1Two jars
The problem. Picture two jars, each with a few chemicals reacting in it. The first keeps changing colour, back and forth, on its own, for minutes. Chemists call this a chemical clock, and it is real (the Belousov–Zhabotinsky reaction is the famous one). The second jar changes once and then stays the same. Most jars are like the second one.
To find out which kind you have, the usual way is to simulate it. But a simulation needs the speed of every reaction, and a living cell has hundreds of reactions whose speeds nobody has measured. So the real question is: can you tell a clock from a jar that settles by reading only the recipe?
A toy chemical clock (the Brusselator) with two chemicals, X and Y. Move the slider past 2 and it starts to tick.
2A map of the jar
The idea. Take a jar with only two chemicals, A and B, and draw a map: going right means more A, going up means more B. Now the whole jar is one dot on the map. As the chemicals react, the dot moves. A clock is a dot that goes round and round forever. A jar that settles is a dot that stops.
Say the recipe is “one A splits into two B”. Every time that happens, A goes down by one and B goes up by two. On the map that is always the same little step: one left, two up, (−1, 2). It does not matter how fast the reaction runs; the step always points the same way.
Pick a reaction. Drag the orange dot: the step is the same wherever the jar is.
3A bead on a wire
The claim. If the reaction can also run backwards (two B join back into one A), that is the same step in reverse. Both steps lie along one line, so the dot can never leave that line. It is a bead on a wire. This is just counting atoms: they are never made or destroyed. Each jar, depending on how it started, has its own wire (chemists say stoichiometric compatibility class).
Where does the bead stop? Where splitting and joining happen equally often. All those balance points form a curve (for the usual rule that a reaction's speed is proportional to the amounts that meet, it is k₁·A = k₂·B²). The bead slides along its wire until it reaches the curve. Every wire crosses the curve exactly once, so every jar settles, in exactly one place. A bead on a wire cannot go round in circles: this recipe can never be a clock.
Tap anywhere to start a new jar. Each bead slides along its dashed wire to the green curve. Change the two speeds: the curve bends, but every wire still crosses it exactly once.
Read the detail: the equations behind figure 3
Mass action for A ⇌ 2B gives dA/dt = −k₁A + k₂B² and dB/dt = 2k₁A − 2k₂B². Add the first to half the second and you get zero: that is the line. Setting both to zero gives A = (k₂/k₁)B²: that is the curve. The figure integrates the equations with a fourth-order Runge–Kutta step.
4A test you do by counting
The theorem. With five chemicals you cannot draw the map any more. In the 1970s the chemical engineer Martin Feinberg found a way to get the same answer by counting three things in the recipe:
- Pieces: the different things written on either side of an arrow. In our recipe, “A” and “2B”: two.
- Groups: how many separate parts the recipe falls into when you draw it. Ours: one.
- Directions: how many independent ways the steps point. Ours: one (the wire).
Pieces minus groups minus directions is called the deficiency. Ours is 2 − 1 − 1 = 0. Feinberg’s theorem: if it is zero and every reaction can be undone (directly or by a roundabout path), every jar settles to exactly one resting place, and it stays there. No clocks, no switches, whatever the speeds are. If it is zero and some reaction cannot be undone, there is no balanced state with everything present: something always runs out.
Pick a recipe. The counts are computed live from the reactions.
5What it does not tell you
The limits. The test says nothing when the count is above zero. The Lotka network, a chemical cousin of predators and prey, counts 1, and it does tick. The recipe B → A, A + 2B → 3B also counts 1. Depending on how much material is in the beaker, it has zero, one or two resting points. Feinberg’s deficiency one theorem and computer checks take over there.
“Stable” here means locally stable: start close and you return. The claim that every start ends at that point, the global attractor conjecture, is still open in general.
Why we liked it. It has nothing to do with robots. But the move is the one we want everywhere: find what cannot change (the line) and where things must stop (the curve) before you simulate anything. A count on the wiring diagram can then rule out whole families of behaviour.
Sources. Casual Physics Enjoyer, Chemistry is Actually Geometry (Substack, 2026). M. Feinberg, Foundations of Chemical Reaction Network Theory (Springer, 2019). M. Feinberg, “Chemical reaction network structure and the stability of complex isothermal reactors” (Chem. Eng. Sci., 1987).
Part of Unrelated learnings: notes we write while reading outside robotics. Corrections are welcome at founders@stormy.ai.