9.8 Higher Dimensions: Membranes & Drums ended on a limitation. Every region there was bounded, so the modes formed a discrete list and the solution was a sum. A wake behind a wing, a pulse on a very long wire, or a puff of heat in open air has no wall to fix the modes. A flight simulator spends most of its time in exactly this situation: the domain is large compared with anything of interest. This lesson takes the Fourier series of 8.3 Fourier Series to its limit, where the list of modes becomes a continuum and the sum becomes an integral. Two payoffs follow: the heat equation gets a closed-form answer for any initial data, and waves get a vocabulary (dispersion, group velocity) that no bounded problem could teach.
Transcript
- Bounded regions gave a list of modes and a sum. With no walls, take a function of period two L and let L grow. The wavenumbers sit pi over L apart, so they crowd together, and the series, a Riemann sum, becomes an integral. That is the transform pair: F of k is the integral of f times e to the minus i k x, and f is one over two pi times the integral of F times e to the i k x. The two pi sits in the inverse.
- The Gaussian is the pair worth memorising. A Gaussian of width s transforms to a Gaussian of width one over s: narrow in space means wide in wavenumber. Measuring both spreads, the product is 0.5000 for every Gaussian. For the hyperbolic secant it is 0.5236, which is pi over 6, a little above the bound. You cannot localise a signal in both space and wavenumber.
- Two rules do the heavy lifting. Differentiating in x multiplies by i k, and a second derivative multiplies by minus k squared, so a constant coefficient PDE becomes an ODE in time for each wavenumber, with no coupling. And convolution in space is plain multiplication in wavenumber. A kink such as e to the minus absolute x transforms to two over one plus k squared, falling slowly because of the corner. Energy is preserved too: both sides of Plancherel give 1.772454 for the unit Gaussian.
- Now heat on the whole line. Transform, and each wavenumber decays at its own rate: U equals F times e to the minus alpha k squared t. That factor is itself a Gaussian pair, so the answer is the initial data convolved with the heat kernel, a Gaussian of variance two alpha t. The width grows like the square root of t. For a box of heat, the centre reads 0.9747, 0.8427, 0.5205 and 0.2763 at times 0.1, 0.25, 1 and 4, while the total stays 2.0000.
- The variance of the box gains 0.2000, 0.5000, 2.0000 and 8.0000: exactly two alpha t each time. At the centre the box obeys the error function of one over two root t, which falls to one half at t about 1.10. And there is no finite front: at t equals 1, five units beyond the edge, the temperature is 2.0 times ten to the minus 4. Heat multiplies by a real decay, whose kernel has infinite tails; a wave only rotates a phase.
- Waves by transform: u t t equals c squared u x x gives U equals F times cosine of c k t, which is two travelling copies of the initial shape. That is d'Alembert's formula, and the spectral solution of a Gaussian pulse matches it to below ten to the minus twelve. The reason is the dispersion relation, omega equals c k: every wavenumber travels at the same speed. Other equations give other omega of k.
- A plane wave's crests move at the phase velocity, omega over k. But a packet of nearby wavenumbers has an envelope, and the envelope moves at the group velocity, d omega d k. For u t plus u x x x equals zero, omega is minus k cubed. At k equals 3 the crests move at minus 9 and the envelope at minus 27: three times faster. For deep water the crests go at 0.5774 and the energy at 0.2887, half as fast. Measured on the grid, the packet centroid moves at minus 27.094 and 0.2891.
- When omega is not linear, the envelope also spreads. The packet of width 2.8284 keeps it when omega equals k, even at time 10. Under minus k cubed it grows to 4.2594 by time 1 and 6.9692 by time 2: more than double. Deep water spreads slowly, 2.8297 at time 10. That is dispersion: a splash sorts itself by wavelength, and a ship's wake is the pattern.
- On a periodic box the same rule gives spectral differentiation: take the FFT, multiply by i k, transform back. For e to the sine x, second order differences lose a factor four per doubling, and fourth order a factor sixteen. The FFT error is 0.00432 at 8 points, 1.8 times ten to the minus 7 at 16, and rounding by 32. Sixteen points beat 128 points of fourth order. But it needs smoothness: for absolute sine cubed the FFT error also drops only by a factor 4.
- To recap. Letting the period go to infinity turns the series into the transform pair. Gaussians trade width for spectrum, derivatives become multiplication, and heat is a Gaussian blur of variance two alpha t. Waves only rotate phases, and dispersion separates the crests from the envelope. The FFT turns this into a derivative that beats finite differences on smooth data. Next: wave equation numerics and the CFL condition, where the grid's own dispersion decides what is stable.
