JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

The Laplacian in Curved Coordinates

Why does a vibrating drum summon Bessel functions and a charged sphere summon Legendre polynomials? Because when you write Laplace's equation in the coordinates that fit the shape and then separate the variables, the leftover ordinary equations ARE the named equations of the earlier rungs — the geometry chooses the special functions for you.

The puzzle this guide solves

Throughout this rung you have met the three great linear PDEs — the heat equation, the wave equation, and the steady-state Laplace equation nabla^2 u = 0 — and you tamed them on rectangles by separation of variables, which produced ordinary sine and cosine modes and a Fourier series. But the world is rarely a rectangle. A drumhead is a disk, a heated wire is a cylinder, a planet's gravitational field lives outside a sphere. The honest question is: what happens to the same method when the domain is round?

Here is the punchline you can hold onto for the whole guide. Separation of variables still works — but only because we first rewrite the Laplacian in coordinates that match the boundary, and when we do, the leftover one-variable equations are no longer the gentle y'' + lambda y = 0. They are the Bessel equation and the Legendre equation you studied back in the Series-Solutions and Special-Functions rungs. The round geometry does not merely complicate the algebra; it reaches into your toolbox and selects exactly which special functions are allowed to live on that shape. By the end you will see why a disk forces Bessel functions and a sphere forces Legendre polynomials, and the choice will feel inevitable rather than memorized.

What the Laplacian becomes when space curves

First, what IS the Laplacian, intuitively? In Cartesian coordinates nabla^2 u = d^2u/dx^2 + d^2u/dy^2 + d^2u/dz^2, a sum of partial second derivatives. But that formula is the local face of a coordinate-free idea: nabla^2 u measures how much u at a point differs from the average of u on a tiny sphere around it. If a point is cooler than its surroundings, nabla^2 u is positive there and heat will flow in; Laplace's equation nabla^2 u = 0 is precisely the statement 'every point already equals the average of its neighbours' — a perfectly relaxed, harmonic field with no hot spots or cold spots left to even out. That averaging picture is the same in any coordinate system; only the bookkeeping changes.

Why must the bookkeeping change at all? Because in curved coordinates the grid lines are not evenly spaced. Step one tick in the angle theta near the centre of a disk and you barely move; take the same one-tick step near the rim and you sweep a long arc. The conversion factors that record how far a real distance corresponds to one tick of each coordinate are the scale factors, and they are the whole story. For polar coordinates (r, theta) the scale factors are h_r = 1 and h_theta = r: moving in r is honest distance, but a step in theta covers an arc of length r·d(theta), longer where r is bigger. The Laplacian must divide by these factors to undo the stretching of the grid.

The disk: how the 1/r terms breed the Bessel equation

Take the flattest interesting case: Laplace's (or Helmholtz's) equation on a disk, written in polar coordinates. Plugging the scale factors h_r = 1, h_theta = r into the recipe gives nabla^2 u = d^2u/dr^2 + (1/r) du/dr + (1/r^2) d^2u/d(theta)^2. Those two extra pieces with 1/r and 1/r^2 out front are the fingerprint of curvature; they are absent in the flat Cartesian form and they are about to do all the work. They are not errors or approximations — they are the exact price of the angle theta covering more ground the farther out you go.

Now separate: guess a product u(r, theta) = R(r) Theta(theta), substitute, and divide through by R Theta. The angle and the radius cleanly split onto opposite sides of an equals sign, which can only happen if both sides equal the same constant — call it m^2. The angular half gives Theta'' + m^2 Theta = 0, the same friendly oscillator as on a rectangle, with solutions cos(m theta) and sin(m theta). But here the geometry adds a demand the rectangle never made: going once around the disk, theta to theta + 2pi, returns you to the very same physical point, so Theta must repeat with period 2pi. That single-valuedness forces m to be a whole number 0, 1, 2, 3, ... — the integer is handed to you by the topology of the circle, not chosen by hand.

The radial half is where the named function is born. After replacing the angular derivative by -m^2 and (for the vibrating drum) carrying a frequency constant k^2, the radius obeys R'' + (1/r) R' + (k^2 - m^2/r^2) R = 0. Substitute x = k r and this is, character for character, the Bessel equation of order m: x^2 R'' + x R' + (x^2 - m^2) R = 0. You did not import Bessel's equation from outside; the 1/r and 1/r^2 left over from the curved Laplacian, plus the integer m supplied by the circle, assembled it in front of you. Its bounded solution is the Bessel function J_m(k r), and demanding that the drum be clamped at the rim — R(a) = 0 — forces k a to be a zero of J_m, which is precisely how the drum's allowed frequencies (its overtones) get quantized.

Laplacian on a disk (polar):  nabla^2 u = u_rr + (1/r) u_r + (1/r^2) u_thetatheta

Separate   u(r,theta) = R(r) Theta(theta) ,  divide by R Theta:

     R''/R + (1/r) R'/R  +  (1/r^2) Theta''/Theta  =  -k^2
     |__________ depends on r only _________|   |__ theta only __|

  =>  Theta'' + m^2 Theta = 0      (period-2pi  =>  m = 0,1,2,...)
  =>  R'' + (1/r) R' + (k^2 - m^2/r^2) R = 0

Let x = k r :   x^2 R'' + x R' + (x^2 - m^2) R = 0   <-- BESSEL of order m
     bounded solution  R = J_m(k r) ;  clamp R(a)=0  =>  k a = zero of J_m
Separation on a disk: the angular factor's 2pi-periodicity quantizes m to integers, and the radial factor turns out to be the Bessel equation of that same order m.

