Regular Singular Points and Frobenius

What happens to a linear differential equation at a point where its coefficients blow up? Near an ordinary point, solutions are as smooth as the equation itself. Power series work without complication there. But equations like Bessel's, Legendre's, and the hypergeometric equation all have points where a leading coefficient vanishes. At these points, the naive power-series method fails outright. The astonishing fact, discovered by Ferdinand Georg Frobenius in 1873, is that even at these singular points a controlled failure occurs. As long as the singularity is not too severe, generalized power series still converge. This holds provided one allows a fractional or negative leading exponent. In a degenerate case, a logarithm is also required. This matters because most of the special functions of mathematical physics arise this way. Bessel functions, Legendre functions, Laguerre and Hermite polynomials, and hypergeometric functions are examples. They arise exactly as Frobenius series solutions near a regular singular point. This occurs for a second-order linear ODE. Understanding this theory converts an intimidating list of "special functions" into instances of a single, unified construction. The big picture is a trichotomy. At an ordinary point, ordinary power series work. At a regular singular point, the singularity is "mild." Coefficients have at worst simple/double poles after normalizing. A modified series method — Frobenius' method — still produces solutions there. This method is governed by an algebraic indicial equation. Its roots dictate the exponents of the series. At an irregular singular point, even this fails. One then needs asymptotic or more delicate tools. These lie beyond the present scope. This article makes the classification of singular points precise. It proves the existence of Frobenius series solutions under the regular singular point hypothesis. It works out the three qualitatively different cases governed by the indicial roots. It demonstrates the method on genuinely asymmetric examples. Every algebraic step is made visible and checkable.

Key takeaways
  • Near an ordinary point, solutions are analytic; near a singular point they may not be, unless the singularity is regular.
  • A point x0x_0 is a regular singular point of y′′+p(x)y′+q(x)y=0y'' + p(x)y' + q(x)y = 0 if (x−x0)p(x)(x-x_0)p(x) and (x−x0)2q(x)(x-x_0)^2 q(x) are analytic there.
  • The indicial equation r(r−1)+p0r+q0=0r(r-1) + p_0 r + q_0 = 0 gives exponents r1≥r2r_1 \ge r_2 that start the Frobenius series.
  • Three cases by r1−r2r_1 - r_2: distinct roots not differing by an integer, equal roots, and roots differing by a positive integer (the last two may force a logarithmic term).
  • The method converts singular ODEs (Bessel, Legendre-type) into solvable recursions for series coefficients.
  • Key idea: at a regular singular point, multiplying by (x−x0)r(x-x_0)^r tames the singularity enough that Frobenius series still converge.

Where this lives today

The Frobenius method is the engine behind essentially every classical special function used in physics and engineering: Bessel functions (wave propagation in cylindrical geometry, heat conduction in cylinders, quantum mechanics of a particle in a cylindrical well), Legendre functions (angular momentum eigenfunctions, gravitational and electrostatic multipole expansions), and Laguerre/Hermite-type equations (radial hydrogen atom wavefunctions, quantum harmonic oscillator). In complex analysis, regular singular points of linear ODEs are exactly the points where a Fuchsian differential equation can have nontrivial local monodromy, connecting Frobenius' construction to the study of the Wronskian and to global questions about analytic continuation of solutions around singularities. In numerical analysis and applied mathematics, recognizing a regular singular point tells a practitioner exactly what kind of coordinate transformation or series ansatz will regularize a numerical solver near a coordinate singularity (e.g., the origin in polar or spherical coordinates). The indicial equation also appears, in disguise, in the study of asymptotic expansions of integrals and in the theory of hypergeometric-type equations underlying orthogonal polynomials used throughout statistics and numerical approximation. Understanding this material builds directly on the ideas from Series Solutions Near Ordinary Points and prepares one to recognize the structure of the classical special functions wherever they arise.

Ask the tutor: How do I determine, without fully solving, whether the log term actually appears when the indicial roots differ by a positive integer?

DefinitionOrdinary, singular, and regular singular points

Consider the linear second-order ODE, written in standard form,

y′′+p(x) y′+q(x) y=0,y'' + p(x)\,y' + q(x)\,y = 0,

or equivalently in the form

P(x) y′′+Q(x) y′+R(x) y=0,P(x)\,y'' + Q(x)\,y' + R(x)\,y = 0,

where P,Q,RP, Q, R are analytic at x0x_0 with p=Q/Pp = Q/P, q=R/Pq = R/P.