From a sum to an integral
Take a function of period and write the series in the complex form of 8.3 Fourier Series, with and spacing :
Put . Then , so the series reads . This is a Riemann sum. Let : the spacing , and if decays fast enough the sum becomes an integral. That is the Fourier transform pair:
This lesson uses this convention throughout: is the angular wavenumber (radians per unit length), the forward transform carries no factor, and the inverse carries . Other books put on both, or use and no factors at all, so the formulas below change by factors of if you swap conventions. The orientation matches np.fft.fft, which is the unnormalized sum with . is the amplitude of the wave needed to build ; is the total.
The Gaussian and the uncertainty tradeoff
The one transform worth knowing by heart is the Gaussian. For ,
(Complete the square in and use ; the shift into the complex plane is justified in a complex analysis course.) A Gaussian transforms to a Gaussian, and the widths are inverse: narrow in means wide in . Scaling says why. Squeezing with gives , which is wider in by the same factor. A sharp spike needs every wavenumber to build it.
Measure width by the standard deviation of the energy density, , and likewise from . The sim computes both by FFT on a grid. For the Gaussian with it gives and , product ; at the widths are and , still . For the product is at every (it equals ; the sim’s is the integral). I recall that for every signal, with equality only for Gaussians (from memory), which is the mathematical core of the Heisenberg principle. For a simulation it has a plain meaning: a grid that resolves features of width must carry wavenumbers up to about , and you cannot localize in both and .
Derivatives become multiplication
Integrate by parts, with the boundary terms vanishing because decays:
Differentiation in is multiplication by . This is why Fourier methods exist: a constant-coefficient PDE becomes an ODE in for each , with no coupling between wavenumbers. It also reads smoothness off the spectrum: has transform (Problem 1), falling only like , because the kink in means jumps and cannot decay faster than . The Laplace transform of 6.1 The Laplace Transform did the same to ODEs in time; a Fourier transform in time is a Laplace transform along , so the two are one idea.
Convolution theorem. The convolution has transform (substitute and the double integral factors). A product of spectra is a blend in space, exactly as in 6.5 Convolution & Transfer Functions. We use it next.
Heat on the whole line
Solve on with . Transform in : , so
Each wavenumber decays at its own rate, as the modes of 9.3 Solving the Heat Equation: Separation of Variables did, except that now every is present. The factor is the Gaussian pair with , whose inverse transform is the heat kernel
The solution is the initial data blurred by a Gaussian of variance : the width grows like , not like . A point of heat at spreads into a bell centred at , and the answer is the sum of the bells.
Worked example. Start from with . Convolving two Gaussians adds variances, , and the area is conserved, so
At the peak is and the width ; the area is unchanged. The area is the mode, which does not decay.
A box. For on , the integral of the kernel over gives . The sim solves this by FFT instead, with (all quoted to four digits):
| 0.1 | 0.25 | 1 | 4 | |
|---|---|---|---|---|
| 0.9747 | 0.8427 | 0.5205 | 0.2763 |
The variance of grows by , which is each time, and throughout. Multiplying by and convolving with the sampled kernel agree to rounding error, below : the convolution theorem, checked.
What to try. In the heat view slide from and watch the right panel: high wavenumbers die first, which is the smoothing. Then open the other views as you read on.
The contrast with waves
Heat has no finite speed. At the box solution at , five units from the edge, is , small but positive; at it is . Every point feels every other instantly, because each Gaussian has infinite tails. The wave equation behaves differently, as 9.5 d'Alembert's Solution & Characteristics showed: data with compact support gives exactly zero outside the cone . The multiplier is the difference. Heat multiplies each by a real , which shrinks it; waves multiply by a phase, which only rotates it.
Waves by transform
Transform in : . With , . A factor is the transform of (shift theorem), so : d’Alembert’s formula, recovered. The sim agrees: for the pulse at the FFT solution and d’Alembert differ by less than (rounding).
The reason it works for every at the same speed is that the relation between frequency and wavenumber is linear, . In general, substitute a plane wave into a linear PDE with constant coefficients. The result is the dispersion relation :
| equation | |
|---|---|
| (imaginary: decay) | |
| (a beam) |
A real means oscillation, an imaginary part means growth or decay. Each plane wave has phase velocity . Now take a packet of wavenumbers near and expand . The exponent splits,
The first factor is the carrier, moving at . The second depends only on , so the envelope moves at the group velocity . The energy travels at . When the linear approximation fails at the next order, and the envelope also spreads. That is dispersion.
