---
title: 'Normal Modes: Why Coupled Things Beat, and When They Become Chaos'
date: '2026-06-17T00:00:00.000Z'
description: >-
Two pendulums linked by a spring look hopelessly complicated. But there is a
change of coordinates in which the same system is just two independent
pendulums that never talk to each other. The mess is an illusion of the basis
you chose.
labels: 'physics,mechanics,normal modes,chaos'
release: true
author: Ben Ebsworth
heroImage: /blog/normal-modes-to-chaos/hero.webp
markdown_url: /blog/normal-modes-to-chaos/
canonical_url: 'https://benebsworth.com/blog/normal-modes-to-chaos/'
---
Take two identical pendulums and tie a weak spring between them. Pull one aside, let go, and watch. The first pendulum swings, slows, and falls still — while the second, untouched, somehow winds itself up to full amplitude. Then the second donates everything back, and the first swells again. Energy sloshes between them on and on, a slow heartbeat riding the fast swing, and the motion of either pendulum on its own looks like nothing you could write a formula for. Two coupled masses, and already the picture seems hopeless.
That hopelessness is a lie about the coordinates, not a fact about the system. There exists a change of variables — a rotation of your axes, nothing more — in which the *same* two pendulums become two completely independent oscillators that never exchange a single joule. In those coordinates each one swings at one fixed frequency forever, clean as a metronome, and the sloshing simply does not happen. The beat you saw was never in the physics. It was an artifact of watching the system in the wrong basis. Find the right basis and the coupling evaporates.
These privileged coordinates are the **normal modes**, and the whole apparatus of small oscillations — molecular spectra, bridge resonances, the phonons in a crystal — is the claim that *any* system of coupled oscillators becomes a set of independent ones in the right coordinates. Then we break it. Add one more pendulum on a single arm and the trick collapses, because normal modes only live near equilibrium. Beyond that small neighbourhood lies the double pendulum, and the double pendulum is chaos. One post spans the most-solvable and least-solvable systems in mechanics, separated by a single act of linearisation.
> [LabSide component] Side-by-side lab layout: the same interactive lab effect as LabCanvas (referenced by its `effect` slug) rendered in one column with the post's prose (`children`) beside it, stacking vertically on mobile. `reverse` swaps the columns; `params` override defaults and `controls={false}` hides the effect's controls. Used to weave explanation and visualisation together rather than dropping the lab as an isolated figure. The rendered post has the live version; this is a placeholder for the markdown-only sibling.
Start with **Initial condition** set to *Left displaced*. The left mass (teal) starts pulled aside, the right (purple) at rest. Watch the energy pour across: the left amplitude shrinks to nothing while the right grows to full, then the trade reverses. Neither mass keeps a steady amplitude — this is the "hopeless" regime.
Now switch **Initial condition** to *Normal mode 1*. Both masses start displaced the *same* way and released together. The beating vanishes. Both swing in lockstep at one unchanging frequency, the coupling spring never stretching, forever. Switch to *Normal mode 2* — masses displaced *opposite* — and again the beat is gone, just at a higher pitch. You have found the two coordinates in which the system is trivial.
## The trick: diagonalise the coupling
Write the physics down and the structure jumps out. Two masses $m$ on springs of stiffness $k$ to the walls, coupled by a spring of stiffness $k_c$. Let $x_1$ and $x_2$ be the displacements from rest. Newton's second law on each mass gives two equations, and the only thing tying them together is the coupling term $k_c(x_1 - x_2)$:
> [Equation component] Labeled display-math block (KaTeX-rendered). Wraps a `$$...$$` math expression with an optional `id` for cross-references, an explicit `number` like "(3.2)", and a short `caption` shown below in monospace muted text. The math is rendered server-side via `remark-math` + `rehype-katex` (Katex is the rendering engine, not MathJax). Use this for the *important* equations — the ones the reader should remember, the ones the post's argument hinges on. A 2,000-word post should have 3-5 numbered equations, not 30; the rest stay as inline `$...$` math in running prose. Cross-reference via `equation (1)`.