  • x0x_0 is an ordinary point if pp and qq are analytic at x0x_0 (equivalently P(x0)≠0P(x_0)\ne 0).
  • x0x_0 is a singular point if it is not ordinary (equivalently P(x0)=0P(x_0) = 0, assuming Q,RQ,R have no common zero canceling it).
  • A singular point x0x_0 is a regular singular point if both
(x−x0) p(x)and(x−x0)2 q(x)(x - x_0)\,p(x) \quad \text{and} \quad (x-x_0)^2\, q(x)

are analytic at x0x_0 (i.e., have convergent Taylor series in a neighborhood of x0x_0).

  • A singular point that is not regular is called an irregular singular point.

Equivalently, writing p0(x)=(x−x0)p(x)p_0(x) = (x-x_0)p(x) and q0(x)=(x−x0)2q(x)q_0(x) = (x-x_0)^2 q(x) as their Taylor series

p0(x)=∑n=0∞pn(x−x0)n,q0(x)=∑n=0∞qn(x−x0)n,p_0(x) = \sum_{n=0}^\infty p_n (x-x_0)^n, \qquad q_0(x) = \sum_{n=0}^\infty q_n (x-x_0)^n,

the indicial equation at the regular singular point x0x_0 is the quadratic

F(r):=r(r−1)+p0 r+q0=0,F(r) := r(r-1) + p_0\, r + q_0 = 0,

whose roots r1≥r2r_1 \ge r_2 (real parts, if complex) are the exponents of singularity.

Intuition

Think of p(x)p(x) and q(x)q(x) as measuring how strongly the equation "feels" the point x0x_0. At an ordinary point they are perfectly finite there, and a solution can be built as an ordinary Taylor series ∑an(x−x0)n\sum a_n(x-x_0)^n, with the recursion for ana_n coming straight from matching coefficients. At a singular point, pp or qq blows up, and naively substituting a power series produces a contradiction — the x0x_0-th coefficient equation may force a0=0a_0 = 0, then a1=0a_1 = 0, and so on, or may simply not close. The key insight is how badly they blow up. If pp has at worst a simple pole and qq at worst a double pole at x0x_0 — precisely the regular singular point condition — then multiplying through by (x−x0)2(x-x_0)^2 clears the singularity into a well-behaved, analytic equation. This suggests trying a solution of the form (x−x0)r(x-x_0)^r times an ordinary power series, where rr need not be a nonnegative integer. Substituting y=(x−x0)r∑an(x−x0)ny = (x-x_0)^r\sum a_n (x-x_0)^n into the ODE and collecting the lowest order term forces a0≠0a_0 \ne 0 to satisfy exactly the indicial equation F(r)=0F(r) = 0: a quadratic condition on the allowed leading exponent. Each root rr of this quadratic is a candidate starting exponent for a genuine series solution, and the recursion for higher coefficients closes up provided F(r+n)≠0F(r+n) \ne 0 for n=1,2,…n = 1, 2, \dots — a condition that can fail exactly when the two roots differ by a positive integer.

Figure
Classification of points for x2y′′+x y′+(x2−ν2)y=0x^2 y'' + x\,y' + (x^2 - \nu^2) y = 0 (Bessel's equation), singular at x=0x=0
x = 0 (singular) ordinary ordinary p(x) = 1/x, q(x) = 1 - ν²/x² x·p(x) = 1 (analytic at 0) x²·q(x) = x² - ν² (analytic at 0) ⟹ x = 0 is a REGULAR singular point Indicial eq: r(r−1) + r − ν² = 0 ⟹ r = ±ν

The Frobenius construction, precisely

We now make the intuition above into a theorem. Throughout, fix a regular singular point x0x_0 and, without loss of generality (by translating), take x0=0x_0 = 0. Write

x2y′′+x[x p(x)]y′+[x2q(x)]y=0,x^2 y'' + x\big[x\,p(x)\big] y' + \big[x^2 q(x)\big] y = 0,

so that with p0(x)=xp(x)=∑pnxnp_0(x) = x p(x) = \sum p_n x^n and q0(x)=x2q(x)=∑qnxnq_0(x) = x^2 q(x) = \sum q_n x^n analytic on ∣x∣<ρ|x| < \rho for some ρ>0\rho > 0, the equation reads

x2y′′+x p0(x) y′+q0(x) y=0.(∗)x^2 y'' + x\, p_0(x)\, y' + q_0(x)\, y = 0. \tag{$\ast$}
TheoremFrobenius' Theorem

Let x=0x = 0 be a regular singular point of (∗)(\ast), with p0(x)=∑n≥0pnxnp_0(x) = \sum_{n\ge0} p_n x^n and q0(x)=∑n≥0qnxnq_0(x)=\sum_{n\ge0} q_n x^n convergent for ∣x∣<ρ|x| < \rho. Let

F(r)=r(r−1)+p0r+q0F(r) = r(r-1) + p_0 r + q_0

be the indicial polynomial, with roots r1,r2r_1, r_2 (indexed so that Re⁡r1≥Re⁡r2\operatorname{Re} r_1 \ge \operatorname{Re} r_2; assume both real for concreteness).

