Where the equation comes from
The previous guide solved Legendre's equation, the master equation of anything with spherical symmetry. Now swap the sphere for a cylinder: a vibrating drumhead, the cross-section of an optical fibre, heat creeping inward through a round pipe. Set up the wave equation or heat equation on a disc, push it through separation of variables, and the angular part is easy (just sines and cosines around the circle) — but the radial part, the piece that depends on the distance r from the centre, keeps producing the same stubborn equation. That equation is Bessel's equation. Legendre is to spheres what Bessel's equation is to cylinders.
Written out, Bessel's equation of order n is x^2 y'' + x y' + (x^2 - n^2) y = 0, where x is essentially a scaled radius and the parameter n (the order) is whatever the angular separation handed you — it need not be a whole number. Notice the leading coefficient x^2: it vanishes at x = 0, the very centre of the drum. Divide through to put the equation in standard form y'' + P(x) y' + Q(x) y = 0 and you get P(x) = 1/x and Q(x) = 1 - n^2/x^2, both of which blow up at the origin. So x = 0 is a singular point — and it sits exactly where the physics lives, on the axis of the cylinder. A plain Taylor series cannot be trusted to start there.
Building the first solution with Frobenius
Since x = 0 is a regular singular point, a plain power series is the wrong ansatz — but the method of Frobenius is exactly right. Posit a series carrying an unknown leading power: y = x^r times (a_0 + a_1 x + a_2 x^2 + ...), with a_0 not zero and the exponent r to be discovered. Substitute this into x^2 y'' + x y' + (x^2 - n^2) y = 0 and look at the very lowest power of x. Because a_0 is not zero, the coefficient of that lowest power must vanish on its own, and it gives the indicial equation r^2 - n^2 = 0. Its two roots are r = +n and r = -n — the two possible leading exponents with which a solution can emerge from the singular centre.
Take the larger root r = +n first; it always yields a clean solution. Feeding it back in and demanding that every higher power of x also vanish produces a recurrence relation linking the coefficients: a_2m = -a_{2m-2} / (4 m (m + n)). Only even-indexed coefficients survive (the odd ones are all forced to zero), and each is the previous one divided by a tidy product. Crank the recurrence from a_0 and the pattern that tumbles out involves m! and the product (n+1)(n+2)...(n+m) in the denominator — a product that begs to be written with the Gamma function, since Gamma(m + n + 1) packages exactly that factorial-like growth even when n is not a whole number.
Bessel: x^2 y'' + x y' + (x^2 - n^2) y = 0, singular point x = 0
Frobenius ansatz: y = x^r * SUM_{m>=0} a_m x^m, a_0 != 0
lowest power of x -> indicial equation r^2 - n^2 = 0 -> r = +n, -n
take r = +n, match each power -> recurrence a_{2m} = - a_{2m-2} / (4 m (m+n))
(odd a's all zero)
choose a_0 = 1 / (2^n Gamma(n+1)) -> the standard first-kind solution
J_n(x) = SUM_{m>=0} (-1)^m / ( m! Gamma(m+n+1) ) * (x/2)^{2m+n}
leading term ~ (x/2)^n / Gamma(n+1) -> J_n(0)=0 for n>0, J_0(0)=1 (finite!)Fix the free constant a_0 to the conventional value 1/(2^n Gamma(n+1)) and the series has a name: J_n(x), the Bessel function of the first kind of order n. It is the radial analogue of cosine and sine for a world with a centre. Crucially, every term carries the factor x^n out front (the +n exponent we chose), so near the origin J_n(x) behaves like (x/2)^n / Gamma(n+1): for n > 0 it starts at zero and rises smoothly, while J_0(0) = 1. Either way it is finite at the centre — exactly the behaviour a real drumhead must have on its axis.
The second solution, and why it misbehaves
A second-order equation needs two independent solutions, so where is the other one? The indicial equation handed us a second root, r = -n. For many values of n you can run the same recurrence with that exponent and get a genuinely independent series J_{-n}(x). But there is a catch baked into the indicial roots themselves. The two roots are +n and -n, and their difference is 2n. When n is an integer (or a half-integer makes 2n an integer), that difference is a whole number — and the indicial-equation trichotomy warns that whenever the roots differ by an integer, the second Frobenius series can break down, and a logarithm may be forced into the solution. For integer order, J_{-n} turns out to be just a constant multiple of J_n, not independent at all — so the naive second series collapses.
