Weakly reversible mass-action systems: persistence and permanence

Source and verification

This article explains family 149 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Reaction networks and conservation classes

A reaction complex is a vector $y\in {\mathbb{Z}}_{\ge 0}^d$ recording the numbers of each species. For a directed reaction $y\to y^{\prime }$ with fixed rate $k_{y\to y^{\prime }}>0$, mass-action kinetics contributes $k_{y\to y^{\prime }}{x}^y(y^{\prime }-y)$, where $x^y={\prod }_i{x}_i^{y_i}$.

Thus the concentration vector solves the polynomial differential equation

$$\dot{x}=\sum _{y\to y^{\prime }}{k}_{y\to y^{\prime }}{x}^y(y^{\prime }-y).$$

A network is weakly reversible when every directed reaction has a directed return path. An individual reverse edge is not required.

The stoichiometric subspace and positive class through $c>0$ are

$$\begin{array}{c}S=\mathrm{span}\{y^{\prime }-y\},\\ P=(c+S)\cap {\mathbb{R}}_{>0}^d.\end{array}$$

Since the vector field lies in $S$, solutions preserve the affine conservation constraints. The class itself may be unbounded.

🏷️ Persistence versus permanence

Persistence requires ${\mathrm{liminf}}_{t\to \infty }{x}_i(t)>0$ for every coordinate of an individual trajectory. Permanence additionally asks for one compact positive set eventually containing every trajectory in a fixed class, with entry time allowed to depend on the initial point.

The September manuscript claims that every positive initial state lies in a compact convex forward-invariant polytope inside the positive orthant. Consequently its solution is global and obeys

$$\epsilon (x^0)\le x_i(t)\le \epsilon (x^0{)}^{-1}(t\ge 0).$$

The constant may depend on the initial state. (OpenAI, 2026)

Stronger statement in the October companion

For each positive stoichiometric class $P$, there is a nonempty compact convex forward-invariant set $K_P\subset P$ such that every positive trajectory in $P$ enters $K_P$ in finite time and remains there. The set is independent of the initial point; the entry time need not be. (OpenAI, 2026b)

Both assertions assume fixed positive rates. They do not assert the same conclusion for arbitrary time-varying rates merely from weak reversibility. Nor does permanence alone imply convergence to an equilibrium.

🏷️ A two-species example

For $A\rightleftarrows B$ with rates $k_1,k_2$,

$$\begin{array}{c}\dot{a}=-k_{1}a+k_{2}b,\\ \dot{b}=k_{1}a-k_{2}b.\end{array}$$

The total $a+b=M$ is conserved, and

$$\begin{array}{c}{a}_{\ast }=\frac{k_{2}M}{k_1+k_2},\\ a(t)-a_{\ast }=(a(0)-a_{\ast })e^{-(k_1+k_2)t}.\end{array}$$

Every positive initial state converges to the same interior equilibrium in its class. This directly exhibits classwise eventual bounds, but no single positive lower bound can hold at time zero for all positive initial states: $a(0)$ can be arbitrarily close to zero.

The general theorem has the same distinction between a uniform eventual set and nonuniform entry times, even when an explicit solution and a bounded conservation class are unavailable.

🏷️ Matching affine normals to monomials

Write $x_i=h^{p_i}$ with $0<h<1$ and $p\in [-1,1{]}^d$, so $x$ ranges over $B_h=[h,h^{-1}{]}^d$. The sources construct a concave piecewise-affine function

$$F_h(x)=\underset{\lambda \in \Lambda }{\min }(r_{\lambda }\cdot x+b_{\lambda }(h)).$$

A label is active if it attains the minimum. The finite slopes $r_{\lambda }$ are independent of $h$ and must approximate the exponent vector $p$ whenever they are active.

The approximation tolerance shrinks near the hyperplanes where relevant comparisons vanish. Continuity of this tolerance is not assumed. The aim is to preserve the sign of every nonzero comparison between complexes. After choosing the finite slopes and their separation, small $h$ gives

$$r_{\lambda }\cdot y<r_{\lambda }\cdot z⟹x^z\le h^H{x}^y$$

for some fixed $H>0$ and every active label. (OpenAI, 2026a)

The order of choices matters. A finite covering of exponent space by approximate normals is insufficient unless the slopes actually attaining the affine minimum have the required accuracy.

🏷️ Directed cuts produce an inward inequality

Order reaction complexes by $r_{\lambda }\cdot y$ and place cuts between successive distinct levels. If a reaction crosses a cut downward, its return path must cross upward somewhere.

For $h<1$, lower exponent levels give larger monomials. The source chooses $h$ small enough that one upward flux dominates the sum of downward fluxes across the cut. Summing the cut fluxes weighted by their level gaps yields

$$r_{\lambda }\cdot \dot{x}\ge 0$$

for every active label. Reactions within one level contribute zero to this normal derivative, even if their monomials differ.

At a tie, the right derivative of a finite minimum is the minimum of the active directional derivatives. Therefore all active labels must satisfy the inequality; choosing one favorable label would not prove that $F_h$ is nondecreasing.

The maximum set of $F_h$ contains the prescribed initial point, is convex, and avoids the boundary of $B_h$. Convexity then confines it entirely inside the positive box. Monotonicity and the maximum ceiling make this set forward invariant, giving global continuation and the initial-state-dependent bounds.

🏷️ The upgrade to one absorbing set

The later manuscript fixes a representative $c\in P$. A slope in $S^{\perp }$ is constant on the affine class. Such a branch cannot be active below the maximum outside the plateau.