```latex
\begin{aligned} m\,\ddot{x}_1 &= -k\,x_1 - k_c\,(x_1 - x_2) \\ m\,\ddot{x}_2 &= -k\,x_2 - k_c\,(x_2 - x_1) \end{aligned}
```
$$
\begin{aligned}
m\,\ddot{x}_1 &= -k\,x_1 - k_c\,(x_1 - x_2) \\
m\,\ddot{x}_2 &= -k\,x_2 - k_c\,(x_2 - x_1)
\end{aligned}
$$
The mess lives entirely in the cross-terms. Equation 1 says the acceleration of mass 1 depends on the position of mass 2, and vice versa — the mathematical fingerprint of "they talk to each other". You cannot solve either equation without the other.
The way out is to stop tracking $x_1$ and $x_2$ and track two clever combinations instead. Define the **sum** coordinate $q_1 = x_1 + x_2$ and the **difference** coordinate $q_2 = x_1 - x_2$. Add the two lines of equation 1 and the coupling cancels; subtract them and it doubles, but still touches only $q_2$:
$$
m\,\ddot{q}_1 = -k\,q_1, \qquad m\,\ddot{q}_2 = -(k + 2k_c)\,q_2
$$
Look at what happened. Each new coordinate obeys a *single* equation mentioning only itself. No cross-terms, no sloshing. $q_1$ and $q_2$ are two independent simple harmonic oscillators that have never heard of each other. The coupling didn't disappear from the world — it disappeared from the *coordinates*. Each $q$ is a normal mode, and its frequency falls straight out:
> [Equation component] Labeled display-math block (KaTeX-rendered). Wraps a `$$...$$` math expression with an optional `id` for cross-references, an explicit `number` like "(3.2)", and a short `caption` shown below in monospace muted text. The math is rendered server-side via `remark-math` + `rehype-katex` (Katex is the rendering engine, not MathJax). Use this for the *important* equations — the ones the reader should remember, the ones the post's argument hinges on. A 2,000-word post should have 3-5 numbered equations, not 30; the rest stay as inline `$...$` math in running prose. Cross-reference via `equation (1)`.
```latex
\omega_1 = \sqrt{\frac{k}{m}} \quad \text{(in phase)}, \qquad \omega_2 = \sqrt{\frac{k + 2k_c}{m}} \quad \text{(out of phase)}
```
$$
\omega_1 = \sqrt{\frac{k}{m}} \quad \text{(in phase)}, \qquad \omega_2 = \sqrt{\frac{k + 2k_c}{m}} \quad \text{(out of phase)}
$$
The in-phase mode $\omega_1$ is the masses moving together: the coupling spring rides along at constant length, never stretches, stores no energy, and so leaves the frequency untouched — exactly an uncoupled pendulum. The out-of-phase mode $\omega_2$ is the masses moving against each other: now the coupling spring fights the motion, and the extra stiffness $2k_c$ pushes the frequency up. Higher pitch, exactly what you saw switching to *Normal mode 2* in the lab.
> [Callout component] Styled info-block component (ported from the feelingdesigner project at ~/projects/feelingdesigner). Renders a rounded card with a tinted background, a 1px left accent bar in the type-specific colour, a quarter-circle SVG in the top-left corner that visually "cuts" the corner, and a floating icon badge that sits half-off the top edge. Seven types are available, each with its own accent colour and icon: info (blue, Info icon, neutral information), warning (yellow, AlertCircle, subtle caution), success (blue, CheckCircle, positive confirmation), error (red, XCircle, something is wrong), thinking (orange, Brain, an insight or mental model), feeling (red, Heart, a subjective observation), and doing (yellow, Hammer, a practical step to take). Used in the post to highlight key insights, contrasts, and gotchas without breaking the prose flow.
