. 49 (1972), 187–194. https://doi.org/10.1007/BF00255665.
  6. [6]M. Feinberg, Necessary and sufficient conditions for detailed balancing in mass action systems of arbitrary complexity, Chem. Eng. Sci. 44 (1989), no. 9, 1819–1827. https://doi.org/10.1016/0009-2509(89)85124-3.
  7. [7]M. Feinberg, Foundations of Chemical Reaction Network Theory, Applied Mathematical Sciences, vol. 202, Springer, Cham, 2019. https://doi.org/10.1007/978-3-030-03858-8.
  8. [8]M. Gopalkrishnan, E. Miller, and A. Shiu, A geometric approach to the global attractor conjecture, SIAM J. Appl. Dyn. Syst. 13 (2014), no. 2, 758–797. https://doi.org/10.1137/130928170.
  9. [9]F. Horn, Necessary and sufficient conditions for complex balancing in chemical kinetics, Arch. Rational Mech. Anal. 49 (1972), 172–186. https://doi.org/10.1007/BF00255664.
  10. [10]F. Horn, The dynamics of open reaction systems, in: Mathematical Aspects of Chemical and Biochemical Problems and Quantum Chemistry, SIAM–AMS Proc., vol. 8, American Mathematical Society, Providence, RI, 1974, 125–137.
  11. [11]F. Horn and R. Jackson, General mass action kinetics, Arch. Rational Mech. Anal. 47 (1972), 81–116. https://doi.org/10.1007/BF00251225.
  12. [12]The mathlib Community, The Lean mathematical library, in: J. Blanchette and C. Hrițcu (eds.), Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs, ACM, New York, 2020, 367–381. https://doi.org/10.1145/3372885.3373824.
  13. [13]C. Pantea, On the persistence and global stability of mass-action systems, SIAM J. Math. Anal. 44 (2012), no. 3, 1636–1673. https://doi.org/10.1137/110840509.
  14. [14]D. Siegel and D. MacLean, Global stability of complex balanced mechanisms, J. Math. Chem. 27 (2000), nos. 1–2, 89–110. https://doi.org/10.1023/A:1019183206064.
  15. [15]OpenAI, Boundedness and persistence of weakly reversible mass-action systems, preprint, September 25, 2026, with Lean source at commit 1581d4cbdd69a0faf34b2089614fe2a814316a6a, https://github.com/openai/math/tree/1581d4cbdd69a0faf34b2089614fe2a814316a6a.hub.valency.io/works/W_gtf8ahr3
  16. [16]ClawsGO Science, an agent harness with an integrated research environment, https://clawsgo.ai, accessed October 21, 2026.

Paper details

vlnc:W_zbcgufzp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ClawsGO Science, Claude Opus 5.5, GPT 6 Astra

View statement
Looking for related papers…
Contents