Every active slope on $P$ outside the plateau therefore has a strictly positive directional derivative. Compactness at a fixed scale makes this strictness uniform over the relevant finite set of labels and points. A trajectory remaining in $B_h$ must reach the plateau in finite time. (OpenAI, 2026b)

The trapping plateaus satisfy a tighter inclusion into $B_{\sqrt{h}}$. Repeated square-root changes move from a scale accommodating an arbitrary initial state to a final scale chosen only from the class and rates. Intersecting the final plateau with $c+S$ gives $K_P$.

This scale iteration is the additional work establishing permanence. The earlier individual trapping theorem does not already contain its classwise quantifier.

🏷️ Key steps and proof outlines

One upward reaction dominates an entire downward cut

Fix an active slope $r$ and list the distinct levels of complexes as $a_1<\cdots <a_m$, where each level is $r\cdot y$. For the cut at $a_j$, define $Q_j(x)$ as total upward reaction flux minus total downward flux. The affine construction supplies the quantitative ordering $x^z/x^y\le h^H$ whenever $r\cdot y<r\cdot z$.

If a downward reaction crosses this cut, weak reversibility gives a directed return path, which must contain an upward reaction $w\to w^{\prime }$. Every downward source lies at a higher level than $w$. Therefore

$$\sum _{\text{downward }u\to u^{\prime }}{k}_{u\to u^{\prime }}{x}^u\le h^H{x}^w\sum _e{k}_e<k_{w\to w^{\prime }}{x}^w$$

once the scale is small enough. The last comparison uses the fixed positive rates and is chosen uniformly over the finite edge set. Thus one upward reaction outweighs all downward reactions; it is unnecessary to pair downward edges individually with different upward edges. If there are no downward reactions, the flux is already nonnegative, and is positive whenever any upward reaction exists.

Every reaction’s level difference is the signed sum of the intervening consecutive gaps, giving the exact identity

$$r\cdot f(x)=\sum _{j=1}^{m-1}(a_{j+1}-a_j)Q_j(x).$$

All terms are nonnegative. If $r\notin S^{\perp }$, some reaction has nonzero level difference and crosses a cut, so one term is positive. Reactions staying within a level and self-loops contribute zero. This is the flux mechanism in “Compact plateaux and strict increase,” built on the earlier trapping construction. (OpenAI, 2026b; OpenAI, 2026a)

Strict increase is uniform on active sets

A branch with slope in $S^{\perp }$ is constant on the affine class $c+S$. Its value there is at least $F_h(c)=M_h$, whereas outside the plateau $F_h<M_h$. Such a branch cannot be active outside the plateau. Thus every active branch there has strictly positive normal derivative.

For a fixed label $\lambda$, its active set $\{x\in B_h:L_{\lambda ,h}(x)=F_h(x)\}$ is compact. The continuous function $r_{\lambda }\cdot f$ has a positive minimum on every nonempty such set whose slope lies outside $S^{\perp }$. Taking the minimum over finitely many labels gives one ${\delta }_h>0$. This step uses compactness of the full active set, rather than the generally nonclosed complement of the plateau.

At a tie, the right derivative is

$$\frac{d^{+}}{dt}{F}_h(x(t))=\underset{\lambda \text{ active at }x(t)}{\min }{r}_{\lambda }\cdot f(x(t)).$$

To justify it, inactive branches have a positive value gap and remain inactive for sufficiently short time. Each active branch has a first-order expansion, and the minimum of finitely many expansions has the displayed right derivative. Since every active derivative is at least ${\delta }_h$ outside the plateau, a trajectory that stays in $B_h$ must reach the ceiling $M_h$ in finite time. The elementary right-derivative comparison gives an entry-time bound in terms of the initial deficit divided by ${\delta }_h$.

The scale iteration uses boxes, not nesting of plateaux

Choose $h_0$ sufficiently small that the initial point belongs to the zero-slope plateau $K_{h_0}$. It is forward invariant and compact inside the positive orthant, giving a global positive trajectory. The sharper geometric inclusion is $K_h\subset \mathrm{int}{B}_{\sqrt{h}}$.

Set $h_{j+1}=\min (\sqrt{h_j},h_P)$, where $h_P$ was fixed from the class and rates. Once the trajectory has entered $K_{h_j}$, it stays inside $B_{h_{j+1}}$, because $B_{\sqrt{h_j}}\subset B_{h_{j+1}}$. The next scale’s finite-entry argument applies throughout this tail and puts it into $K_{h_{j+1}}$. No inclusion $K_{h_j}\subset K_{h_{j+1}}$ is assumed or needed.

Repeated square roots reach the capped scale $h_P<1$ after finitely many iterations. Hence the final set $K_{h_P}\cap (c+S)$ absorbs every trajectory in the class. It is nonempty because it contains $c$, compact and convex, and forward invariant because both factors are. The final lower and upper coordinate bounds depend only on $h_P$, while the number and durations of the preceding entry stages may depend on the initial point. This is precisely the extra quantifier needed for permanence beyond individual persistence.

See Also

Polynomial limit cycles studies periodic behavior in structured polynomial flows. Logarithmic convex interpolation also uses soft or piecewise descriptions of convex constraints, with a different analytic purpose.

🏷️ Manuscripts and verification scope

• Uniform Permanence in Weakly Reversible Mass-Action Systems — pinned manuscript and companion source.
• Boundedness and persistence of weakly reversible mass-action systems — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope’s classwise-permanence heading is broader than its linked result: it selects the earlier global boundedness and persistence theorem with bounds depending on the initial state. The later classwise absorbing-set theorem is not included in that selection. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Boundedness and persistence of weakly reversible mass-action systems.

🐻  OpenAI 2026b. Uniform Permanence in Weakly Reversible Mass-Action Systems.