9.7 Nonhomogeneous Problems & Eigenfunction Expansions ended with a promise: the machinery of project, solve one ODE per mode, and sum goes through unchanged on a drum, but the modes themselves are no longer sines. This lesson pays that debt. A flight simulator or a wind tunnel lives in two and three space dimensions, so the rods and strings of the last lessons were a rehearsal. The plan is the same: find the shapes that keep their form, find how fast each oscillates or decays, and add them up.
Transcript
- A drumhead is the string problem in two dimensions. Tension pulls from four sides, so the wave equation becomes u t t equals c squared times the Laplacian, and heat becomes u t equals alpha times the Laplacian. Separate time from space and both reduce to one question: minus the Laplacian of phi equals lambda phi, with phi zero on the edge. The frequency is c root lambda, the heat decay rate alpha lambda. On a grid this is a symmetric matrix: the modes are its eigenvectors.
- On a rectangle each factor is a string, so the modes are products of sines, with lambda m n equal to pi squared times m squared over a squared plus n squared over b squared. On the square, the frequency ratios to the lowest are 1, 1.5811, 2, 2.2361 and 2.5495. Not whole numbers, except that the two-two mode is the one-one mode with all lengths halved: an exact octave.
- On the square, the one-two and two-one modes share a frequency, a degenerate pair, because swapping x and y turns one into the other. So any mixture is also a mode. The identity, with sine pi x sine pi y times cosine pi y plus or minus cosine pi x, shows that the sum vanishes along x plus y equals 1, and the difference along x equals y. Numerically the sum is below ten to the minus fifteen there: rounding noise.
- Degeneracy can also be arithmetic. Count pairs with m squared plus n squared equal to s: for s equals 5 there are 2, for 25 still 2, but 50 gives 3, and 65 gives 4. Symmetry is fragile, though. Stretch the square to 1 by 1.1 and the eigenvalues separate, 42.496 against 47.635, a 12 percent gap, so the nodal line drifts.
- Now heat. A plate starts at 1 with cold edges. The lowest eigenvalue on the unit square is 19.739, a time constant of 0.0507. At the centre, all the modes give 0.99837 at time 0.01, while the lowest mode alone says 1.33074, above the starting value. By 0.05 one mode is within 1.3 percent, by 0.1 within 0.02 percent. High modes die first, and the plate relaxes to the lowest eigenvector.
- Shape matters. Lambda one one is pi squared times 1 plus 1 over b squared. At b equal 0.5 it is 49.348, and the centre is only 0.13660 at time 0.05. At b equal 2 it is 12.337, and the centre is still 0.76989. The narrowest dimension sets the rate. At equal area, the square has 19.739, the thin rectangle 41.946, and the disk 18.168, the slowest. In a duct, the same eigenvalue sets the viscous spin-down.
- A circular drum separates in polar coordinates: cosine m theta around, and Bessel's equation along the radius. The clamped rim needs J m to vanish, so the frequencies are Bessel zeros: 2.404826, 3.831706, 5.135622. Divided by the first they give 1, 1.5933, 2.1355, 2.2954, 2.6531. No integer pattern: a drum gives a thump, not a pitch. Mode m n has m nodal diameters and n minus 1 nodal circles.
- With no formula, replace the Laplacian by the five-point stencil. On the square the error in the lowest eigenvalue is second order: 19.4868, 19.6759 and 19.7234 at N of 8, 16 and 32, against 19.7392, and each halving cuts the error by 4. On the disk, the staircase boundary is slower: 109, 437 and 1789 unknowns give errors of 9.9, 3.3 and 2.1 percent. The grid also splits the exact pair 26.3746 into 25.758 and 25.782.
- Can you hear the shape of a drum? Weyl's law counts the eigenvalues below lambda as about area times lambda over 4 pi, less a perimeter correction. For the unit square the exact counts are 71, 300 and 764 below 1000, 4000 and 10000. Area alone gives 79.6, 318.3 and 795.8; with the perimeter, 69.5, 298.2 and 763.9. So the spectrum reveals area and perimeter, though in 1992 two different polygons were found with identical eigenvalues. Cavities and tanks use this same eigenproblem.
- To recap. Two dimensions change the geometry, not the plan: minus the Laplacian of phi equals lambda phi. Rectangles give product modes and moving nodal lines, disks give Bessel modes with no pitch, and heat relaxes to the lowest mode. Where there is no formula, the five-point matrix computes the eigenvalues. Next: infinite domains and the Fourier transform, where the eigenvalues form a continuum and the sum becomes an integral.
From a line to a plane
A membrane is a drumhead: a thin sheet under tension (force per unit edge length) with mass per unit area. On a tiny square of side , the left and right edges pull like a string, giving the vertical force , and the top and bottom edges add . Newton’s law with mass gives
the 2D wave equation. The symbol is the Laplacian: the operator of 9.6 Laplace's Equation: Rectangles & Disks. The heat equation of 9.2 The Heat (Diffusion) Equation becomes . In both, on the edge (a clamped rim, a cold wall).
