Differential equations and dynamical systems

Chaos theory

Deterministic systems that are extremely sensitive to initial conditions, producing unpredictable long-term behavior.

IntuitionWhy can't we predict the weather two weeks out?

Weather forecasting uses the same physical laws — Newton's equations, thermodynamics — whether we predict tomorrow or next month, yet forecasts become useless after about ten days. The reason is not that the equations are wrong or the computers too slow: the atmosphere is chaotic. A butterfly flapping its wings in Brazil, changing the air temperature by a fraction of a degree, can in principle change whether a hurricane forms in the Atlantic weeks later. This is not mysticism; it is a precise mathematical property called sensitive dependence on initial conditions, discovered independently by Henri Poincaré while studying the three-body problem in the 1890s and rediscovered by the meteorologist Edward Lorenz in 1963 while rounding a computer printout from six decimal places to three.

Cobweb diagram of the logistic map

x_{n+1} = r x_n(1 - x_n)

: increase the control parameter

r

from

2.5

(stable fixed point) through

3.2

(period-

2

cycle) up to

3.8

(chaos).

UndergraduateSensitive dependence and the Lorenz system

Definition: Sensitive dependence on initial conditions

A dynamical system has sensitive dependence on initial conditions if two trajectories starting from nearby points, a distance $\delta_0$ apart, separate at an exponential rate $|\delta_n| \approx |\delta_0|\,e^{\lambda n}$ , where $\lambda$ is the Lyapunov exponent. When $\lambda>0$ , even a minuscule measurement error $\delta_0$ grows to a macroscopic, forecast-ruining size after a finite time — exactly why weather forecasts break down after about ten days.

\begin{aligned} \dot x &= \sigma(y-x) \\ \dot y &= x(\rho - z) - y \\ \dot z &= xy - \beta z \end{aligned}

This is the Lorenz system, a drastic simplification of atmospheric convection: $x$ is proportional to the convective roll velocity, $y$ to the temperature difference between rising and falling currents, and $z$ to the distortion of the vertical temperature profile from linear. The parameter $\sigma$ is the Prandtl number, $\rho$ is proportional to the Rayleigh number (how strongly the fluid is heated from below), and $\beta$ is a geometric factor. Lorenz's classic chaotic choice is $\sigma=10$ , $\rho=28$ , $\beta=8/3$ — for these values every trajectory eventually settles onto the same butterfly-shaped fractal set (the Lorenz attractor), yet which wing it is on at any given time is essentially unpredictable far in advance.

x_{n+1} = r\,x_n(1-x_n)

A far simpler system shows the same phenomenon: the logistic map $x_{n+1} = r\,x_n(1-x_n)$ , a one-dimensional model of population growth with limited resources (here $x_n\in[0,1]$ is the population fraction of its maximum, and $r$ controls the growth rate). For small $r$ every orbit settles to a single fixed point; as $r$ increases the fixed point loses stability and is replaced by a stable 2-cycle, then a 4-cycle, then an 8-cycle — a cascade of period-doubling bifurcations that accumulates at $r\approx3.5699$ (the Feigenbaum point), beyond which the orbit typically never repeats at all. Remarkably, the rate at which the bifurcation points bunch up, $\delta\approx4.6692$ (the Feigenbaum constant), is the same for an enormous class of unrelated one-dimensional maps — a rare instance of a truly universal constant in nonlinear dynamics.

Long-term behavior of the logistic map as r increases
Parameter	Long-term behavior
$r<3$	Single stable fixed point
$r=3.2$	Stable period-2 cycle
$r=3.5$	Stable period-4 cycle
$r=3.57$	Onset of chaos (accumulation of period doublings)
$r=4$	Fully chaotic; Lyapunov exponent $\ln 2$ > 0

UndergraduateQuantifying chaos: the Lyapunov exponent

\lambda = \lim_{n\to\infty}\frac{1}{n}\sum_{i=0}^{n-1}\ln\left|f'(x_i)\right|

The Lyapunov exponent $\lambda$ is the average, along a typical orbit $x_0,x_1,x_2,\dots$ , of the logarithm of the local stretching factor $|f'(x_i)|$ at each step: it measures how fast nearby points separate on average, in the same way an interest rate measures average exponential growth. $\lambda>0$ means neighboring points diverge exponentially — the hallmark of chaos; $\lambda<0$ means they converge, as at a stable fixed point; $\lambda=0$ is the borderline (periodic or quasi-periodic) case.