(1) For the larger root r1r_1, there is a solution of (∗)(\ast) of the form

y1(x)=xr1∑n=0∞anxn,a0=1,y_1(x) = x^{r_1}\sum_{n=0}^\infty a_n x^n, \qquad a_0 = 1,

convergent for 0<x<ρ0 < x < \rho, with the ana_n determined uniquely by a recursion coming from (∗)(\ast).

(2) If r1−r2r_1 - r_2 is not a nonnegative integer, there is a second, independent solution of the same form with r1r_1 replaced by r2r_2:

y2(x)=xr2∑n=0∞bnxn,b0=1,y_2(x) = x^{r_2}\sum_{n=0}^\infty b_n x^n, \qquad b_0 = 1,

also convergent for 0<x<ρ0 < x < \rho.

(3) If r1=r2=:rr_1 = r_2 =: r, a second, independent solution has the form

y2(x)=y1(x)ln⁡x+xr∑n=1∞cnxn.y_2(x) = y_1(x)\ln x + x^{r}\sum_{n=1}^\infty c_n x^n .

(4) If r1−r2=Nr_1 - r_2 = N is a positive integer, a second, independent solution has the form

y2(x)=C y1(x)ln⁡x+xr2∑n=0∞cnxn,c0=1,y_2(x) = C\, y_1(x)\ln x + x^{r_2}\sum_{n=0}^\infty c_n x^n, \qquad c_0 = 1,

for some constant CC (possibly 00). In all cases the series involved converge for 0<x<ρ0 < x < \rho, and {y1,y2}\{y_1, y_2\} is a fundamental set of solutions on 0<x<ρ0 < x < \rho.

Proof

We give the proof by direct substitution and coefficient-matching (the classical method of Frobenius), justifying convergence via a majorant (comparison) series argument. This is a self-contained argument; it uses only the ordinary-point Series Solutions Near Ordinary Points existence machinery as an auxiliary tool for the analytic companion equation, not for the singular equation itself, so there is no circularity.

Step 1: Setting up the ansatz and indicial equation. Seek y=xr∑n=0∞anxny = x^r \sum_{n=0}^\infty a_n x^n with a0≠0a_0 \ne 0, formally differentiating term by term:

y′=∑n=0∞(n+r)anxn+r−1,y′′=∑n=0∞(n+r)(n+r−1)anxn+r−2.y' = \sum_{n=0}^\infty (n+r) a_n x^{n+r-1}, \qquad y'' = \sum_{n=0}^\infty (n+r)(n+r-1) a_n x^{n+r-2}.

Substitute into (∗)(\ast), using p0(x)=∑pkxkp_0(x) = \sum p_k x^k, q0(x)=∑qkxkq_0(x) = \sum q_k x^k:

∑n=0∞(n+r)(n+r−1)anxn+r+(∑k=0∞pkxk)(∑n=0∞(n+r)anxn+r)+(∑k=0∞qkxk)(∑n=0∞anxn+r)=0.\sum_{n=0}^\infty (n+r)(n+r-1) a_n x^{n+r} + \Big(\sum_{k=0}^\infty p_k x^{k}\Big)\Big(\sum_{n=0}^\infty (n+r) a_n x^{n+r}\Big) + \Big(\sum_{k=0}^\infty q_k x^k\Big)\Big(\sum_{n=0}^\infty a_n x^{n+r}\Big) = 0.

Divide by xrx^r and collect the coefficient of xnx^n (n≥0n \ge 0) using the Cauchy product:

(n+r)(n+r−1)an+∑k=0npk(n−k+r)an−k+∑k=0nqkan−k=0.(n+r)(n+r-1)a_n + \sum_{k=0}^{n} p_k (n-k+r) a_{n-k} + \sum_{k=0}^n q_k a_{n-k} = 0.

