A common safe region for every trajectory of a weakly reversible reaction network

← All digests

Family 149 · Dynamical systems and ergodic theory
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-21

1. The question

Can the wiring diagram of a chemical reaction network prevent any of its species from disappearing? A reaction can consume one substance rapidly, yet a chain of other reactions may regenerate it. The relevant structural condition is weak reversibility: every reaction has a directed route back to its starting combination of species. The reverse need not be a single reaction.

Family 149 asks whether this condition controls all long-term concentrations under mass-action kinetics. Two manuscripts give progressively stronger claims. The earlier one claims that each positive trajectory stays inside some bounded region away from zero. The later one claims that every trajectory with the same conservation constraints eventually enters one common compact region. That uniformity is the distinction between individual persistence and classwise permanence.

The conclusion includes networks whose conservation constraints allow arbitrarily large concentrations. Thus boundedness is an output, rather than an assumption that a finite supply of material has already settled the problem.

2. Background and essential definitions

Reactions become polynomial equations

Suppose there are $d$ species with concentration vector $x=(x_1,\ldots,x_d)$. A complex is a vector $y$ of nonnegative integers recording a combination of species. For example, with species $A,B$, the complex $2A+B$ is $(2,1)$.

A reaction $y\to y'$ has a positive constant rate $k_{y\to y'}$. Mass action assigns it flux $k_{y\to y'}x^y$, where $x^y=\prod_i x_i^{y_i}$. The resulting differential equation is

$$\dot x=\sum_{y\to y'}k_{y\to y'}x^y(y'-y).$$

Each term multiplies the reaction frequency by the change in species amounts. For $2A\rightleftarrows B$, writing concentrations $a,b$ and rates $k_+,k_-$, this gives

$$\dot a=-2k_+a^2+2k_-b,\qquad \dot b=k_+a^2-k_-b.$$

Consequently $a+2b$ is constant. This elementary example has bounded concentration classes, but general networks need not.

Conservation classes and three kinds of control

Let $S$ be the vector space spanned by all reaction changes $y'-y$. Starting from a strictly positive vector $c$, the solution remains in the affine space $c+S$. Its positive stoichiometric compatibility class is

$$P=(c+S)\cap(0,\infty)^d.$$

The adjective “stoichiometric” refers to the species bookkeeping encoded in the reaction changes. Two initial states in the same class obey the same linear conservation constraints.

A trajectory is bounded if no concentration grows without limit. It is persistent if every concentration has a strictly positive lower limit as time tends to infinity. The bounds may depend on the trajectory.

Permanence on $P$ is stronger: one compact subset of $P$ eventually contains every positive trajectory in that class. Different initial states may take different times to arrive. A set is forward invariant if a trajectory starting there stays there for every later time.

These are statements about deterministic concentrations. They do not say that a finite stochastic population has zero probability of extinction.

3. Development history and previous partial results

Horn and Jackson’s 1972 theory and Feinberg’s work established the special role of complex balance. At a complex-balanced equilibrium, the total flux entering each complex equals the total leaving it. This provides an entropy-like function decreasing along trajectories and a unique positive equilibrium in each positive class. Complex balance depends on rate constants as well as the reaction graph; weak reversibility alone imposes fewer conditions.

Feinberg’s discussion of weakly reversible systems raised the question of excluding limiting states with zero concentrations. Persistence strengthens this by excluding boundary approach even along a sequence of times. Anderson subsequently distinguished the boundedness question from persistence for already bounded trajectories. A proof of one does not automatically supply the other.

Several substantial special cases were known. Craciun, Nazarov, and Pantea established permanence for two-species weakly reversible systems within a broader geometric class called endotactic networks. Pantea established persistence of bounded trajectories when the stoichiometric space has dimension two. These are different restrictions: a system can have many species but only two independent reaction directions.

Gopalkrishnan, Miller, and Shiu proved permanence for strongly endotactic systems, a geometric class including weakly reversible networks with a single connected component of the reaction graph. Boros and Hofbauer later gave another proof in that single-component case. Several such results allow rates to vary with time within fixed positive upper and lower bounds.

The manuscripts also discuss Craciun’s toric differential inclusions. They distinguish the September 2026 revision’s persistence statement for bounded weakly reversible trajectories from its complex-balanced global-convergence application. The general persistence statement does not itself supply the missing upper bound.

The earlier manuscript in this family notes a previous general claim by August and Barahona and identifies an issue in its monomial-dominance argument. This is the manuscript’s account of the literature; the present exposition does not independently assess that proof. Its own construction is presented as an independent route to simultaneous boundedness and persistence.

Finally, the later manuscript explicitly builds on the earlier one’s affine approximation lemma. Its additional contribution is the passage from a trapping region chosen for one initial state to an absorbing region shared by an entire class.

4. What the new manuscripts claim

Individual bounds, then uniform eventual bounds

M2, dated 25 September 2026, considers any finite weakly reversible network, in any species dimension, with positive constant reaction rates. For every strictly positive initial state, it claims a unique solution for all forward time and a number $\varepsilon>0$ such that

$$\varepsilon\le x_i(t)\le\varepsilon^{-1} \quad\text{for every }t\ge0\text{ and every species }i.$$

Here $\varepsilon$ may depend on the initial state. More geometrically, the manuscript constructs a compact convex positive polytope containing that state and invariant under the dynamics.

M1, dated 5 October 2026, strengthens the quantifiers. For each fixed positive class $P$, it claims a nonempty compact convex forward-invariant set $K_P\subset P$ such that every initial state $x_0\in P$ has a finite entry time $T(x_0)$, after which its trajectory stays in $K_P$. In particular,

$$\varepsilon_P\le x_i(t)\le\varepsilon_P^{-1} \quad\text{whenever }t\ge T(x_0).$$

The same $\varepsilon_P$ works for all initial states in that class. It may depend on the network, its fixed rates, and the class. Neither the entry time nor the bound is asserted to be uniform over different classes or rate choices.

