Hear the Shape of a Drum

After reading this you will know what "the sound of a drum" means mathematically, how a computer finds those frequencies, and why two drums of different shape can ring at exactly the same pitches.

What the question actually asks

Strike a drum and it rings at a fixed set of frequencies. Those frequencies do not depend on how hard you hit it, only on the drumhead: its shape, how tightly it is clamped at the rim, and the material. In 1966 Mark Kac phrased a clean question about this. Suppose you write down every frequency a drumhead produces, the full list. Does that list pin down the drum's outline? Put visually: if two drums produce the identical list of frequencies, must they have the same shape?

The honest answer, found in 1992, is no. Carolyn Gordon, David Webb and Scott Wolpert built two drums with different shapes that share every single frequency. This tool lets you hear that. You draw a membrane, the computer solves for its vibration modes, and you listen. The default pair is the Gordon-Webb-Wolpert counterexample, so you can take a blind test and try to tell them apart.

The membrane here is idealized. It has no stiffness, no air loading, and it is clamped rigidly at the edge, meaning the rim never moves. That is the exact object Kac asked about, and it is simpler than a real snare drum. That simplicity is what makes the frequencies computable and the theorem clean.

The equation behind the sound

A clamped membrane obeys the two-dimensional wave equation. Its standing waves separate into a spatial shape multiplied by a sine in time. The spatial shapes are the solutions of an eigenvalue problem called the Helmholtz problem:

-\Delta u = \lambda u, \quad u = 0 \text{ on the rim}

Here u(x,y) is the shape of one vibration mode, the height of the drumhead at each point. The symbol \Delta is the Laplacian, \partial^2 u/\partial x^2 + \partial^2 u/\partial y^2, which measures how curved the surface is. The condition u = 0 on the rim says the edge is nailed down. Only special values of \lambda allow a nonzero solution. Those values are the eigenvalues, and there are infinitely many of them: \lambda_1 \lt \lambda_2 \le \lambda_3 \le \cdots.

Each eigenvalue gives one audible frequency by

f_n = \frac{c}{2\pi}\sqrt{\lambda_n}

where c is the wave speed on the membrane, set by tension and mass. Doubling the tension speed multiplies every frequency by the same factor, so it shifts pitch without changing the pattern of ratios. The shape lives entirely in the \lambda_n. That is why "the spectrum" of a drum means the list \lambda_1, \lambda_2, \lambda_3, \dots, and hearing the drum means hearing the ratios \sqrt{\lambda_n/\lambda_1}.

How the browser computes the modes

You cannot solve -\Delta u = \lambda u in closed form for a general polygon. The disk and the rectangle have known answers; almost nothing else does. So the tool discretizes. It covers the drum with thousands of small triangles (a Delaunay mesh), and on each triangle it approximates u by a quadratic polynomial. This turns the continuous problem into a matrix problem:

K \mathbf{u} = \lambda M \mathbf{u}

K is the stiffness matrix, encoding the Laplacian, and M is the mass matrix, encoding area. Both are sparse: each mesh node touches only its neighbors, so most entries are zero. A sparse Cholesky factorization of M plus a Lanczos iteration extracts the smallest few eigenvalues fast, which are exactly the low modes you can hear as distinct pitches. Finer meshes give more accurate \lambda_n; the tradeoff is compute time in your browser.

Finite elements always overestimate eigenvalues slightly, because forcing u to be piecewise polynomial stiffens the membrane. Refining the mesh lowers every computed \lambda_n toward the true value. Treat the last digit of any frequency as uncertain.

A worked example: the disk

Pick the Circle preset and you can check the computer against exact theory. For a disk of radius R, the eigenvalues are \lambda = (j_{m,k}/R)^2, where j_{m,k} is the k-th positive zero of the Bessel function J_m. Every preset is scaled to the same area, 3.5 square units (the area of the Gordon-Webb-Wolpert drums with unit legs), so take R = \sqrt{3.5/\pi} \approx 1.0555.

Reproducing the first four disk modes

  1. The first Bessel zeros are j_{0,1} = 2.405, j_{1,1} = 3.832, j_{2,1} = 5.136, j_{0,2} = 5.520.
  2. Square them and divide by R^2 = 3.5/\pi \approx 1.1141. Fundamental: \lambda_1 = 2.405^2 / 1.1141 = 5.192.
  3. Next mode is doubly degenerate (two shapes, same pitch): \lambda_2 = \lambda_3 = 3.832^2 / 1.1141 = 13.18.
  4. Then \lambda_4 = \lambda_5 = 5.136^2 / 1.1141 = 23.68, and \lambda_6 = 5.520^2 / 1.1141 = 27.35.
  5. Frequency ratios to the fundamental: \sqrt{13.18/5.192} = 1.593 and \sqrt{23.68/5.192} = 2.136.

Those ratios are not whole numbers, which is why a drum sounds inharmonic and un-pitched compared to a string, whose overtones are 2, 3, 4 times the fundamental. The tool's computed values should match these to within a fraction of a percent.

The draft-mesh values sit just above the exact eigenvalues, by about 0.01 %, the expected finite-element overestimate. The gap shrinks further as you refine the mesh.