Theorem: Period-doubling of the logistic map at r = 3

For the logistic map $f(x)=rx(1-x)$ , the nontrivial fixed point $x^* = 1-\tfrac{1}{r}$ is stable for $1<r<3$ and loses stability exactly at $r=3$ , where it undergoes a period-doubling bifurcation into a stable 2-cycle.

Why is it true?

The derivative of the map at a fixed point measures how a tiny perturbation there grows or shrinks after one step; when its magnitude crosses 1 the point flips from attracting to repelling, and since the map's second iterate has slope $(-1)^2=1$ exactly at that threshold, a new period-2 orbit is precisely what can bifurcate off.

Proof

Step 1 (find the fixed points). Solve $rx(1-x)=x$ , i.e. $x[r(1-x)-1]=0$ . This gives $x=0$ or the nontrivial fixed point $x^* = 1-\tfrac{1}{r}$ , which lies in $(0,1)$ precisely when $r>1$ .

Step 2 (linearize). Differentiating, $f'(x)=r-2rx$ . Evaluating at the nontrivial fixed point, $f'(x^*) = r - 2r\left(1-\tfrac1r\right) = r - 2r + 2 = 2-r$ , so $f'(x^*) = 2-r$ .

Step 3 (stability criterion). A fixed point is locally stable exactly when $|f'(x^*)|<1$ , i.e. $|2-r|<1$ , which rearranges to $1<r<3$ . So $x^*$ attracts nearby orbits throughout this whole range.

Step 4 (the bifurcation at r = 3). At $r=3$ exactly, $f'(x^*)=-1$ : the linearization is marginal. For $r$ slightly above 3, $|f'(x^*)|>1$ and $x^*$ becomes repelling. Meanwhile the second-iterate map $f\circ f$ has, by the chain rule, derivative $f'(x^*)^2=1$ at $x^*$ exactly at $r=3$ ; expanding $f\circ f$ to higher order shows this degenerate point splits into two genuine new fixed points of $f\circ f$ (a stable 2-cycle of $f$ ) as $r$ increases past 3 — the period-doubling bifurcation.

Theorem: The Lyapunov exponent of the fully chaotic logistic map equals ln 2

At $r=4$ , the logistic map $x_{n+1} = r\,x_n(1-x_n)$ has Lyapunov exponent $\ln 2$ for Lebesgue-almost every initial condition $x_0\in(0,1)$ .

Why is it true?

The map at r = 4 looks nothing like a simple doubling map at first glance, but a clever change of variables (a smooth conjugacy) turns it into exactly the doubling map, whose stretching rate is obviously 2 at every point — the extra distortion introduced by the change of variables averages out to nothing over long times.

Proof

Step 1 (the conjugacy). Substitute $x=\sin^2(\pi y)$ . Using the double-angle identity, $4x(1-x)=4\sin^2(\pi y)\cos^2(\pi y)=\sin^2(2\pi y)$ . If we also set $y' = T(y)$ for the doubling map $T(y)=2y \bmod 1$ , then $\sin^2(\pi y') = \sin^2(2\pi y)$ matches exactly the right-hand side above, so $f(h(y)) = h(T(y))$ with $h(y)=\sin^2(\pi y)$ : the logistic map at r = 4 is smoothly conjugate to the doubling map.