Choosing $q_1 = x_1 + x_2$ and $q_2 = x_1 - x_2$ feels like a trick pulled from a hat. It is not. Cast equation 1 as a matrix problem, $m\ddot{\mathbf{x}} = -\mathbf{K}\mathbf{x}$, where $\mathbf{K}$ is the symmetric stiffness matrix carrying $k + k_c$ on its diagonal and $-k_c$ off it. The normal modes are the **eigenvectors** of $\mathbf{K}$, and the mode frequencies come from its **eigenvalues**. For two identical masses, symmetry forces the eigenvectors to be $(1, 1)$ and $(1, -1)$ — the sum and the difference. The "trick" is just diagonalising a matrix, and the modes are whatever the eigenvectors happen to be.
The coupled system is a single linear operator $\mathbf{K}$ acting on the displacement vector. Diagonalising means finding the basis in which $\mathbf{K}$ has no off-diagonal entries — and in *that* basis the equations of motion uncouple, by construction. The eigenvalue problem is:
> [Equation component] Labeled display-math block (KaTeX-rendered). Wraps a `$$...$$` math expression with an optional `id` for cross-references, an explicit `number` like "(3.2)", and a short `caption` shown below in monospace muted text. The math is rendered server-side via `remark-math` + `rehype-katex` (Katex is the rendering engine, not MathJax). Use this for the *important* equations — the ones the reader should remember, the ones the post's argument hinges on. A 2,000-word post should have 3-5 numbered equations, not 30; the rest stay as inline `$...$` math in running prose. Cross-reference via `equation (1)`.
```latex
\mathbf{K}\,\mathbf{v}_i = m\,\omega_i^2\,\mathbf{v}_i, \qquad \mathbf{K} = \begin{pmatrix} k + k_c & -k_c \\ -k_c & k + k_c \end{pmatrix}
```
$$
\mathbf{K}\,\mathbf{v}_i = m\,\omega_i^2\,\mathbf{v}_i, \qquad \mathbf{K} = \begin{pmatrix} k + k_c & -k_c \\ -k_c & k + k_c \end{pmatrix}
$$
Every coupled-oscillator problem you will ever meet — a chain of 3 masses, a 2D lattice of 10⁶ of them, carbon dioxide with its 3 atoms and 4 vibrational modes — is the same sentence: build the stiffness matrix, find its eigenvectors, and the impossible-looking motion resolves into a sum of independent metronomes. Normal modes feel like magic only because the linear algebra has already done the work, handing you the answer in a basis where nothing is coupled.
> [PullQuote component] Editorial pull-quote. Renders a striking sentence from the surrounding prose as a large, italicised blockquote with a branded accent border. The quote text follows this placeholder verbatim, so the LLM reader still sees the highlighted sentence.
The coupling never disappears from the world. It disappears from the coordinates — and a normal mode is just the basis in which the mess was hiding.
## Beats are two modes interfering
So where did the sloshing come from, if the modes never slosh? It comes from watching a *superposition* of two modes through the original coordinates $x_1, x_2$. When you displace only the left mass, you have not excited a single mode — you have lit up *both* at once, in equal measure, because the lopsided start $(x_1, x_2) = (A, 0)$ is the average of mode 1 $(A, A)/2$ and mode 2 $(A, -A)/2$. Each mode then runs at its own frequency, $\omega_1$ and $\omega_2$, and the left mass is their sum:
$$
x_1(t) = \frac{A}{2}\cos(\omega_1 t) + \frac{A}{2}\cos(\omega_2 t)
$$
Two cosines at nearby frequencies add to a product — the standard trigonometric identity — of a fast carrier at the average frequency and a slow envelope at half the difference:
> [Equation component] Labeled display-math block (KaTeX-rendered). Wraps a `$$...$$` math expression with an optional `id` for cross-references, an explicit `number` like "(3.2)", and a short `caption` shown below in monospace muted text. The math is rendered server-side via `remark-math` + `rehype-katex` (Katex is the rendering engine, not MathJax). Use this for the *important* equations — the ones the reader should remember, the ones the post's argument hinges on. A 2,000-word post should have 3-5 numbered equations, not 30; the rest stay as inline `$...$` math in running prose. Cross-reference via `equation (1)`.