The sim tests this on the packet with .
| model | measured centroid speed | envelope width at , later | ||
|---|---|---|---|---|
| 1 | 1 | 1.0000 | 2.8284 at : 2.8284 | |
| 2.8284, then 4.2594 at and 6.9692 at | ||||
| deep water | 0.5774 | 0.2887 | 0.2891 | 2.8284 at : 2.8297 |
For the envelope runs at three times the speed of the crests and more than doubles its width in two time units; for nothing changes shape. The small excess over is not an error: the packet has a spread of wavenumbers, and the centroid moves at the energy-weighted average of , which here is (the average formula is from memory; the sim reproduces its number).
Energy: Plancherel
The transform preserves energy. The same Riemann-sum argument that took Parseval’s identity of 8.3 Fourier Series to infinity gives
The sim prints both sides for the Gaussian with : and , which is . Apply it in time. Waves multiply each mode by a phase, so never changes and the energy is constant. Heat multiplies by , so shrinks and the energy decays, fastest where is large. The same measure that separated waves from heat above is the energy.
Spectral differentiation on a periodic box
A computer has a finite box, so replace the line by a periodic domain of length with samples . The discrete transform is the DFT and the FFT computes it in operations (8.5 Convergence & the Gibbs Phenomenon, which also explains aliasing). The recipe for is the derivative rule applied to samples: FFT, multiply by , inverse FFT. The one care is the Nyquist mode , whose derivative would be a complex number; set that multiplier to .
The sim compares this with the central difference of 10.1 Finite Differences: Discretizing Derivatives (second order) and a five-point central difference (fourth order), on , with maximum error over the nodes:
| 8 | 16 | 32 | 64 | 128 | |
|---|---|---|---|---|---|
| 2nd-order FD | 0.340 | 0.0936 | 0.0258 | 0.00651 | 0.00163 |
| 4th-order FD | 0.138 | 0.0158 | 0.00113 | ||
| FFT | 0.00432 | rounding | rounding | rounding |
The finite differences lose a fixed factor (, then ) per doubling of ; the FFT error falls faster than any power, and by it has reached rounding error near , which is why no digits are quoted. Sixteen points with the FFT beat points of fourth-order differences by a factor above . This is spectral accuracy, and it needs a smooth periodic function. For , whose third derivative jumps, the FFT error at is : a fixed factor again. Accuracy is set by smoothness, as in the convergence table of 8.5 Convergence & the Gibbs Phenomenon.
Check your understanding
Problems
Find for with , and say how fast decays at large and why.
Show worked solution
Split at . , and the half on is the complex conjugate. Sum: . It falls like : has a corner at , jumps, and can fall no faster than , so .
With and on , show and find when it falls to .
Show worked solution
. Put : . The inverse error function satisfies , so , giving .
For find , and . For and a packet at , which is faster, the crests or the envelope, and by how much?
Show worked solution
Substituting : , so . On the branch and . At : , . The envelope is twice as fast as the crests, and shorter waves travel faster still, so a beam hit by an impulse rings with the high frequencies arriving first (from memory).
Computation
The exercise is the whole spectral method in a few lines: the wavenumbers, the derivative, and one exact heat step.
Second, a playground: a packet under three dispersion relations. The envelope’s centre moves at the group velocity.
Change and and compare the centroid with .
Key Takeaways
- Letting the period turns the Fourier sum into the pair , . The convention (no factor forward, back) matters: other books move the .
- A Gaussian transforms to a Gaussian with inverse width, so (and for ). Narrow in space means wide in wavenumber.
- turns constant-coefficient PDEs into one ODE per wavenumber; convolution in is multiplication in .
- Heat on the line: , so with the Gaussian kernel of variance . Width grows like , the area is conserved, and the tails are positive instantly ( for a box). Waves have a finite front.
- The dispersion relation gives phase velocity and group velocity ; for at they are and . A nonlinear also spreads the envelope. Deep water has (from memory).
- Plancherel: . Phases conserve energy; the real exponential of heat dissipates it.
- On a periodic grid, FFT differentiation is spectrally accurate for smooth data ( at for , against for fourth-order differences) and falls back to second order when the third derivative jumps.
What comes next
The dispersion relation just derived assumed the continuum. A computer replaces by a difference and by another, so the scheme has its own , and the same plane-wave test shows which wavelengths it moves at the wrong speed and when it blows up. That is the content of 10.5 Wave Equation Numerics & the CFL Condition: the stability limit of the leapfrog wave scheme, the CFL condition, and the numerical dispersion that spreads a grid-scale pulse into ripples (from memory; I have not computed it here). The same wavenumber bookkeeping will carry through, only with the discrete in place of .