15.4 Chemical Kinetics & Stiff Systems ended with a determinant going negative. The Brusselator with , is a stable spiral alone, yet once the two species diffuse at different speeds, has determinant at and a mode grows. That was a promise. This lesson cashes it: where the unstable band is, which wavelength wins, what the nonlinear pattern looks like and what it costs to compute. On the way we settle an IOU from 15.3 Epidemic Models: the front speed quoted there from memory is now measured.
Transcript
- Last lesson promised a pattern. The Brusselator at b equal 1.9 is stable alone, yet with D x equal 1 and D y equal 10 the determinant of J minus q squared D is minus 0.6 at q squared equal 0.4, and a mode grows. One diffusing species only smooths; a pair with unequal diffusivities can make structure. First, an IOU: the front speed 2 root D r from the epidemics lesson, now measured.
- Take logistic growth with diffusion, u t equals D u x x plus r u times one minus u. Far ahead of a front u is tiny, so a tail e to the minus lambda times x minus c t gives c of lambda equals D lambda plus r over lambda. Steep tails are limited by diffusion, shallow ones by growth, with the minimum at lambda star equal root r over D. The speed is c star equals 2 root D r: for D and r equal 1, 2. Doubling D multiplies it by root 2, not 2.
- Explicit FTCS with dx equal 0.1, over t from 20 to 40, gives 1.946 against 2.000. For D equal 2 it gives 2.752 against 2.828, and for r equal 0.5, 1.330 against 1.414: 3 percent to 6 percent low, since pulled fronts approach c star slowly, reaching 1.983 for t from 80 to 100. The tail matters: e to the minus 0.5 x runs at 2.499 and e to the minus 0.7 x at 2.127, but a steeper e to the minus 2 x runs at 1.947, near 2, not 2.5. That selection rule is from memory.
- Perturb a uniform state by cos of q x. Then sigma v equals J minus q squared D, times v, and the top eigenvalue is the dispersion relation. One species can only be stabilised. For two, stable without diffusion, only the determinant can go negative: D 1 D 2 q to the fourth minus B q squared plus det J. That needs B above 2 root D 1 D 2 det J: a self-activating species and a faster inhibitor, local activation with long range inhibition. For the Brusselator, b must exceed 1 plus a root D x over D y, squared.
- At a equal 1, D y equal 10 and b equal 1.9, J has eigenvalues minus 0.05 plus or minus 0.9987 i, a stable spiral. The determinant vanishes at q squared equal 0.1551 and 0.6449: the band is q from 0.394 to 0.803, wavelengths 7.82 to 15.96, with threshold b equal 1.7325. The fastest mode has q squared 0.3216, growth 0.1425, wavelength 11.08. From b equal 1.75 to 1.95 peak growth rises about twelvefold, while the wavelength stays near 11. Chemistry and diffusion set it, not the domain.
- Linear theory says which modes grow, not where growth stops. The nonlinear Brusselator on 0 to 100, with 201 nodes and noise of size 0.01, settles into stationary peaks by t equal 400; seed 1 spaces them 11.773 apart. Over 12 seeds, 7 gave 11.7, 3 gave 11.1 and 2 gave 10.5: all in the band, within 6.3 percent of 11.08. Modes 17, 18 and 19 grow within 3 percent of each other, so the nonlinear terms decide, and the linear winner came out only 3 times in 12.
- Explicit stepping needs the checkerboard mode inside the stability interval; the faster diffusivity binds. With the reaction the most negative eigenvalue is minus 160.99, so h must stay below 0.012423. The checkerboard factor is 0.9962 at h equal 0.0124 and 1.0284 at 0.0126, which blows up within about 800 steps. At h equal 0.01 the run takes 40000 steps. IMEX treats diffusion implicitly: at h equal 0.2 it needs 2000 steps, 20 times fewer, with spacing 11.776 against 11.773. At h equal 0.3 the pattern is lost, for a reason I did not isolate.
- Gray-Scott is a poor model to analyse: its empty state is stable for every F and k, so no noise grows and patterns need a finite seed, like a Fisher front. With D u 0.16, D v 0.08 and a periodic 128 by 128 grid, F 0.030 and k 0.062 give 89 blobs at t equal 5000, and 110 at t equal 15000: still changing. F 0.040 and k 0.060 give one connected labyrinth covering 52.9 percent. The explicit limit is 1.5267: the checkerboard decays at 0.965 for h equal 1.5 and grows at 1.030 for 1.55.
- Turing proposed in 1952 that morphogens could pattern an embryo. The mathematics is exact; whether a given animal is patterned this way is a hypothesis. The diffusivity ratio must exceed 6.98, the wavelength depends on diffusivities only through a fourth root, and noise changes the outcome. Nothing here was fitted to an animal. Next, a neuron rests stably, but a big kick sends a pulse down the axon, then it recovers: an excitable medium, with a slow recovery variable that need not diffuse. The next lesson, on neuron models, supplies the reaction terms. Whether 2 root D r survives is the first question to ask.