```latex
x_1(t) = A\,\cos\!\left(\frac{\omega_2 - \omega_1}{2}\,t\right)\cos\!\left(\frac{\omega_2 + \omega_1}{2}\,t\right)
```
$$
x_1(t) = A\,\cos\!\left(\frac{\omega_2 - \omega_1}{2}\,t\right)\cos\!\left(\frac{\omega_2 + \omega_1}{2}\,t\right)
$$
The fast factor is the swing you see; the slow factor is the beat envelope that pumps the amplitude up and down. The beat frequency is $(\omega_2 - \omega_1)/2$ — it depends *only* on how far apart the two mode frequencies are, which is set by the coupling strength $k_c$. Weak coupling means $\omega_1$ and $\omega_2$ are nearly equal, the envelope is glacially slow, and the energy takes a long time to crawl across. Strong coupling spreads the modes apart and the beat speeds up.
> [Callout component] Styled info-block component (ported from the feelingdesigner project at ~/projects/feelingdesigner). Renders a rounded card with a tinted background, a 1px left accent bar in the type-specific colour, a quarter-circle SVG in the top-left corner that visually "cuts" the corner, and a floating icon badge that sits half-off the top edge. Seven types are available, each with its own accent colour and icon: info (blue, Info icon, neutral information), warning (yellow, AlertCircle, subtle caution), success (blue, CheckCircle, positive confirmation), error (red, XCircle, something is wrong), thinking (orange, Brain, an insight or mental model), feeling (red, Heart, a subjective observation), and doing (yellow, Hammer, a practical step to take). Used in the post to highlight key insights, contrasts, and gotchas without breaking the prose flow.
This is why musicians tune by listening for beats: two strings slightly out of tune produce a slow throb whose period is $1/(f_2 - f_1)$, and you tighten the peg until the throb slows to a stop. The beat is the audible image of equation 4. Same mathematics — two oscillators, slightly mistuned, heard as one wavering tone — whether the oscillators are violin strings, laser modes in an interferometer, or two pendulums on a shared rail. The slosh you watched is your eyes performing a frequency comparison.
Go back to the lab for a moment and turn the **Coupling k_c** slider up. The beat gets faster, exactly as equation 4 predicts: a larger $k_c$ pushes $\omega_2$ higher (equation 2), widens the gap $\omega_2 - \omega_1$, and shortens the envelope. Nothing about the modes themselves changed — they are still two clean metronomes — but the rate at which their interference pattern cycles tracks the gap you just widened.
## A mode is a closed ellipse in phase space
There is a second way to *see* that a normal mode is simple, and it lives in **phase space** — the plane whose axes are position and velocity. A single oscillator at one frequency, plotted as the point $(x, \dot{x})$ over time, traces a closed ellipse: position and velocity trade off, energy conserved, the same loop retraced forever. One frequency, one clean closed curve. That is the geometric signature of a normal mode.
> [LabSide component] Side-by-side lab layout: the same interactive lab effect as LabCanvas (referenced by its `effect` slug) rendered in one column with the post's prose (`children`) beside it, stacking vertically on mobile. `reverse` swaps the columns; `params` override defaults and `controls={false}` hides the effect's controls. Used to weave explanation and visualisation together rather than dropping the lab as an isolated figure. The rendered post has the live version; this is a placeholder for the markdown-only sibling.
Each point here carries an arrow, and trajectories follow the flow. For a *linear* normal mode the picture would be the dullest possible: one perfect ellipse, closed and unchanging, period and shape fixed for all time regardless of amplitude. That dullness is the whole gift — it is what lets you superpose modes by simple addition (equation 4) and guarantees the motion never surprises you.
