Recommended:Britton (2003), Chapter 7, Sections 7.2 to 7.4.1 (pp. 206 to 217), for the derivation with spatial modes, and Murray (2003), Chapter 2, Sections 2.1 to 2.6.
Turing (1952) showed that diffusion can destabilize a uniform state that is stable without diffusion. This is called diffusion-driven or Turing instability. This page derives the four conditions for it in the notation of Murray (2003, sec. 2.3).
The Question
Take the linearized system from Laplacian Eigenmodes. Each mode with wavenumber \(k\) grows like \(e^{\sigma t}\), where \(\sigma\) solves
\[
\det\left(\gamma J - k^2 D - \sigma I\right) = 0, \qquad D = \operatorname{diag}(1, d).
\]
You want the uniform state to be:
stable to the flat mode \(k = 0\) (stable without diffusion), and
Stability at \(k = 0\) requires \(\operatorname{tr} J < 0\) and \(\det J > 0\). Then \(\operatorname{tr} A < 0\) for every \(k\), so the only way to get \(\operatorname{Re}\sigma > 0\) is \(h(k^2) < 0\) for some \(k^2 > 0\), which gives one real positive \(\sigma\).
\(h\) is a quadratic in \(k^2\) with positive constant term. It becomes negative only if its minimum lies at \(k^2 > 0\) and is negative. The minimum is at
All derivatives are evaluated at the steady state. The conditions do not contain \(\gamma\) or the domain size. They are necessary and sufficient for the linearized problem to have a band of growing wavenumbers \(k > 0\) while the flat mode is stable. They are not necessary for patterns in general: the Gray-Scott patterns of Gray-Scott Parameters grow from finite-amplitude perturbations. On a finite domain you also need an admissible wavenumber inside the unstable band (see Dispersion Relation and Domain Size and Mode Selection).
Britton’s Notation
Britton (2003, sec. 7.4.1) derives the same conditions with diffusion coefficients \(D_1\), \(D_2\) and the scale \(\alpha\) (his version of \(\gamma\)), and with the spatial eigenvalue \(\lambda = k^2\). His mode matrix is \(\alpha J - \lambda D\), and he writes its characteristic polynomial as \(\sigma^2 + a_1(\lambda)\sigma + a_2(\lambda) = 0\) with
So \(-a_1\) and \(a_2\) are the trace and determinant of the mode matrix. With \(D_1 = 1\), \(D_2 = d\) and \(\alpha = \gamma\) they are \(-\operatorname{tr} A\) and \(h(k^2)\) of Equation 1. Both roots have negative real part if and only if \(a_1 > 0\) and \(a_2 > 0\). Requiring this at \(\lambda = 0\) gives \(\operatorname{tr} J < 0\) and \(\det J > 0\), which also makes \(a_1(\lambda) > 0\) for every \(\lambda\). Hence a Hopf (oscillatory) instability with \(k \ne 0\) cannot occur in a system of two equations, and the only way to lose stability is \(a_2 < 0\). The remaining two inequalities of Equation 3 are his conditions \(D_2 f_u + D_1 g_v > 0\) and \((D_2 f_u + D_1 g_v)^2 > 4 D_1 D_2 \det J\), which are Equation 3 with \(D_1 = 1\), \(D_2 = d\). The critical wavenumber on an infinite domain is \(\lambda_c = \tfrac{\alpha}{2}\left(f_u/D_1 + g_v/D_2\right)\), which equals \(k_m^2\) in Equation 2. Turing’s own symbols \(a, b, c, d, \mu, \nu\) are compared with both on Turing’s Original Analysis.
What the Signs Tell You
The first and third conditions force \(d \neq 1\). If both species diffuse equally, no Turing instability exists.
They also force \(f_u\) and \(g_v\) to have opposite signs: the trace condition excludes two positive entries, and two negative entries would make \(d f_u + g_v\) negative. Call the species with \(f_u > 0\) the activator. Then \(g_v < 0\) and \(d > -g_v / f_u > 1\): the inhibitor must diffuse faster than the activator.