Step 2 (Lyapunov exponent of the doubling map). The doubling map is piecewise linear with $T'(y)=2$ everywhere it is differentiable, so along every orbit $\tfrac1n\sum_{i=0}^{n-1}\ln|T'(y_i)| = \ln 2$ exactly, for every $n$ — the limit is trivially $\ln 2$ .

Step 3 (transporting the exponent through the conjugacy). Differentiating $f(h(y))=h(T(y))$ by the chain rule gives $f'(h(y))h'(y)=h'(T(y))T'(y)$ , i.e. $\ln|f'(x)| = \ln|h'(T(y))| + \ln|T'(y)| - \ln|h'(y)|$ at $x=h(y)$ . Summing this identity along an orbit $y_0,y_1,\dots,y_{n-1}$ and dividing by $n$ , the middle terms telescope: $\tfrac1n\sum_{i=0}^{n-1}\ln|f'(x_i)| = \ln2+\tfrac1n\big[\ln|h'(y_n)|-\ln|h'(y_0)|\big]$ .

Step 4 (the boundary term vanishes). Since $h'(y)=\pi\sin(2\pi y)$ is bounded (its magnitude never exceeds $\pi$ ), the bracketed term $\ln|h'(y_n)|-\ln|h'(y_0)|$ stays bounded as $n\to\infty$ for any orbit that avoids the countably many zeros of $h'$ — a set of Lebesgue measure zero. Dividing a bounded quantity by $n$ sends it to $0$ , so the Lyapunov exponent of the logistic map at r = 4 equals $\ln 2$ for almost every initial condition.

UndergraduateReal-World Applications and Worked Examples

Chaos theory is not just about weather: it explains why heart-rhythm monitors watch for the onset of chaotic fibrillation, why secure chaotic communication systems mix a message into a chaotic carrier signal that only a receiver with the exact same parameters can decode, why population-ecology models with a high reproduction rate $r$ can swing wildly even with no external randomness, and why engineers design electronic circuits (such as Chua's circuit) specifically to generate chaotic signals for random-number generation and encryption. In every case the same two computations recur: locate and linearize the equilibria or periodic orbits, and estimate the Lyapunov exponent that tells you how fast a small error explodes.

Example: A population that oscillates forever: the period-2 cycle at r = 3.2

A fishery models next year's stock fraction by the logistic map $x_{n+1} = r\,x_n(1-x_n)$ with growth parameter $r=3.2$ (above the period-doubling threshold $r=3$ ). Find the two population levels the stock alternates between, once transients die out.

Solution

Points of a period-2 cycle satisfy $f(f(x))=x$ without being fixed points themselves, so $f(f(x))-x$ must vanish. Expanding $f(f(x))-x$ and dividing out the two roots that are the ordinary fixed points $x=0$ and $x^*=1-1/r$ (which already solve $f(x)=x$ , hence trivially $f(f(x))=x$ ) leaves the genuinely new factor $r^2x^2-r(r+1)x+(r+1)=0$ .

This is a quadratic in $x$ . By the quadratic formula, $x=\dfrac{(r+1)\pm\sqrt{(r-3)(r+1)}}{2r}$ .

Plugging in $r=3.2$ : the numerator terms are $r+1=4.2$ and $(r-3)(r+1)=0.2\times4.2=0.84$ , so $\sqrt{0.84}\approx0.9165$ and $2r=6.4$ . This gives $x\approx\{0.513,\ 0.799\}$ .

So the fishery's stock settles into an oscillation, alternately about $51.3\%$ and $80.0\%$ of carrying capacity every other year — a genuinely periodic population cycle produced by a completely deterministic rule, well before the system becomes chaotic.

Example: Why ten days? Estimating the weather's forecast horizon

Atmospheric models have an estimated Lyapunov exponent of about $\lambda\approx0.9$ per day. Two forecast runs start $\delta_0=10^{-6}$ (in normalized units) apart, representing an essentially perfect initial measurement. Estimate how large this tiny difference has grown after $n=10$ days, and comment on what it means for forecasting.