Now switch **System** to *Duffing* and watch the curves bend. This Duffing oscillator has a *nonlinear*, double-well restoring force — a cubic $\beta x^3$ term over an inverted parabola, so the potential has two minima rather than one — and its phase portrait is no longer a single ellipse but a warped, amplitude-dependent family of loops threading both wells. Switch to *Van der Pol* and trajectories from everywhere spiral onto a single fat loop, a limit cycle that a linear system can never produce. These are the shapes that appear the moment you leave the small-oscillation neighbourhood where modes live.
The ellipse is fragile. It stays a perfect closed curve only because the restoring force is *linear* — proportional to displacement, $F = -kx$. That linearity is itself an approximation: a real pendulum's restoring force is $-mg\sin\theta$, and only for small angles does $\sin\theta \approx \theta$ make it linear. Normal modes are eigenvectors of the stiffness matrix you get by *linearising* the forces about equilibrium. Stay in the small-amplitude neighbourhood and the ellipses hold, the modes stay independent, the system is exactly solvable. Push the amplitude up and the approximation frays: the ellipse warps, the modes leak into each other, and the clean decomposition fails.
> [Callout component] Styled info-block component (ported from the feelingdesigner project at ~/projects/feelingdesigner). Renders a rounded card with a tinted background, a 1px left accent bar in the type-specific colour, a quarter-circle SVG in the top-left corner that visually "cuts" the corner, and a floating icon badge that sits half-off the top edge. Seven types are available, each with its own accent colour and icon: info (blue, Info icon, neutral information), warning (yellow, AlertCircle, subtle caution), success (blue, CheckCircle, positive confirmation), error (red, XCircle, something is wrong), thinking (orange, Brain, an insight or mental model), feeling (red, Heart, a subjective observation), and doing (yellow, Hammer, a practical step to take). Used in the post to highlight key insights, contrasts, and gotchas without breaking the prose flow.
The entire theory of small oscillations carries a hidden caveat in its name: *small*. Linearise the forces about equilibrium, diagonalise the resulting matrix, and you get modes — but only out to the radius where the linearisation is honest. There is no normal-mode decomposition of a pendulum swinging through 170°, no fixed eigenfrequencies for a spring stretched past its elastic limit. The modes are the tangent picture at the bottom of the well; far from the bottom, the well has a shape the tangent never captured, and that shape is where the trouble starts.
## Where it breaks: the double pendulum
Here is the cliff edge. Take the coupled system and make the coupling *strong* and the amplitudes *large* by hanging one pendulum directly off the end of another — a **double pendulum**. The two bobs are now coupled through the rigid arm and through gravity, and there is no weak-spring approximation to hide behind. The equations of motion are still exact Newtonian mechanics, no randomness anywhere. But they are violently nonlinear, full of $\sin(\theta_1 - \theta_2)$ and $\cos(\theta_1 - \theta_2)$ terms that no rotation of axes will ever diagonalise.
> [LabSide component] Side-by-side lab layout: the same interactive lab effect as LabCanvas (referenced by its `effect` slug) rendered in one column with the post's prose (`children`) beside it, stacking vertically on mobile. `reverse` swaps the columns; `params` override defaults and `controls={false}` hides the effect's controls. Used to weave explanation and visualisation together rather than dropping the lab as an isolated figure. The rendered post has the live version; this is a placeholder for the markdown-only sibling.
Three double pendulums start from angles a hair apart. For a beat or two they swing as one — that brief synchrony is the linear approximation still holding near the release point. Then they separate and never reconcile. Nudge **Gravity** up and the divergence comes sooner; the system stretches nearby states apart faster, which is the working definition of a larger Lyapunov exponent. There is no change of coordinates that turns this into three independent oscillators, because the motion ranges over angles where $\sin\theta \neq \theta$ and the linearisation that *defines* a normal mode never applies.
For *small* swings, the double pendulum *does* have normal modes — two of them, a slow in-phase mode and a fast out-of-phase mode, found by exactly the eigenvalue recipe of equation 3 applied to its linearised stiffness matrix. Release it gently and you would see clean beats, the same heartbeat as the spring-coupled masses. The chaos is not in the equations; it is in the *amplitude*. Crank the energy up and the system lives far from equilibrium, where the $\sin\theta$ nonlinearity dominates: the eigenvectors stop being constant, modes bleed into one another, and trajectories that start together are stretched apart exponentially.