Reaction plus diffusion
A reaction-diffusion system adds a Laplacian to each rate equation of a reaction network:
Without the Laplacian this is the ODE of the last three lessons at every point; without it is the heat equation of 9.2 The Heat (Diffusion) Equation. One species with diffusion can only spread and smooth. A pair with unequal diffusivities can do the opposite and make structure from uniformity.
Fisher-KPP: a front with a speed
Take one species with logistic growth, as in 15.1 Population Models:
With on the left and on the right, a front invades the unstable state . How fast?
Derivation. Far ahead of the front is tiny, , and the equation is linear: . Try a tail . Then , so
is large for steep tails (diffusion-limited) and shallow ones (growth-limited), with its minimum at . The Kolmogorov-Petrovsky-Piskunov result (1937; the selection rule is from memory) is that compactly supported data run at exactly the minimum. It is a pulled front: the linear leading edge sets the speed and the saturated bulk follows. For , is minimal at with ; doubling multiplies the speed by , not .
Measured. The sim runs explicit FTCS (10.2 Explicit Methods for the Heat Equation) with , , tracks , and takes the speed over . Scipy’s RK45 on the same semi-discrete system is the independent check.
| sim | scipy | |||
|---|---|---|---|---|
The measured speeds sit below , by to . Pulled fronts approach slowly. Bramson’s correction (from memory) puts the front at , giving window averages , and : within of scipy for the first two rows and for the third. On a longer run, gives (scipy , correction ). The time step matters: reads in the first row.
The tail picks the speed. If the initial data decay slower than , that tail sets the speed (from memory). Test with :
| initial tail | predicted | sim | scipy |
|---|---|---|---|
| (not ) |
Linear stability with diffusion
Let be a uniform steady state with Jacobian . Perturb it by . The Laplacian turns into , so
The growth rate is the largest real part of the eigenvalues of : the dispersion relation. For one species , so diffusion only stabilises. For two it need not.
Take and (stable without diffusion). Then is negative for every , and only a negative determinant can destabilise:
An upward parabola in with vertex at dips below zero iff
This forces a sign pattern. Since and , one diagonal entry is positive, say : species 1 activates itself. Then : the inhibitor must diffuse faster. The picture is local activation with long-range inhibition.
Turing instability of the Brusselator
Here , activates when , , and the threshold becomes
At fixed , . Take , , , . The eigenvalues of are , a stable spiral, and
The unstable band is , wavelengths to . The threshold is , , and the critical ratio at is . The fastest mode (found numerically) has , , wavelength . Sim and numpy agree on all of these.
| band in | fastest | wavelength | ||
|---|---|---|---|---|
| to | ||||
| to | ||||
| to | ||||
| to |
The band widens (about , , in , roughly like ) around a nearly fixed centre, so the fastest wavelength hardly moves while grows about twelvefold. And the wavelength is set by chemistry and diffusion, not by the domain: , so quadrupling stretches the pattern only by .
A pattern from noise: linear against nonlinear
Linear theory says which modes grow, not where growth stops. The sim runs the nonlinear Brusselator on , ( nodes), no-flux ends, , from the uniform state plus noise of size , explicit to . Each run settles into regular stationary peaks of , between about and for seed 1. I measured the mean spacing of the interior peaks for seeds and recomputed each with scipy’s LSODA from identical noise (the generator, mulberry32, is reproduced in Python): the spacings agree to three decimals.
| spacing | seeds | allowed mode , wavelength |
|---|---|---|
| of | : | |
| of | : | |
| of | : |
All twelve land inside the band and within of the linear optimum ; the mean is . They do not coincide with it. The no-flux ends allow only , but the nearest allowed wavelength (, ) is close, so that alone does not explain the spread. The competition is close: the linear growth rates are (), () and (), within , and the nonlinear terms decide. The linear winner, , was the outcome only times in . Linear theory gives the scale of the pattern, not an oracle for its wavelength.
Sensitivity is the flip side. At , for every : no pattern from any noise. At the band exists but is . A pattern-former lives within a factor of two of a boundary where it forms nothing.
Computing it: the explicit limit and IMEX
Explicit FTCS needs the highest grid mode, the checkerboard with , inside the Euler stability interval (10.2 Explicit Methods for the Heat Equation). For heat alone that is ; with two diffusivities the faster binds: . Including the reaction Jacobian, the checkerboard’s most negative eigenvalue is , so . Three checks on the full scheme:
- Scipy’s eigenvalues of the full discrete Jacobian give the same .
- The sim measures the checkerboard growth factor per step by running the scheme: at , at .
- To , gives the normal pattern ( interior peaks, spacing ); blows up within about steps.
At the run takes steps. Halving quarters the limit and multiplies the work about eightfold.