Solution

By the definition of the Lyapunov exponent, small errors grow (on average) like $\delta_n \approx \delta_0\,e^{\lambda n}$ .

With $\lambda\approx0.9$ per day and $n=10$ days, the exponent is $\lambda n \approx 0.9\times10=9$ , so the growth factor is $e^{9}\approx8103$ .

Starting from $\delta_0=10^{-6}$ , after ten days the error has grown to $\delta_{10}\approx10^{-6}\times8103\approx8.1\times10^{-3}$ — a roughly eight-thousand-fold amplification of an initially minuscule, essentially unmeasurable difference.

Since real measurement and modeling errors are already far larger than $10^{-6}$ in relative terms, this exponential blow-up is exactly why operational forecasts, however good the physics and however powerful the computer, lose all skill within about one to two weeks: it is a mathematical horizon set by $\lambda$ , not an engineering limitation.

Historical note

In 1890, Henri Poincaré, competing for King Oscar II of Sweden's prize on the stability of the solar system, discovered while studying the restricted three-body problem that the equations could produce trajectories so intricately tangled that long-term prediction was hopeless in practice — the first mathematical glimpse of what would later be called chaos. The idea lay largely dormant until 1963, when the meteorologist Edward Lorenz, re-running a 12-variable weather model from a mid-point printout rounded to three decimal places instead of the computer's internal six, found the new run diverged completely from the original within a simulated month; his 1963 paper "Deterministic Nonperiodic Flow" introduced the three-equation Lorenz system and the butterfly-effect metaphor. In 1975, the biologist-turned-mathematician Robert May's survey "Simple mathematical models with very complicated dynamics" showed that the humble logistic map already contains the full period-doubling route to chaos, bringing the subject to a wide scientific audience and igniting the modern field of nonlinear dynamics.

Henri Poincaré

ResearchOpen frontier: rigorous chaos and high-dimensional turbulence

Research frontier as of 2026

Proving that a specific computer-simulated attractor is genuinely chaotic — rather than an extremely long but ultimately periodic numerical artifact — is surprisingly hard, because floating-point round-off is itself a perturbation that a positive Lyapunov exponent will amplify. The Lorenz attractor's existence as a robust, genuinely chaotic (strange) attractor was only proved rigorously in 2002 by Warwick Tucker, using interval arithmetic to track rigorous enclosures of the flow rather than single floating-point trajectories — a computer-assisted proof settling the 14th of Steve Smale's list of problems for the 21st century. Rigorous computer-assisted methods of this kind are now an active research area for higher-dimensional and infinite-dimensional systems: whether the Navier–Stokes equations governing turbulent fluid flow possess a well-defined chaotic attractor with finite dimension is still open in full generality, whether the Palis conjecture (that "most" dynamical systems are either hyperbolic or admit only finitely many attractors) holds is unresolved outside special cases, and machine-learning-based surrogate models for predicting the onset of turbulence and rigorously bounding Lyapunov exponents in large coupled systems (from power grids to neural population models) are a fast-growing subfield as of 2026.

For the logistic map $f(x)=rx(1-x)$ with nontrivial fixed point $x^* = 1-\tfrac{1}{r}$ , what is the derivative $f'(x^*)$ in terms of $r$ ?

Two weather-forecast runs start $\delta_0=10^{-6}$ (normalized units) apart. Using $\delta_n\approx\delta_0 e^{\lambda n}$ with $\lambda\approx0.9$ per day, by roughly what factor has the separation grown after $n=10$ days?

At which parameter value $r$ does the logistic map's nontrivial fixed point $x^* = 1-\tfrac{1}{r}$ first lose stability via period-doubling?

A dynamical system has Lyapunov exponent $\lambda>0$ . What does this indicate?

References

Steven H. Strogatz (2015). Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering
Edward N. Lorenz (1963). Deterministic Nonperiodic Flow
Robert M. May (1976). Simple mathematical models with very complicated dynamics
Warwick Tucker (2002). A Rigorous ODE Solver and Smale's 14th Problem