The mirror case is equally possible. If \(f_u < 0\) then \(g_v > 0\), the species \(v\) is the activator, and \(d < -g_v/f_u < 1\): the activator again diffuses more slowly. Turing’s own numerical example is of this type, with his self-activating morphogen \(Y\) diffusing more slowly than \(X\) (Turing’s Original Analysis).
With \(f_u g_v < 0\), the second condition needs \(f_v g_u < 0\). Either the activator promotes the inhibitor and the inhibitor suppresses the activator (\(g_u > 0\), \(f_v < 0\)), or the opposite cross signs hold (Murray 2003, sec. 2.3).
The usual summary is short-range activation with long-range inhibition (Britton 2003, chap. 7 summary and Section 7.1; Kondo and Miura 2010). In this model the longer range of the inhibitor comes from its faster diffusion. Kondo and Miura (2010) point out that the range can also be produced by other forms of signalling, so \(d > 1\) is the diffusive realisation of the requirement and not the requirement itself.
The Critical Diffusion Ratio
At the onset \(h_{\min} = 0\), so the critical ratio \(d_c\) solves
taking the root with \(d_c f_u + g_v > 0\)(Murray 2003, sec. 2.3). For \(d > d_c\) a band of wavenumbers is unstable.
Worked Example: Gierer-Meinhardt
From Linearization, \(f_u = b(1-a)/(1+a)\), \(f_v = -b^2/(1+a)^2\), \(g_u = 2(1+a)/b\), \(g_v = -1\), and \(\det J = b\). So \(f_u > 0\) needs \(a < 1\), and \(u\) is the activator. With \(a = 0.4\) and \(b = 1\):
So \(d = 20\) gives no Turing instability and \(d = 30\) does. You will test this prediction in the laboratory.
NoteExercise
For Gierer-Meinhardt with \(b = 1\), show that the fourth condition reads \(d f_u - 1 > 2\sqrt{d}\), and solve it for \(\sqrt{d}\).
TipSolution
With \(g_v = -1\) and \(\det J = b = 1\), the third and fourth conditions together say \(d f_u - 1 > 2\sqrt{d}\). This is a quadratic inequality in \(s = \sqrt d\): \(f_u s^2 - 2s - 1 > 0\). For \(f_u > 0\) it holds for \(s > \left(1 + \sqrt{1 + f_u}\right)/f_u\). With \(a = 0.4\), \(f_u = 3/7\) and \(s > 5.122\), so \(d > 26.24\).
Kondo, Shigeru, and Takashi Miura. 2010. “Reaction-Diffusion Model as a Framework for Understanding Biological Pattern Formation.”Science 329 (5999): 1616–20. https://doi.org/10.1126/science.1179047.
Murray, James D. 2003. Mathematical Biology II: Spatial Models and Biomedical Applications. 3rd ed. Vol. 18. Interdisciplinary Applied Mathematics. Springer. https://doi.org/10.1007/b98869.
Turing, Alan M. 1952. “The Chemical Basis of Morphogenesis.”Philosophical Transactions of the Royal Society B 237 (641): 37–72. https://doi.org/10.1098/rstb.1952.0012.