> [Callout component] Styled info-block component (ported from the feelingdesigner project at ~/projects/feelingdesigner). Renders a rounded card with a tinted background, a 1px left accent bar in the type-specific colour, a quarter-circle SVG in the top-left corner that visually "cuts" the corner, and a floating icon badge that sits half-off the top edge. Seven types are available, each with its own accent colour and icon: info (blue, Info icon, neutral information), warning (yellow, AlertCircle, subtle caution), success (blue, CheckCircle, positive confirmation), error (red, XCircle, something is wrong), thinking (orange, Brain, an insight or mental model), feeling (red, Heart, a subjective observation), and doing (yellow, Hammer, a practical step to take). Used in the post to highlight key insights, contrasts, and gotchas without breaking the prose flow.
The coupled oscillator is the most-solvable system in mechanics: you write its motion as an exact, finite sum of independent modes, valid for all time, computed by diagonalising a matrix once. The double pendulum is among the least-solvable: no closed form, sensitive dependence on initial conditions, a prediction horizon you can compute but never beat. They are the *same kind of object* — masses, gravity, coupling — sitting on opposite sides of a single dividing line. On one side, $\sin\theta \approx \theta$ and the world is linear algebra. On the other, $\sin\theta$ is itself, and the world is chaos. Normal modes are the gift you get for staying near the bottom of the well.
> [StatGroup component] Editorial metric row — a wrapper for 2-4 `` components, rendered as a horizontal band that breaks up long prose. The individual stats follow as their own placeholders.
> [Stat component] Editorial stat callout. Renders one key metric as large `value` text under a `label` header, with optional smaller `context` subtext beneath. Used inside a `` to surface the numbers the post hinges on.
> [Stat component] Editorial stat callout. Renders one key metric as large `value` text under a `label` header, with optional smaller `context` subtext beneath. Used inside a `` to surface the numbers the post hinges on.
> [Stat component] Editorial stat callout. Renders one key metric as large `value` text under a `label` header, with optional smaller `context` subtext beneath. Used inside a `` to surface the numbers the post hinges on.
## The same idea in other rooms
Normal modes are not a pendulum curiosity. They are the universal language for any system of coupled linear oscillators, and the same eigenvalue problem (equation 3) appears in fields that never speak to each other.
In **molecular spectroscopy**, a molecule of $N$ atoms is $3N$ coupled oscillators — every atom tugged by chemical bonds to its neighbours. Strip out the six rigid-body motions (three translations, three rotations) and diagonalise the bond stiffness matrix, and you get $3N-6$ vibrational normal modes, each at its own frequency. A linear molecule loses one rotation and so keeps $3N-5$ of them: carbon dioxide, three atoms in a line, has four. Those frequencies are exactly what an infrared (IR) spectrometer reads: shine IR light through the sample, and the molecule absorbs at the photon energies that match its mode frequencies. The "fingerprint region" of an IR spectrum is a direct readout of the eigenvalues of equation 3 for that molecule. Carbon dioxide's bending mode near 15 µm — one of those four eigenmodes, and the band that makes it a greenhouse gas — sits almost exactly where the warm Earth radiates hardest.
In **electronics**, replace masses-on-springs with capacitors-and-inductors. Two LC resonant circuits coupled by a shared inductor obey the same matrix equation as equation 1, with charge playing the role of displacement. They beat, they have in-phase and out-of-phase modes, and the band-pass filter at the front of every radio is two coupled LC tanks tuned so their normal-mode frequencies straddle the channel you want. The engineer tuning a coupled-resonator filter is solving an eigenvalue problem with a screwdriver.