Separate the k=0k=0 term (which involves ana_n itself) from the rest:

[(n+r)(n+r−1)+p0(n+r)+q0]an=−∑k=1n[pk(n−k+r)+qk]an−k.(R)\Big[(n+r)(n+r-1) + p_0(n+r) + q_0\Big] a_n = -\sum_{k=1}^{n}\big[p_k(n-k+r) + q_k\big] a_{n-k}. \tag{R}

Define F(x):=x(x−1)+p0x+q0F(x) := x(x-1) + p_0 x + q_0, so the bracket on the left is F(n+r)F(n+r). At n=0n=0, the right side of (R) is empty (sum from k=1k=1 to 00), giving

F(r) a0=0.F(r)\, a_0 = 0.

Since we require a0≠0a_0 \ne 0 (this is what "leading term" means for the series), we must have

F(r)=0,F(r) = 0,

which is exactly the indicial equation. This proves that any Frobenius solution's leading exponent must be a root r1r_1 or r2r_2 of FF.

Step 2: Constructing y1y_1 for r=r1r = r_1, and its recursion. Set r=r1r = r_1, a0=1a_0 = 1. For n≥1n \ge 1, equation (R) reads

F(n+r1) an=−∑k=1n[pk(n−k+r1)+qk]an−k.(R1)F(n + r_1)\, a_n = -\sum_{k=1}^n \big[p_k(n-k+r_1) + q_k\big] a_{n-k}. \tag{R$_1$}

Since r1r_1 is the larger root, n+r1>r1≥r2n + r_1 > r_1 \ge r_2 for every n≥1n \ge 1, wait — we must check F(n+r1)≠0F(n+r_1) \ne 0. Because F(x)=(x−r1)(x−r2)F(x) = (x - r_1)(x-r_2), we have F(n+r1)=n(n+r1−r2)F(n+r_1) = n(n + r_1 - r_2). Since Re⁡(r1−r2)≥0\operatorname{Re}(r_1 - r_2) \ge 0 and n≥1n \ge 1, this vanishes only if r1−r2=0r_1 - r_2 = 0 is impossible for n≥1n\ge1 real positive factor n≠0n\ne0... more carefully: F(n+r1)=0F(n+r_1)=0 would require n=0n=0 (excluded) or n=r2−r1n = r_2 - r_1, impossible since Re⁡(r2−r1)≤0<n\operatorname{Re}(r_2-r_1)\le 0 < n. Hence F(n+r1)≠0F(n+r_1) \ne 0 for all n≥1n \ge 1, and (R1_1) determines ana_n uniquely in terms of a0,…,an−1a_0,\dots,a_{n-1}, by induction on nn. This proves existence and uniqueness of the formal series coefficients, establishing part (1) at the formal level.

Step 3: Convergence of the series for y1y_1 (majorant argument). Since p0,q0p_0, q_0 are analytic on ∣x∣<ρ|x|<\rho, for any 0<ρ′<ρ0 < \rho' < \rho there is M>0M>0 with ∣pk∣≤M/ρ′k|p_k| \le M/\rho'^k and ∣qk∣≤M/ρ′k|q_k|\le M/\rho'^k for all k≥0k \ge 0 (Cauchy estimates for analytic functions). From (R1_1),

∣an∣≤1∣F(n+r1)∣∑k=1n(∣pk∣(n−k+∣r1∣)+∣qk∣)∣an−k∣.|a_n| \le \frac{1}{|F(n+r_1)|}\sum_{k=1}^n \big(|p_k|(n - k + |r_1|) + |q_k|\big)|a_{n-k}|.

For nn large, ∣F(n+r1)∣=n∣n+r1−r2∣≥n2/2|F(n+r_1)| = n|n + r_1 - r_2| \ge n^2/2 eventually (since r1−r2r_1 - r_2 fixed), and the right side is bounded by CMn2∑k=1nρ′−k(n+∣r1∣)∣an−k∣\frac{CM}{n^2}\sum_{k=1}^n \rho'^{-k}(n+|r_1|)|a_{n-k}| for a suitable constant CC. One shows by comparison to the equation for

g(x)=∑bnxn,b0=1,bn=An2∑k=1nρ′−kbn−kg(x) = \sum b_n x^n, \qquad b_0=1,\quad b_n = \frac{A}{n^2}\sum_{k=1}^n \rho'^{-k} b_{n-k}