The fix is to assemble a genuinely independent second solution by hand. Define Y_n(x), the Bessel function of the second kind (sometimes called the Neumann function), as a particular combination of J_n and J_{-n} engineered to survive the integer-order limit. The price of that engineering is a logarithm: near the origin Y_n(x) contains a term proportional to ln(x) (and, for n > 0, a piece blowing up like 1/x^n on top). So Y_n(x) is unbounded at the centre — it dives to minus infinity as x approaches zero. That logarithmic blow-up is the visible fingerprint of equal-or-integer-separated indicial roots, the exact case the Frobenius theory warned us about.
How they actually behave: decaying cosines
The series tells you everything near x = 0, but it is hopeless for picturing the function far out — you would need hundreds of terms. The far-field behaviour comes from a separate asymptotic analysis (the same large-argument machinery from the asymptotics rung), and it is wonderfully concrete. For large x, J_n(x) is approximately sqrt(2/(pi x)) times cos(x - n pi/2 - pi/4), and Y_n(x) is the same amplitude times sin(x - n pi/2 - pi/4). Read that out loud: a cosine (or sine) wave, shifted by a phase, whose amplitude fades like 1/sqrt(x). They are decaying cosines — exactly the picture of a circular ripple that oscillates as it travels outward while its height drops, because the same energy is spread around an ever-larger circle.
Be honest about what that approximation is. The large-x cosine formula is an asymptotic expansion, not a convergent series: keep adding correction terms and it will eventually diverge for any fixed x. Yet truncated at the right place it is astonishingly accurate, and the larger x is, the better the leading term alone does. This is the recurring lesson of advanced applied calculus — a divergent series, used wisely, can out-predict a convergent one. The power series for J_n is the exact tool near the origin; the asymptotic form is the exact tool far away; neither is 'the formula', and a real computation stitches them together.
Because both functions oscillate forever, each crosses zero infinitely many times, and those zeros are where the physics gets pinned down. Call the m-th positive zero of J_n by the name j_{n,m}: J_n(j_{n,m}) = 0. On a drum of radius R clamped at the rim, the boundary condition is that the displacement vanishes there, which forces the scaled radius to land on a zero — and that selects a discrete set of allowed vibration frequencies, the drum's overtones. Unlike a string, those overtones are not whole-number multiples of a fundamental (the zeros of J_0 are spaced unevenly), which is precisely why a drum sounds like a drum and not like a plucked string.
Solving a real drum, step by step
Let us put the pieces together on the original problem: a circular drumhead of radius R, clamped at its edge, vibrating in a perfectly circular (angle-independent) mode. The vertical displacement u depends on radius r and time t through the wave equation, and the whole job is to find which shapes and frequencies are allowed. Here is the full arc, from the partial differential equation down to a number you could measure.
- Separate variables. Write u(r, t) = F(r) G(t) and feed it into the wave equation in polar form. The time part gives G(t) = cos(omega t) (simple oscillation at frequency omega), and the radial part is forced to satisfy F'' + (1/r) F' + k^2 F = 0, where k = omega/c ties the spatial scale to the frequency.
- Recognize Bessel. Substitute x = k r to rescale, and the radial equation becomes x^2 F'' + x F' + x^2 F = 0 — Bessel's equation of order zero (n = 0, because the mode has no angular variation). Its two solutions are J_0(x) and Y_0(x).
- Demand finiteness at the centre. The domain includes r = 0, the middle of the drum, where the displacement must be finite. But Y_0 blows up logarithmically there, so its coefficient must be zero. Only F(r) = J_0(k r) survives — the geometry has thrown away half the solution space.
- Apply the clamped edge. The rim r = R is held still, so F(R) = 0, meaning J_0(k R) = 0. That forces k R to be one of the zeros of J_0: k R = j_{0,m} for m = 1, 2, 3, ... The allowed frequencies are omega_m = c j_{0,m} / R — a discrete spectrum, read straight off the zeros of a Bessel function.
- Build the general motion. Each m gives one mode J_0(j_{0,m} r/R) cos(omega_m t). The full vibration is a superposition of all of them, and the weights are fixed from the initial shape by a Fourier-Bessel series — the cylindrical cousin of an ordinary Fourier series, expanding the starting profile over the J_0 modes.
Step back and admire what just happened. A messy partial differential equation about a physical object dissolved, in five honest moves, into looking up the zeros of one special function. The quantization of the drum's pitches — that there is a lowest tone and a discrete ladder above it — was not imposed; it fell out of the single requirement that the solution stay finite at the centre and vanish at the rim. That is the deep payoff of this whole rung: the famous functions of physics are not exotic, they are simply the natural shapes that survive when a series solution must obey the geometry it lives in.