---title: "Turing Conditions"subtitle: "Module 4: Space and Pattern Formation"format: html---::: {.callout-note}## Reading- **Compulsory:** @turing1952chemical, Sections 6 to 9, read together with the guide on [Turing's Original Analysis](turing-original-analysis.qmd).- **Recommended:** @britton2003essential, Chapter 7, Sections 7.2 to 7.4.1 (pp. 206 to 217), for the derivation with spatial modes, and @murray2003mathematical, Chapter 2, Sections 2.1 to 2.6.:::@turing1952chemical showed that diffusion can destabilize a uniform state that is stable without diffusion. This is called diffusion-driven or Turing instability. This page derives the four conditions for it in the notation of @murray2003mathematical [Section 2.3].## The QuestionTake the linearized system from [Laplacian Eigenmodes](laplacian-eigenmodes.qmd). Each mode with wavenumber $k$ grows like $e^{\sigma t}$, where $\sigma$ solves$$\det\left(\gamma J - k^2 D - \sigma I\right) = 0, \qquad D = \operatorname{diag}(1, d).$$You want the uniform state to be:- stable to the flat mode $k = 0$ (stable without diffusion), and- unstable to at least one mode with $k \neq 0$.## Trace and Determinant of the Mode MatrixFor $A(k^2) = \gamma J - k^2 D$,$$\operatorname{tr} A = \gamma (f_u + g_v) - (1 + d)k^2,$$$$h(k^2) \equiv \det A = d\,k^4 - \gamma\,(d f_u + g_v)\,k^2 + \gamma^2 (f_u g_v - f_v g_u).$$ {#eq-tc-h}Stability at $k = 0$ requires $\operatorname{tr} J < 0$ and $\det J > 0$. Then $\operatorname{tr} A < 0$ for every $k$, so the only way to get $\operatorname{Re}\sigma > 0$ is $h(k^2) < 0$ for some $k^2 > 0$, which gives one real positive $\sigma$.$h$ is a quadratic in $k^2$ with positive constant term. It becomes negative only if its minimum lies at $k^2 > 0$ and is negative. The minimum is at$$k_m^2 = \frac{\gamma\,(d f_u + g_v)}{2d}, \qquadh_{\min} = \gamma^2\left[\det J - \frac{(d f_u + g_v)^2}{4d}\right].$$ {#eq-tc-hmin}## The Four ConditionsCollecting the requirements gives the Turing conditions [@murray2003mathematical, Section 2.3]:$$\begin{aligned}&f_u + g_v < 0, \\&f_u g_v - f_v g_u > 0, \\&d f_u + g_v > 0, \\&(d f_u + g_v)^2 - 4 d\,(f_u g_v - f_v g_u) > 0.\end{aligned}$$ {#eq-tc-conditions}All derivatives are evaluated at the steady state. The conditions do not contain $\gamma$ or the domain size. They are necessary and sufficient for the linearized problem to have a band of growing wavenumbers $k > 0$ while the flat mode is stable. They are not necessary for patterns in general: the Gray-Scott patterns of [Gray-Scott Parameters](gray-scott-parameters.qmd) grow from finite-amplitude perturbations. On a finite domain you also need an admissible wavenumber inside the unstable band (see [Dispersion Relation](dispersion-relation.qmd) and [Domain Size and Mode Selection](domain-size-mode-selection.qmd)).## Britton's Notation@britton2003essential [Section 7.4.1] derives the same conditions with diffusion coefficients $D_1$, $D_2$ and the scale $\alpha$ (his version of $\gamma$), and with the spatial eigenvalue $\lambda = k^2$. His mode matrix is $\alpha J - \lambda D$, and he writes its characteristic polynomial as $\sigma^2 + a_1(\lambda)\sigma + a_2(\lambda) = 0$ with$$\begin{aligned}a_1(\lambda) &= (D_1 + D_2)\lambda - \alpha (f_u + g_v), \\a_2(\lambda) &= D_1 D_2 \lambda^2 - \alpha (D_2 f_u + D_1 g_v)\lambda + \alpha^2 \det J.