There is no restriction on the number of connected components, usually called linkage classes. The class $P$ can be unbounded. Initial concentrations must nevertheless be strictly positive.

M2 also recovers the known existence of a positive equilibrium in every positive class, crediting Boros. With the additional complex-balance assumption, it derives convergence to the class’s unique positive equilibrium. General weak reversibility is not claimed to force such convergence.

5. Why the advance matters

Provisional importance: 5/5. If correct, the family resolves central persistence, boundedness, and permanence questions in chemical reaction network theory for fixed positive rates. Its significance comes from replacing dimension and graph-component restrictions with a basic structural condition.

The common absorbing set is particularly useful. A separate bound for each experiment says that each individual run behaves safely, but gives no shared eventual range for a whole class of starting conditions. Permanence supplies that shared range. It controls the long-term dynamics before one knows whether trajectories settle to equilibria, oscillate, or exhibit more complicated behavior.

The methods also address a general difficulty in polynomial differential equations: some coordinates may become tiny while others become large. A single estimate based on total concentration can miss that mixed behavior. Here the construction compares the individual monomials responsible for the reactions.

6. The main idea

Compare reaction strengths on logarithmic scales

Choose a small number $0<h<1$ and write positive concentrations in a box as

$$x_i=h^{p_i},\qquad -1\le p_i\le1.$$

Then $x^y=h^{p\cdot y}$. Because $h<1$, a smaller exponent means a larger monomial. Thus ordering complexes by a linear function of their exponent vectors reveals which reaction fluxes dominate.

The construction uses a concave piecewise-affine function

$$F_h(x)=\min_{\lambda}\bigl(r_\lambda\cdot x+b_\lambda(h)\bigr).$$

There are finitely many labels $\lambda$. A label is active where its affine expression achieves the minimum. The central approximation lemma chooses the slopes $r_\lambda$ so that every active slope closely matches the exponent vector $p$ at that point.

The required accuracy is especially delicate near ties between complexes. A careless small error could reverse their order. The lemma permits a tolerance that changes discontinuously near comparison hyperplanes, so all nonzero comparisons have the correct sign. It controls every active label, including several that tie.

Return paths make a barrier point inward

Order the complexes by their value under an active slope and cut the ordered list between consecutive levels. If a reaction crosses the cut in one direction, weak reversibility supplies a crossing in the opposite direction along a return path. The monomial comparison makes an appropriately directed flux dominate the sum of adverse fluxes when $h$ is small enough.

Summing over cuts shows that $F_h$ cannot decrease along a trajectory in the relevant box. Its maximum set is a compact convex polytope. Once a trajectory lies on this maximum plateau, nondecrease prevents it from leaving. Choosing the scale to include a prescribed initial state gives M2’s individual trapping region.

M1 adds strictness outside a class’s plateau. A slope perpendicular to $S$ is constant on $c+S$, so it cannot be the active obstruction outside the desired maximum set. Every remaining relevant slope gives strict increase. At a fixed scale, compactness turns this into a positive lower bound on the increase, forcing finite-time entry.

The final step connects scales. A trapping set at scale $h$ lies inside the box associated with $\sqrt h$. Repeatedly replacing $h$ by its square root, stopping at a scale fixed by the class, brings every trajectory into one common region after finitely many stages. This uses containments in successive boxes; it does not assume that all the trapping polytopes are nested.

7. Scope, limitations, and open questions

The rates are fixed positive constants. Earlier special-case theorems sometimes handle bounded time-varying rates, but that stronger rate flexibility should not be transferred to these general claims.

Permanence also does not give a common entry time for initial states arbitrarily close to the boundary or arbitrarily far away in an unbounded class. The stated uniformity concerns the eventual region. Practical estimates for its size and for entry times are separate quantitative questions.

The claims exclude extinction and escape to infinity from positive initial states. They do not classify attractors or imply uniqueness of equilibrium for arbitrary weakly reversible systems. Nor do they cover boundary initial states, non-mass-action kinetics, infinite reaction networks, or stochastic reaction counts without additional work.

The conceptual separation remains useful: complex balance supports convergence to a particular equilibrium; weak reversibility supplies the broader geometric protection expressed by permanence.

8. Sources and further reading

Horn and Jackson, General mass action kinetics (1972), is a primary source for complex balance and its stability theory. Feinberg’s Chemical reaction network structure and the stability of complex isothermal reactors—I (1987) develops the structural setting and the boundary question.

For earlier partial results, see Craciun, Nazarov, and Pantea, Persistence and permanence of mass-action and power-law dynamical systems (2013); Pantea, On the persistence and global stability of mass-action systems (2012); and Gopalkrishnan, Miller, and Shiu, A geometric approach to the global attractor conjecture (2014). Boros and Hofbauer’s Permanence of weakly reversible mass-action systems with a single linkage class (2020) is the closest earlier absorbing-set result cited here.

Craciun’s toric differential inclusion manuscript, version 3 supplies the separate bounded-trajectory and complex-balanced context. The precise claims and construction summarized above are those of the two pinned manuscripts below.

Source manuscripts

• M1. OpenAI, Uniform Permanence in Weakly Reversible Mass-Action Systems (October 5, 2026).
• M2. OpenAI, Boundedness and persistence of weakly reversible mass-action systems (September 25, 2026).

Sources consulted on 21 October 2026 at repository revision 1581d4c. Download the manuscript citations.