IMEX. Treat the stiff diffusion implicitly (the tridiagonal solve of 10.4 Implicit Methods & Crank–Nicolson) and the mild reaction explicitly: , then . It is the operator splitting of 15.4 Chemical Kinetics & Stiff Systems in miniature, first order in . At it takes steps, times fewer, with spacing against explicit and range against (a dense-solve numpy version agrees). At the pattern is lost: no peaks, confined to . The splitting’s linear growth rate for the fastest mode is still there against the exact , so mode analysis does not explain it and I did not isolate the cause. IMEX removes the limit; another limit, found by experiment, takes over.
Gray-Scott in 2D, and why the Brusselator does the analysis
The Gray-Scott model,
is the standard demonstration: a feed and a kill rate give spots, labyrinths and splitting blobs. It is a poor model to analyse here. The empty state has Jacobian , stable for every and every diffusivity, so no noise grows: patterns start only when a finite seed of invades, like a Fisher front, and then splits. The Brusselator is the opposite: last lesson’s Jacobian, a closed-form threshold, growth from noise. So we analyse the Brusselator in 1D and play with Gray-Scott in 2D.
The sim uses , , a periodic grid of spacing , a central square seed, and . In 2D the checkerboard has , so the limit is about , and with the reaction at . Measured by running the scheme (and in numpy): checkerboard growth at , at , and blows up within steps. The default has margin.
| , | after | after |
|---|---|---|
| , | blobs of , of area | blobs, |
| , | connected structure, | structure, |
(Blobs are 4-connected with periodic wrap; sim and numpy/scipy agree exactly.) The first row is spots whose count is still changing at , so it is a snapshot; the second is one connected labyrinth. Spirals and other regimes exist in the - plane (from memory, not shown here).
What this does and does not show
Turing proposed in 1952 that morphogens obeying such equations could pattern an embryo. The mathematics is exact: the band, threshold and wavelength are facts about a PDE. Whether a given animal is patterned this way is a hypothesis to be tested against biology, not against the PDE. Support is mixed (from memory): some systems have been argued to fit and others, leopard spots among them, remain interpretations, and real tissue has growth, curvature and more than two species. Our own numbers show three limits. The diffusivity ratio must exceed a threshold ( here), hard to arrange when small molecules diffuse at similar speeds (from memory). The wavelength depends on diffusivities only through a fourth root. And noise and geometry change the outcome, as the twelve seeds showed. Nothing here was fitted to an animal.
Where this goes next
So far we have a front (Fisher) and a stationary pattern (Turing). A neuron is different: its membrane rests stably, but a large enough kick sends a stereotyped pulse down the axon, after which it recovers. That is an excitable medium, again , with a fast voltage variable and a slow recovery variable that, unlike the Turing inhibitor, need not diffuse at all. The Hodgkin-Huxley and FitzHugh-Nagumo models of 15.6 Neuron Models supply , and the front becomes a travelling pulse. Whether anything like survives there is the first question to ask; it should not be assumed.
The exercise: Laplacian, Fisher speed, dispersion relation and step limit
The playground first grows the 1D Brusselator pattern with explicit steps and tests the step limit.
Now build the pieces yourself: the no-flux Laplacian, the Fisher front speed, the dispersion relation, the unstable band and the explicit limit.
Concept checks
Problems
A species spreads with and (consistent units). Find the asymptotic front speed and the time to cover length units. Is the real time shorter or longer?
Show worked solution
, and . The real time is longer: the measured speed approaches from below, as in the table above.
A reaction has at its uniform state, with . (a) Confirm it is stable without diffusion. (b) Find the smallest for a Turing instability. (c) Find there.
Show worked solution
(a) , . (b) The condition is . With : , so and . (A numerical scan finds an unstable mode at and none at .) (c) .
For the Brusselator with , , , on , find the explicit step limit and the steps to at and , and the ratio of work (steps times nodes).
Show worked solution
The limit is for : at ( steps, nodes) and at ( steps, nodes). The work is against , a ratio of , close to : steps , nodes .
For the tail gives . (a) Show , with equality only at . (b) Predict the speed for data decaying like and like . (c) The sim gave and . Do these support (b)?
Show worked solution
(a) . (b) For the tail carries the front at ; for the front selects the minimum, , not (both from memory). (c) Yes: matches , and is near (below it, as in the table) and far from . One window is evidence, not proof.
Key Takeaways
- Adding to a reaction gives a PDE; one diffusing species only smooths, but two with unequal diffusivities can create structure.
- A Fisher-KPP front is pulled by its linear tail: is minimal at , giving ; measured speeds are to below it at to and reach near .
- A mode grows at the top eigenvalue of ; only can destabilise a state stable without diffusion.
- Turing needs an activator and a faster inhibitor: for the Brusselator, band and fastest wavelength at .
- The nonlinear wavelength lands in the band and near, not at, the fastest mode: twelve seeds gave , or , at most away.
- Explicit schemes need of the grid checkerboard, about : here, measured as growth factors and .
- IMEX took with fewer steps and the same pattern, but failed at for a reason I did not isolate.
- Gray-Scott patterns need a finite seed because the empty state is always stable, and have long transients ( then spots).
- Turing’s mechanism as the origin of real animal markings is a hypothesis; the mathematics is exact, the biology unsettled.