Weyl's law: what the outline still gives away

You cannot hear the full shape, but you can hear a lot. Count how many eigenvalues fall below a cutoff \lambda and call it N(\lambda). Weyl's law says this count grows in a way fixed by area and perimeter:

N(\lambda) \approx \frac{A}{4\pi}\lambda - \frac{L}{4\pi}\sqrt{\lambda} + C

A is the area, L the perimeter, and C a constant set by the corners and topology. So the area is audible: it controls how densely the modes crowd together at high frequency. The perimeter is audible too, in the second term. For the tool's disk, of area 3.5, A/4\pi = 3.5/(4\pi) = 0.279, so by \lambda = 60 you expect roughly 0.279 \cdot 60 = 16.7 modes minus the perimeter correction, about 12.8 in all. The exact count is 12, and the tool finds the same 12.

The Gordon-Webb-Wolpert drums share the same area, the same perimeter, and the same corner constant, so their Weyl counts agree at every order. And beyond the approximation, every individual eigenvalue matches. That is the whole point of the counterexample: nothing in the spectrum distinguishes them.

The smoothed count from Weyl's law. Real eigenvalues make this a staircase that jumps at each mode, tracking this curve. The leading slope is the area over 4 pi.

Why two drums can sound identical

The counterexample is built from seven identical right triangles (half-squares) glued edge to edge, in two arrangements that cannot be rotated or reflected into each other. Given a vibration mode on the first drum, you can cut it into seven pieces, one per triangle, and reassemble those pieces on the second drum with a fixed pattern of signs and reflections. The reassembled function turns out to be a valid mode of the second drum at the same eigenvalue. This "transplantation" maps modes one to one, so the two spectra are equal term by term. That argument, not a computation, is what settled the question.

The two Gordon-Webb-Wolpert drums are each made of seven congruent half-square triangles glued in different patterns. Any mode of one drum can be cut and reflected piece by piece to build a matching mode of the other at the identical eigenvalue, which is why every frequency agrees.

Reading the results and avoiding traps

When you strike a point, the tool sounds a weighted sum of modes. "Real strike" weights each mode by its height at the point you clicked, so a mode with a nodal line under your finger is silent. Hit near a nodal line and the timbre changes even though the drum's spectrum has not. "Pure spectrum" ignores the strike point and weights modes by frequency alone. That is the fair blind test, because it removes timbre and leaves only the pitches Kac cared about.

Do not conclude two drums differ just because a real strike sounds different. Isospectral drums can have different timbres depending on where you hit them, because the mode shapes differ even when the eigenvalues match. Only "pure spectrum" tests the theorem. That is where the two drums become genuinely indistinguishable by ear.

Two more traps. First, degeneracy: the disk's modes 2 and 3 are one pitch heard once, not two. Counting distinct pitches instead of eigenvalues will confuse you. Second, mesh resolution: if you compare a coarse mesh of one drum to a fine mesh of another, small numerical gaps can masquerade as real spectral differences. Use the same resolution on both.

Eigenvalue \lambda_n
A number for which the membrane has a standing-wave solution. Sets the pitch through f \propto \sqrt{\lambda}.
Nodal line
The curve where a mode has zero height. The drumhead does not move there.
Isospectral
Two shapes with the identical list of eigenvalues, hence the identical spectrum.
Spectrum
The full ordered list of eigenvalues, the mathematical stand-in for "the sound".

Related tools

To see nodal lines form physically instead of computed, vibrate a real square plate in the Chladni Plate Patterns tool and watch sand settle onto the lines where the surface is still. For the one-dimensional cousin of this problem, drive a string and watch it lock onto its harmonics in Standing Waves on a String. To build up two-dimensional wave patterns from moving sources rather than fixed modes, drop ripples in the Wave Interference tank. And for a different lesson about how sensitive a system is to its exact starting shape, watch trajectories diverge in the Double Pendulum.

Frequently asked questions

Can you hear the shape of a drum?

Not in general. In 1992 Gordon, Webb and Wolpert built two differently shaped drums with the exact same spectrum, so the frequencies do not determine the shape. You can hear the area and perimeter (through Weyl's law), but not the full outline.

What do the frequencies depend on if not shape alone?

Pitch scales with the wave speed c, set by tension and mass per area, through f = c\sqrt{\lambda}/2\pi. Change the tension and every frequency scales together. The shape controls only the ratios \sqrt{\lambda_n/\lambda_1}, which is why the presets are normalized to equal area.

Why does a drum sound un-pitched compared to a guitar string?

A string's overtones are integer multiples of its fundamental (2, 3, 4 and so on), which the ear fuses into one pitch. A membrane's ratios like 1.593 and 2.136 are irrational and inharmonic, so the modes do not fuse and you hear a thud rather than a note.

Why does hitting different spots change the sound if the spectrum is fixed?

A strike excites a mode in proportion to that mode's height at the strike point. Hit a nodal line and that mode stays silent. So the mix of loudnesses, the timbre, depends on where you strike, even though the available frequencies never change. Switch to "pure spectrum" to remove this and test only the pitches.

Is the disk special?

Yes. Among all shapes of a given area, the disk has the lowest possible fundamental frequency. That is the Faber-Krahn inequality. Any polygon of equal area rings higher at its first mode.