\end{aligned}$$ {#eq-tc-britton}So $-a_1$ and $a_2$ are the trace and determinant of the mode matrix. With $D_1 = 1$, $D_2 = d$ and $\alpha = \gamma$ they are $-\operatorname{tr} A$ and $h(k^2)$ of @eq-tc-h. Both roots have negative real part if and only if $a_1 > 0$ and $a_2 > 0$. Requiring this at $\lambda = 0$ gives $\operatorname{tr} J < 0$ and $\det J > 0$, which also makes $a_1(\lambda) > 0$ for every $\lambda$. Hence a Hopf (oscillatory) instability with $k \ne 0$ cannot occur in a system of two equations, and the only way to lose stability is $a_2 < 0$. The remaining two inequalities of @eq-tc-conditions are his conditions $D_2 f_u + D_1 g_v > 0$ and $(D_2 f_u + D_1 g_v)^2 > 4 D_1 D_2 \det J$, which are @eq-tc-conditions with $D_1 = 1$, $D_2 = d$. The critical wavenumber on an infinite domain is $\lambda_c = \tfrac{\alpha}{2}\left(f_u/D_1 + g_v/D_2\right)$, which equals $k_m^2$ in @eq-tc-hmin [@britton2003essential, Equations 7.4.27 to 7.4.30]. Turing's own symbols $a, b, c, d, \mu, \nu$ are compared with both on [Turing's Original Analysis](turing-original-analysis.qmd).## What the Signs Tell You- The first and third conditions force $d \neq 1$. If both species diffuse equally, no Turing instability exists.- They also force $f_u$ and $g_v$ to have opposite signs: the trace condition excludes two positive entries, and two negative entries would make $d f_u + g_v$ negative. Call the species with $f_u > 0$ the activator. Then $g_v < 0$ and $d > -g_v / f_u > 1$: the inhibitor must diffuse faster than the activator.- The mirror case is equally possible. If $f_u < 0$ then $g_v > 0$, the species $v$ is the activator, and $d < -g_v/f_u < 1$: the activator again diffuses more slowly. Turing's own numerical example is of this type, with his self-activating morphogen $Y$ diffusing more slowly than $X$ ([Turing's Original Analysis](turing-original-analysis.qmd)).- With $f_u g_v < 0$, the second condition needs $f_v g_u < 0$. Either the activator promotes the inhibitor and the inhibitor suppresses the activator ($g_u > 0$, $f_v < 0$), or the opposite cross signs hold [@murray2003mathematical, Section 2.3].The usual summary is short-range activation with long-range inhibition [@britton2003essential, Chapter 7 summary and Section 7.1; @kondo2010reaction]. In this model the longer range of the inhibitor comes from its faster diffusion. @kondo2010reaction point out that the range can also be produced by other forms of signalling, so $d > 1$ is the diffusive realisation of the requirement and not the requirement itself.## The Critical Diffusion RatioAt the onset $h_{\min} = 0$, so the critical ratio $d_c$ solves$$f_u^2\,d_c^2 + 2\,(2 f_v g_u - f_u g_v)\,d_c + g_v^2 = 0,$$ {#eq-tc-dc}taking the root with $d_c f_u + g_v > 0$ [@murray2003mathematical, Section 2.3]. For $d > d_c$ a band of wavenumbers is unstable.## Worked Example: Gierer-MeinhardtFrom [Linearization](linearization.qmd), $f_u = b(1-a)/(1+a)$, $f_v = -b^2/(1+a)^2$, $g_u = 2(1+a)/b$, $g_v = -1$, and $\det J = b$. So $f_u > 0$ needs $a < 1$, and $u$ is the activator. With $a = 0.4$ and $b = 1$:```{python}from amlab.turing import critical_diffusion_ratio, gierer_meinhardt_jacobian, turing_conditionsfu, fv, gu, gv = gierer_meinhardt_jacobian(a=0.40, b=1.00)print(f"f_u = {fu:.4f}, f_v = {fv:.4f}, g_u = {gu:.4f}, g_v = {gv:.4f}")print(f"d_c = {critical_diffusion_ratio(fu, fv, gu, gv):.2f}")for d in (20, 30):print(f"d = {d}: Turing conditions hold? {bool(turing_conditions(fu, fv, gu, gv, d))}")```So $d = 20$ gives no Turing instability and $d = 30$ does. You will test this prediction in the laboratory.::: {.callout-note}## ExerciseFor Gierer-Meinhardt with $b = 1$, show that the fourth condition reads $d f_u - 1 > 2\sqrt{d}$, and solve it for $\sqrt{d}$.:::::: {.callout-tip collapse="true"}## SolutionWith $g_v = -1$ and $\det J = b = 1$, the third and fourth conditions together say $d f_u - 1 > 2\sqrt{d}$. This is a quadratic inequality in $s = \sqrt d$: $f_u s^2 - 2s - 1 > 0$. For $f_u > 0$ it holds for $s > \left(1 + \sqrt{1 + f_u}\right)/f_u$. With $a = 0.4$, $f_u = 3/7$ and $s > 5.122$, so $d > 26.24$.:::