Separate time from space as in 9.4 The Wave Equation & Vibrating Strings: . Then must be a constant, , and
Each eigenvalue gives a wave of angular frequency and a heat mode that decays like . All the geometry sits in one question: the eigenfunctions and eigenvalues of on this shape. This is the Sturm-Liouville problem of 8.2 Eigenvalue Problems & Sturm–Liouville Theory with two space variables: the eigenvalues are positive, eigenfunctions with different are perpendicular (), and every initial shape is a sum of them.
The rectangular membrane
Take the rectangle , and separate again, . Then , so each term is its own constant and each factor is a string problem with fixed ends:
with The general solution is , with from a double sine series (the series of 8.3 Fourier Series once in , once in ). Mode has vertical and horizontal nodal lines, curves where the sheet does not move; sand on a vibrating plate collects there. (Problem 1 puts numbers on a square drum.)
The ratios to the lowest, , are for (from the sim): not whole numbers, except that is with all lengths halved, an exact octave.
Degeneracy and the lines that move
On a square, : two different shapes with one frequency, a degenerate pair. The reason is symmetry, since swapping and turns one mode into the other. The consequence is new in 2D. Any combination oscillates at the shared frequency, so it is also a mode, with its own nodal curve. Using ,
The factor vanishes on the edge. The bracket vanishes where , which is the line for the plus sign and for the minus. The sim checks both: the sum of the two modes along is below at every sample, which is rounding noise. Intermediate bends the nodal line, giving Chladni patterns (from memory: real Chladni figures are made on plates, which obey a fourth-order equation, so a membrane gives the idea, not the exact patterns).
Degeneracy can also be arithmetic. Counting ordered pairs with (the sim’s counter): gives only , since is not allowed, while gives three, , and gives four. The coincidence of with the others is number theory, not symmetry.
Symmetry is fragile. Stretch the square to and the sim gives against : a 12% gap. The pair separates, and the combination is no longer a single mode: it beats, and its nodal line drifts.
What to try. In the rectangle view press , then set the aspect to and animate. Switch to the disk, set , then , and compare the frequency ladder with the string’s dashed harmonics.
Heat on a rectangle
Heat uses the same eigenproblem with decay in place of oscillation. A plate , starts at with its edges held at (). The coefficient of is , which for odd gives (zero if either is even, as in the 1D rod). So
For the unit square , so the time constant of the slowest mode is . The centre temperature from the sim’s 59-term sum is
| 0.01 | 0.02 | 0.05 | 0.10 | |
|---|---|---|---|---|
| all modes | 0.99837 | 0.95094 | 0.59647 | 0.22514 |
| lowest mode only | 1.33074 | 1.09236 | 0.60421 | 0.22519 |
Early on one mode is badly wrong (even above the starting value ): the sharp edges of the initial data need many modes. By one mode is within and by within : high modes die first and the plate relaxes to . Repeated smoothing finds the lowest eigenvector, an idea the numerical method below reuses.
Shape effects. Stretching a side changes . At it is (time constant ) and the centre temperature at is ; at it is (time constant ) and the centre is still . A thin plate cools through its short side: the narrowest dimension sets the rate. At equal area (1), the square has , the rectangle , and the disk (radius ) , with from the next section. The disk cools slowest, and I recall that this is a theorem, Faber-Krahn: among shapes of equal area the disk has the smallest first eigenvalue (from memory).
The circular drum
For a disk of radius use polar coordinates and the Laplacian . Separate, . The angular equation must be periodic, since and are the same point, so or with The radial equation, with , is
Bessel’s equation of order (5.8 Series Solutions). One solution, , is smooth at the centre; the second blows up there and is discarded because a drumhead does not. The clamped rim requires , so is a zero of . Call the -th positive zero . Then
For the cosine and sine modes share a frequency: a degenerate pair caused by rotational symmetry. Mode has nodal diameters and interior nodal circles, so moves as one piece and rocks about a diameter.
The sim evaluates (from memory) by the trapezoid rule, scans for sign changes and bisects:
| (0,1) | (1,1) | (2,1) | (0,2) | (3,1) | (1,2) | (4,1) | (2,2) | (0,3) | (0,4) | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2.404826 | 3.831706 | 5.135622 | 5.520078 | 6.380162 | 7.015587 | 7.588342 | 8.417244 | 8.653728 | 11.791534 | |
| ratio to | 1 | 1.5933 | 2.1355 | 2.2954 | 2.6531 | 2.9173 | 3.1555 | 3.5001 | 3.5985 | 4.9033 |
Why a drum has no pitch
A string’s frequencies are whole multiples of the fundamental, so its sound repeats with one period and the ear hears a pitch. The drum’s lowest ratios are , with no integer pattern: Bessel zeros are roughly apart only far out (from memory). The sound is a cluster of unrelated frequencies, a thump with a vague pitch. I recall that a timpani’s air loading moves toward , , (the sim has ), which is why it can be tuned (from memory). Expanding an initial shape uses a Fourier series in and a Fourier-Bessel series in with weight , as in 8.6 Orthogonal Polynomials & Special Functions.
What if there is no formula?
