The Method of Frobenius

Where the plain power series breaks down

In the previous guide you used the [[power-series-method|power-series method]] with confidence: assume y = sum of a_n x^n, feed it into the equation, and grind out a recurrence for the coefficients. That works beautifully — but only at an [[ordinary-and-singular-points|ordinary point]], a place where the equation's coefficient functions are well behaved. The trouble is that the most important equations of physics misbehave at exactly the point we most want to expand around, usually x = 0. Try a plain power series there and it either collapses or simply cannot represent the true solution.

Take the simplest cautionary example, the Cauchy-Euler equation x^2 y'' - 2y = 0. You may recall from the Cauchy-Euler equation that its solutions are pure powers: y = x^2 and y = x^{-1}. Now the x^2 is a perfectly ordinary power series, but x^{-1} is not — no sum of a_0 + a_1 x + a_2 x^2 + ... can ever produce a 1/x. A plain power series is structurally blind to this second solution. The equation is telling us the solution wants to start with a negative power, and our ansatz forbids it.

So the fix almost writes itself. If the solution insists on starting with x to some power that is not a non-negative integer, then build that power into the ansatz from the start. Multiply a normal power series by a leading factor x^r, where r is an unknown exponent we will solve for. That single idea — letting the series start at a flexible power x^r — is the entire [[frobenius-method|method of Frobenius]].

Regular singular points: damage you can repair

Not every singular point is salvageable, so we must be honest about the dividing line. Write the equation in standard form y'' + P(x) y' + Q(x) y = 0. A point x_0 is a [[regular-singular-point|regular singular point]] if P or Q does blow up there, but only mildly: (x - x_0) P(x) and (x - x_0)^2 Q(x) must both stay finite — in fact analytic — at x_0. In words, P is allowed to diverge no worse than 1/(x - x_0) and Q no worse than 1/(x - x_0)^2. If the singularity is fiercer than that, the point is irregular, and Frobenius offers no guarantee.

With a regular singular point in hand, the Frobenius ansatz is y = x^r times (sum of a_n x^n for n from 0 to infinity), with a_0 not zero by convention — the leading coefficient must be nonzero, otherwise we have just relabeled r. Differentiate term by term, exactly as you did for an ordinary power series in Volume I, except now each term carries the extra exponent r: the n-th term a_n x^{n+r} has derivative (n+r) a_n x^{n+r-1}, and second derivative (n+r)(n+r-1) a_n x^{n+r-2}. The exponent r rides along through every differentiation, and that is what will let it fall out of the lowest-power balance.

The indicial equation fixes the leading exponent

Here is the pivotal move. After substituting the Frobenius series into the equation, collect terms by power of x and demand that every coefficient vanish — that is just the series method's rule again, that a power series is zero only if all its coefficients are. Now look at the single lowest power present, the term in x^{r-2} or x^r depending on how you arrange it. It is built entirely from a_0, and since a_0 is not allowed to be zero, the bracket multiplying it must be zero by itself. That bracket, set to zero, is the [[indicial-equation|indicial equation]].

The indicial equation is always a simple quadratic in r, of the form r(r-1) + p_0 r + q_0 = 0, where p_0 and q_0 are the values at x_0 of those tamed combinations x P(x) and x^2 Q(x). It is the singular-point cousin of the characteristic equation you solved for constant-coefficient ODEs: there a quadratic in a growth rate selected the exponentials, here a quadratic in r selects the leading powers. Its two roots r_1 and r_2 (label them so r_1 is greater than or equal to r_2) are called the exponents of the singularity, and they are the only candidates for how the solutions may begin.

y = x^r ( a_0 + a_1 x + a_2 x^2 + ... ),   a_0 =/= 0

Standard form:  y'' + P y' + Q y = 0
Regular singular point at 0:   p0 = lim_{x->0} x P(x)
                               q0 = lim_{x->0} x^2 Q(x)

Lowest power of x  ->  only a_0 survives  ->

   INDICIAL EQUATION:   r(r-1) + p0 r + q0 = 0

Roots r1 >= r2  =  the two exponents of the singularity.
Higher powers give the recurrence that fixes a_1, a_2, ...

Three faces of the second solution

The larger root always works cleanly. Feeding r = r_1 into the higher-power equations gives a recurrence that determines a_1, a_2, ... in turn, and you get one honest Frobenius solution y_1 = x^{r_1} times a power series. The drama is entirely about the second solution, and which of three forms it takes is decided by the difference r_1 - r_2. That single number is the whole forecast.

1. Case 1 — the roots differ by something other than an integer (r_1 - r_2 is not a whole number). Then both roots behave, and you get two clean independent Frobenius series, y_1 = x^{r_1}(...) and y_2 = x^{r_2}(...). The general solution is their combination; you are done. (Equal roots count as the integer 0, so they fall under Case 2, not here.)
2. Case 2 — the roots are equal (r_1 = r_2). One Frobenius series simply cannot supply a second independent solution — there is only one exponent to start at. The missing partner is forced to carry a logarithm: y_2 = y_1 ln(x) + x^{r_1}(a new series). The log is not a decoration; it is the unavoidable signature of a repeated exponent, exactly as a repeated characteristic root forced an extra factor of x in the constant-coefficient world.
3. Case 3 — the roots differ by a positive integer (r_1 - r_2 = N, a whole number). This is the treacherous case. When you try to build the series from the smaller root r_2, the recurrence hits a step where it would divide by zero exactly at term N — the very place the larger root's series begins. Sometimes that zero is forgiving and you still get a pure series; sometimes it forces a logarithm just like Case 2, with y_2 = c y_1 ln(x) + x^{r_2}(a series), where the constant c may turn out to be zero.

Bessel's equation: the method in the flesh

Let the whole machine run on the equation that practically invented the method, the Bessel equation x^2 y'' + x y' + (x^2 - nu^2) y = 0. In standard form P(x) = 1/x and Q(x) = (x^2 - nu^2)/x^2, so x P = 1 and x^2 Q = x^2 - nu^2, both finite at zero: x = 0 is a textbook regular singular point. The indicial equation is r(r-1) + 1 times r - nu^2 = 0, which simplifies to r^2 = nu^2. The two exponents are therefore r_1 = nu and r_2 = -nu — and their difference 2 nu is exactly the dial that selects which of the three cases we land in.

Watch all three cases appear from one equation just by changing nu. If nu = 1/2, the difference 2 nu = 1 is an integer (Case 3), yet the singularity is forgiving and the solutions turn out to be the elementary sin(x)/sqrt(x) and cos(x)/sqrt(x) — pure series, no log. If nu is, say, 1/3, the difference 2/3 is not an integer (Case 1) and you get two clean independent series, the Bessel functions of order plus and minus 1/3. But if nu = 0 the roots coincide (Case 2), and the inevitable logarithm appears: the second solution, the Bessel function of the second kind Y_0(x), carries a ln(x) term and dives to minus infinity at the origin.

This is why a vibrating circular drumhead is finite at its center but a point heat source is not: the physically allowed solution is the one that stays bounded, the Bessel function of the first kind, and the logarithmic Y_0 is discarded by the boundary condition. The lesson generalizes far past Bessel. Run Frobenius on the same template with different coefficients and you generate the named functions of physics one after another. The method is not a trick for one equation; it is the forge in which a whole catalog of special functions is hammered out, each one the Frobenius solution of its own singular equation.