And in **solid-state physics**, push $N$ to Avogadro's number. A crystal is a 3D lattice of atoms on springs — $10^{23}$ coupled oscillators — and its normal modes are the **phonons**, the quantised lattice vibrations that carry sound and heat through every solid object you have ever touched. The same diagonalisation, just very large.
Which is what makes the **Fermi-Pasta-Ulam-Tsingou problem** of 1955 such a beautiful shock. Enrico Fermi, John Pasta, Stanislaw Ulam and Mary Tsingou ran one of the first scientific computations ever, on the MANIAC computer at Los Alamos: a chain of a few dozen masses on springs (the most-cited run used 32) with a *weak nonlinear* term added. The expectation, from statistical mechanics, was unanimous. Start all the energy in the lowest normal mode, let the nonlinearity couple the modes, and energy should spread until every mode holds its fair share — the chain should *thermalise*, reaching equipartition the way a hot object reaches uniform temperature. Instead the energy refused. It sloshed among a few low modes, then — astonishingly — came almost all the way *back* to the first mode, recurring nearly to its starting state. The normal modes were far more robust than anyone had any right to expect.
> [Callout component] Styled info-block component (ported from the feelingdesigner project at ~/projects/feelingdesigner). Renders a rounded card with a tinted background, a 1px left accent bar in the type-specific colour, a quarter-circle SVG in the top-left corner that visually "cuts" the corner, and a floating icon badge that sits half-off the top edge. Seven types are available, each with its own accent colour and icon: info (blue, Info icon, neutral information), warning (yellow, AlertCircle, subtle caution), success (blue, CheckCircle, positive confirmation), error (red, XCircle, something is wrong), thinking (orange, Brain, an insight or mental model), feeling (red, Heart, a subjective observation), and doing (yellow, Hammer, a practical step to take). Used in the post to highlight key insights, contrasts, and gotchas without breaking the prose flow.
The Fermi-Pasta-Ulam-Tsingou result was so surprising it sat semi-published for years, and untangling it launched two whole fields. The recurrence turned out to connect to *solitons* — stable nonlinear waves — and to the theory of *integrable systems*, where hidden conserved quantities prevent the expected thermalisation. The lesson is the one this whole post has been circling: normal modes are an approximation, but they are a *stubborn* one. Add nonlinearity and they do not instantly dissolve into chaos; sometimes they organise the motion for a remarkably long time before — or instead of — breaking down. The line between solvable and chaotic is not a cliff everywhere. Sometimes it is a long, gentle slope, and FPUT is the experiment that found it.
The coupled pendulums, the IR spectrum, the radio filter, the crystal, and the FPUT chain are one idea wearing five costumes: a system of coupled oscillators is a matrix, its modes are eigenvectors, and the motion is a superposition of independent metronomes — right up until the amplitude grows large enough that the matrix stops being constant, the eigenvectors start to drift, and the linear picture surrenders the system to chaos.
## Reading further
- [Goldstein, Poole & Safko, *Classical Mechanics*, 3rd ed., chapter 6](https://www.pearson.com/en-us/subject-catalog/p/classical-mechanics/P200000006339) — the canonical derivation of small oscillations as the eigenvalue problem of the stiffness and mass matrices; equation 3 in its full Lagrangian generality.
- [A. P. French, *Vibrations and Waves*](https://wwnorton.com/books/9780393099362) — the clearest physical-intuition treatment of coupled oscillators, beats, and normal modes; builds the two-pendulum case slowly and visually before any matrices appear.
- [Fermi, Pasta, Ulam & Tsingou, *Studies of Nonlinear Problems* (Los Alamos report LA-1940, 1955)](https://www.osti.gov/biblio/4376203) — the original computation; the chain that refused to thermalise and quietly opened the study of solitons and integrable systems.
- [Strogatz, *Nonlinear Dynamics and Chaos*, chapters 6–9](https://www.stevenstrogatz.com/books/nonlinear-dynamics-and-chaos) — the bridge from linear normal modes to the nonlinear phase portraits, limit cycles, and chaos that take over once the small-amplitude approximation fails.