The rectangle was never special — the coordinates were
Across this rung you separated the heat equation and Laplace's equation on intervals and rectangles, and every time the spatial factor came back a sine or a cosine. It is tempting to conclude that PDEs simply love trigonometric functions. They do not. The sines showed up because the boundary was a straight segment and the operator was plain d^2/dx^2 — and on that geometry, sines are the shapes that vanish at clamped ends while staying eigenfunctions. Change the shape of the room, and the room's favourite shapes change with it.
Most real domains are not rectangles. A drumhead is a disk; the air in a pipe wraps a cylinder; the gravitational and electric potentials around a star or a charge live outside a sphere. To separate variables there, the first job is to rewrite the Laplacian in coordinates that fit the boundary, so that the boundary becomes a single coordinate held fixed. On a disk that means polar coordinates (r, theta), where the rim is just r = a. The price is that the operator stops looking like u_xx + u_yy and grows new, r-dependent terms — and those terms are exactly what conjure the new special functions.
On a disk: the angle still gives sines, the radius gives something new
In polar coordinates the Laplacian becomes Laplacian u = u_rr + (1/r) u_r + (1/r^2) u_(theta theta). Feed in the product u = R(r) T(theta) and divide through, and the same keystone argument from guide 1 fires again: a function of r alone equals a function of theta alone only if both equal a constant. The angular piece gives T'' + n^2 T = 0. But notice the boundary condition on theta is not a clamp — it is that you must return to where you started after one full turn, the periodic boundary condition T(theta + 2*pi) = T(theta). That periodicity forces n to be a whole number, and T(theta) = cos(n theta) or sin(n theta). So the angle still speaks Fourier — a Fourier series around the circle.
The radial equation is where the geometry bites. For Laplace's equation it reads r^2 R'' + r R' - n^2 R = 0, an equidimensional equation solved by the harmonic building blocks R = r^n (and r^(-n), thrown out because it blows up at the centre). That is why a steady temperature on a disk decays toward the middle like r^n. But for a vibrating drum — the wave equation, where time separation leaves behind the Helmholtz equation Laplacian u + k^2 u = 0 — the radial equation picks up a k^2 r^2 term and becomes r^2 R'' + r R' + (k^2 r^2 - n^2) R = 0. That is Bessel's equation, and its well-behaved solution is the Bessel function J_n(k r), not a sine or a cosine at all.
Polar Laplacian: u_rr + (1/r) u_r + (1/r^2) u_(theta theta)
Guess u = R(r) T(theta) :
angle -> T'' + n^2 T = 0 periodic -> cos(n*theta), sin(n*theta), n = 0,1,2,...
radius -> Laplace: r^2 R'' + r R' - n^2 R = 0 -> R = r^n
Helmholtz: r^2 R'' + r R' + (k^2 r^2 - n^2) R = 0 -> R = J_n(k r) (Bessel)Bessel functions are sines that have been to the gym
Do not let the unfamiliar name frighten you: a Bessel function J_n(r) is morally a sine that has adapted to a spreading, two-dimensional world. It oscillates and crosses zero over and over, just like a sine, but each successive hump is a little shorter and the zeros are not evenly spaced, because energy spreading outward over a growing circumference must thin out. Clamping the drum's rim, R(a) = 0, demands that k a land on a zero of J_n — and J_n has an infinite ladder of zeros, so once again a homogeneous boundary condition selects a discrete sequence of allowed k values. Those are the drum's natural frequencies, the reason a drum sounds different from a string.
And the punchline ties the whole rung together. Just as you expanded an arbitrary starting profile in sines, you can expand an arbitrary radial profile in Bessel functions — a Bessel-function expansion — and the coefficients come out by the very same trick of guide 3: orthogonality. The Bessel functions for a fixed n are mutually orthogonal on the disk, but with a twist worth flagging: the orthogonality carries a weight of r, so the inner product integrates R_1(r) R_2(r) times r dr, not just dr. That r is the radial part of the area element, and it is the signature of a Sturm-Liouville problem in disguise — the unifying framework the next rung makes explicit.
On a sphere: latitude, longitude, and spherical harmonics
Step up to three dimensions with a spherical boundary and the same story plays out one dimension richer. In spherical coordinates (r, theta, phi) — radius, polar angle (latitude), azimuth (longitude) — the product ansatz becomes u = R(r) Y(theta, phi). Separation peels the radial factor R off cleanly, leaving the angular part Y to satisfy its own eigenvalue problem on the surface of the sphere. The natural eigenfunctions of that surface problem are the spherical harmonics Y_l^m(theta, phi), and they are the genuine three-dimensional analogue of sin(n x): the irreducible vibrational patterns of a sphere.
Inside the angular part, one more separation splits longitude from latitude. The azimuth phi is periodic, so it gives the familiar e^(i m phi), Fourier once more. The polar angle theta is the deep one: after the substitution x = cos(theta) its equation becomes the Legendre equation, and the solutions that stay finite at the poles are the Legendre polynomials P_l(x) (and their associated cousins). Just like sines and Bessel functions, the Legendre polynomials are orthogonal — here on the interval [-1, 1] — so a function on the sphere expands as a spherical-harmonic expansion, coefficients again read off by orthogonality. Every special function in this guide is the same idea wearing a different coat.
One method, many alphabets — and where it still stops
Pull back and see the unity. The recipe never changed across this rung — only the room did, and with it the natural alphabet. The procedure is identical every time, and each special function is just the spatial eigenfunction the chosen geometry happens to prefer.
- Rewrite the Laplacian in coordinates fitted to the boundary, so the boundary becomes one coordinate held fixed.
- Insert the product ansatz and divide, letting the equation separate into one ODE per coordinate, joined by a separation constant.
- Impose the homogeneous boundary and finiteness conditions; they quantize the constant and select a discrete ladder of eigenfunctions (sines, Bessel functions, Legendre polynomials).
- Superpose the modes and fix the coefficients by orthogonality — with the right weight — to match the initial or boundary data.
Be honest about the boundaries of this beautiful machine, exactly as in guide 1. It still needs a linear equation, because superposition is the load-bearing pillar and a nonlinear term lets modes interfere rather than add. It still needs homogeneous conditions along the separated directions — a nonzero boundary value or a source must first be subtracted by shifting to a steady state. And it still needs a separable geometry: there is no coordinate system that tidies the Laplacian on a generic blob, so a kidney-shaped membrane or an L-shaped plate yields no clean product solution. For those domains we hand off to the more general machinery — Sturm-Liouville theory to certify that the eigenfunctions are complete and orthogonal, eigenfunction expansions computed numerically, and the Fourier series idea generalized to a generalized Fourier series in whatever basis the operator provides.