The sphere: how cos(theta) breeds the Legendre equation

Climb up to three dimensions and choose spherical coordinates (r, theta, phi), with theta the polar angle from the north pole and phi the longitude. The scale factors are h_r = 1, h_theta = r, h_phi = r sin(theta), and feeding them through the recipe produces a Laplacian with three nested pieces: a radial part, a polar-angle part carrying a 1/sin(theta) factor, and an azimuthal 1/(r^2 sin^2(theta)) d^2u/d(phi)^2. The lone sin(theta) is the new fingerprint of curvature — it is the squeezing of the longitude lines as they crowd together toward the poles, the same crowding that makes a flat world map stretch Greenland.

Separate in three stages: u = R(r) Theta(theta) Phi(phi). The longitude factor splits off first and gives Phi'' + m^2 Phi = 0 with, once again, 2pi-periodicity forcing m to be an integer — the same circular topology as the disk, now acting on longitude. The radius factor separates next and gives an equidimensional equation r^2 R'' + 2r R' - l(l+1) R = 0 whose solutions are the plain powers r^l and r^{-(l+1)} — the familiar fields that grow or decay like powers of distance, the second being exactly the 1/r potential of a point charge or mass when l = 0.

The polar angle is where the second named function appears. The Theta equation still carries that awkward 1/sin(theta), but a single inspired substitution clears it: let x = cos(theta), so the pole-to-pole sweep theta from 0 to pi becomes x running over the tidy interval -1 to 1. Under that change the messy trigonometric operator collapses into d/dx[(1 - x^2) dTheta/dx] + [l(l+1) - m^2/(1 - x^2)] Theta = 0 — the associated Legendre equation. With no longitude dependence (m = 0) it is the plain Legendre equation (1 - x^2) y'' - 2x y' + l(l+1) y = 0. The geometry has, once again, chosen the function for you: the demand that the solution stay finite at the poles x = +1 and x = -1 forces l to be a non-negative integer, and the only solutions that survive are the Legendre polynomials P_l(cos theta).

Why this had to happen: the deeper pattern

Step back and the pattern is the same in every coordinate system. Separation peels the multivariable PDE into one ordinary differential equation per coordinate, and each of those single-variable equations is a Sturm-Liouville problem — a self-adjoint eigenvalue problem on an interval, exactly the structure that occupied an entire rung of this ladder. On the rectangle the interval is plain and the Sturm-Liouville problem is the trivial y'' + lambda y = 0, so the eigenfunctions are sines and cosines. On the disk and the sphere the interval comes with the curved coordinate's weight built in, and the same Sturm-Liouville machine spits out Bessel functions and Legendre polynomials instead. Different geometry, same machine, different named eigenfunctions.

That Sturm-Liouville pedigree is not a footnote — it is what makes the whole method usable. Because each radial or angular problem is self-adjoint, its eigenfunctions are orthogonal and complete, so you can expand any reasonable boundary data as a Bessel-Fourier series in J_m or a Legendre series in P_l, just as you expand on a rectangle with an ordinary Fourier series. The recipe is identical in spirit: separate, find the eigenfunctions, then fix the leftover coefficients by matching the boundary condition through orthogonality. The special functions are simply the right alphabet for round boundaries, and orthogonality is the dictionary that lets you read any function in that alphabet.

  1. Pick coordinates that fit the boundary — polar/cylindrical for round-and-straight shapes, spherical for ball-shaped ones — so the boundary becomes a single coordinate held constant (r = a, say).
  2. Rewrite nabla^2 in those coordinates using the scale factors, then substitute a product u = R(r) Theta(theta) Phi(phi) and divide through to split the variables.
  3. Solve the easy periodic angular pieces first; their single-valuedness quantizes the separation constants (m, l) to integers, which then feed into the radial/polar equations.
  4. Recognize the leftover equation as Bessel or Legendre, keep only the solution that stays finite where it must, and assemble the answer as an orthogonal series whose coefficients are fixed by the boundary data.

Honest caveats, and where the magic stops

Be clear about what made all of this work, because it is fragile. Separation of variables is not a general solver; it succeeds only when the geometry and the equation cooperate. The boundary has to be a level surface of a coordinate (a circle r = a, a sphere r = a), and the coordinate system has to be one of the lucky handful — there are exactly eleven coordinate systems in which the three-dimensional Laplacian separates, no more. Skew the domain into an ellipse or an off-centre annulus and the variables refuse to part; the clean product u = R(r)Theta(theta) is then simply false, and you must reach for numerical methods or a different attack entirely. The named functions are a gift of symmetry, and only of symmetry.

Two more honest points. First, each radial equation has two independent solutions, and we keep only one — the one that does not blow up where the domain reaches. On the solid disk we discard the second-kind Bessel function Y_m because it diverges at the centre r = 0; on a sphere we drop r^{-(l+1)} when the region includes the origin. That discarding is a physical decision (a real drumhead is not infinite at its middle), not a mathematical reflex — and it flips for an exterior problem, where you instead keep the decaying solution and throw away the one that grows at infinity. Always ask which solution your actual domain can tolerate.