Rectangles and disks are the exceptions; for a wing section the eigenvalues must be computed. Put a grid of spacing on the region and replace by the 5-point stencil
with at nodes outside the region. This symmetric matrix has on the diagonal and for each neighbour; its eigenvectors, laid back on the grid, are the discrete modes.
The square. Here the eigenvalues are known, (the sim’s values match it to every printed digit). Expanding the sine gives the relative error for the lowest one: second order. At the formula gives against , an error of . The sim’s iteration, which never uses the formula, gives
| 8 | 16 | 32 | exact | |
|---|---|---|---|---|
| 19.4868 | 19.6759 | 19.7234 | 19.7392 | |
| error |
Each halving of cuts the error by . The pair stays equal (47.2338, 48.8116, 49.2134 against ) because the grid has the square’s symmetry.
How it finds them. The sim solves and repeats with (inverse iteration): each solve divides the component along an eigenvector by its , so the smallest wins, much as the heat decay above leaves the lowest mode. It carries ten vectors and re-diagonalizes in that small subspace each round, since one vector alone stalls on a close pair (10.1 Finite Differences: Discretizing Derivatives has the stencil).
The disk, on a square grid. Keep the nodes strictly inside the circle. The boundary is a staircase. The lowest eigenvalue against the exact :
| 12 | 24 | 48 | |
|---|---|---|---|
| unknowns | 109 | 437 | 1789 |
| 5.2085 | 5.5951 | 5.6634 | |
| error |
This is far slower than : the staircase puts the wall in the wrong place by up to a grid step, so I expect first order at best. Smooth boundaries need cut cells or fitted grids (10.11 2D Grids, Boundary Conditions & Staggered (MAC) Grids). The grid also shows its axes. The pair stays degenerate ( twice at , against ), but the pair, exactly twice, splits into and at ( and at ). A turn only flips the sign of , while a turn gives , and the grid lacks symmetry.
Can you hear the shape of a drum?
Each shape has its own list of , so does the list identify the shape? Mark Kac posed this as “can one hear the shape of a drum?” (1966, from memory). Weyl’s law (from memory) says the number of eigenvalues below is about for area , and with the perimeter the next term is (from memory). For the unit square (, ) the sim counts lattice points:
| 1000 | 4000 | 10000 | |
|---|---|---|---|
| exact count | 71 | 300 | 764 |
| area term only | 79.6 | 318.3 | 795.8 |
| area and perimeter | 69.5 | 298.2 | 763.9 |
So a long list of eigenvalues reveals area and perimeter. But I recall that the full answer is no: in 1992 Gordon, Webb and Wolpert found two different polygons with identical eigenvalues (from memory).
Check your understanding
Problems
A square membrane has side m and m/s. Find the three lowest distinct frequencies in hertz and say which is degenerate.
Show worked solution
Hz. The values belong to , to and , and to , giving , and Hz. The middle one is degenerate.
Show that vanishes on , and find the nodal curve of the difference. Why are both modes at the same frequency?
Show worked solution
The sum is . On , , so the bracket is . For the difference the bracket is , zero on . Both combine and , which share , so one time factor serves both.
A drum has radius m and m/s. Use , , , to find the frequencies of modes , and say which have nodal circles and diameters.
Show worked solution
Hz: , , , . Only has an interior nodal circle (); and have one and two nodal diameters; has no interior nodal line.
(a) Show . (b) Estimate the relative error of the lowest 5-point eigenvalue on the unit square. (c) For compare with the exact .
Show worked solution
(a) with gives . (b) Two directions with : , relative error . (c) At the estimate is . Directly, and , so the error is ; the next series term explains the difference.
Computation
First, the exercise: build the 5-point Laplacian and compute its eigenvalues on the unit square, and compare them with .
Second, a disk on a square grid. The staircase boundary makes the eigenvalues converge slowly, and splits the pair.
Try and watch the fourth and fifth values separate.
Key Takeaways
- The 2D wave and heat equations reduce to with on the edge; frequency is and decay rate . On a grid this is a symmetric matrix eigenproblem.
- On a rectangle the modes are products of sines with . On the square and share a frequency, so are modes with nodal lines and ; stretching to splits the pair ( against ).
- On a disk, and , with nodal diameters and nodal circles. The ratios are not whole numbers, so a drum has no clear pitch.
- Heat on a plate decays at : for the unit square, at , at . The lowest mode takes over after a few time constants.
- The 5-point stencil is second order on the unit square (relative error ; , , against ). On a staircase disk the lowest value is off by , , at and the pair splits.
- Weyl counting (from memory) gives area and perimeter from the spectrum: eigenvalues of the unit square below against predicted.
What comes next
Every region here was bounded, so the spectrum was a discrete list and the solution a sum. A wind-tunnel wake, a pulse on a very long wire or heat from a point in unbounded space has no edge to fix the modes: the eigenvalues form a continuum and the sum becomes an integral. The next lesson, 9.9 Infinite Domains & the Fourier Transform, develops this Fourier transform, with the Gaussian heat kernel and group velocity as payoffs that no bounded region can give.