(for suitable constant AA absorbing the bounded prefactors) that ∣an∣≤bn|a_n| \le b_n for all nn, by induction, and that bnb_n are exactly the Taylor coefficients of a function analytic on ∣x∣<ρ′|x| < \rho' — indeed gg satisfies the algebraic relation g(x)(1−Ax/(1−x/ρ′))≈1g(x)(1 - A x/(1-x/\rho')) \approx 1 near 00, giving a function analytic at x=0x=0, hence radius of convergence ≥\ge some positive number, and letting ρ′→ρ\rho' \to \rho shows the radius of convergence of ∑anxn\sum a_n x^n is at least ρ\rho. (This is the standard majorant technique used identically in the ordinary-point existence proof; we invoke Series Solutions Near Ordinary Points only for the general fact that a recursion of this dominated form produces a convergent series — the singular structure has already been isolated in F(n+r1)F(n+r_1).) Hence y1(x)=xr1∑anxny_1(x) = x^{r_1}\sum a_n x^n converges for 0<x<ρ0 < x < \rho and solves (∗)(\ast) there by construction (all coefficient equations (R1_1) hold). This completes part (1).

Step 4: Case r1−r2r_1 - r_2 not a nonnegative integer — proving part (2). Repeat Steps 2–3 with r=r2r = r_2. The only place the earlier argument used r1r_1 being the larger root was to guarantee F(n+r2)≠0F(n+r_2)\ne 0 for all n≥1n\ge1. Now

F(n+r2)=(n+r2−r1)(n+r2−r2)=n(n+r2−r1)=n(n−(r1−r2)).F(n+r_2) = (n + r_2 - r_1)(n+r_2-r_2)=n(n+r_2-r_1) = n(n - (r_1-r_2)).

This vanishes iff n=r1−r2n = r_1 - r_2. By hypothesis r1−r2r_1 - r_2 is not a nonnegative integer, so in particular it is never equal to any positive integer nn; hence F(n+r2)≠0F(n+r_2)\ne0 for all n≥1n \ge 1, and the same recursion and majorant argument as Steps 2–3 (with r2r_2 in place of r1r_1) produce a second convergent series solution y2=xr2∑bnxny_2 = x^{r_2}\sum b_n x^n, b0=1b_0=1. Since r1≠r2r_1 \ne r_2 and the leading exponents differ (and neither series contains a log⁡x\log x term), y1,y2y_1, y_2 are linearly independent: their Wronskian near x=0x=0 behaves like

W(y1,y2)(x)∼(r2−r1) xr1+r2−1≠0W(y_1,y_2)(x) \sim (r_2 - r_1)\, x^{r_1+r_2 -1} \ne 0

for small x>0x > 0 (leading-order computation of y1y2′−y1′y2y_1 y_2' - y_1'y_2 using only the leading terms xr1,xr2x^{r_1}, x^{r_2}), so they form a fundamental set. This proves part (2).

Step 5: Case r1=r2r_1 = r_2 — proving part (3), via reduction of order. Here F(x)=(x−r)2F(x) = (x-r)^2. We already have y1(x)=xr∑anxny_1(x) = x^r\sum a_n x^n from Step 2 (the argument there only required F(n+r1)≠0F(n+r_1)\ne 0 for n≥1n\ge1, which holds since F(n+r)=n2≠0F(n+r)=n^2\ne0). For a second solution, apply reduction of order: set y2=y1(x)v(x)y_2 = y_1(x) v(x). Substituting into y′′+p(x)y′+q(x)y=0y'' + p(x) y' + q(x) y=0 (recall p=p0(x)/xp = p_0(x)/x, q=q0(x)/x2q=q_0(x)/x^2) and using that y1y_1 solves the equation gives the standard reduction-of-order identity

v′(x)=1y1(x)2exp⁡(−∫p(x) dx).v'(x) = \frac{1}{y_1(x)^2}\exp\Big(-\int p(x)\,dx\Big).

Since p(x)=p0(x)/x=p0/x+p1+p2x+⋯p(x) = p_0(x)/x = p_0/x + p_1 + p_2 x + \cdots, we get ∫p dx=p0ln⁡x+(analytic)\int p\,dx = p_0 \ln x + (\text{analytic}), so

exp⁡(−∫p dx)=x−p0 h(x),h analytic, h(0)≠0.\exp\Big(-\int p\,dx\Big) = x^{-p_0}\, h(x), \qquad h \text{ analytic}, \ h(0)\ne 0.

Also y1(x)2=x2rk(x)y_1(x)^2 = x^{2r} k(x) with kk analytic, k(0)=a02=1≠0k(0) = a_0^2 = 1 \ne 0. Since rr is a double root of F(x)=x(x−1)+p0x+q0F(x) = x(x-1)+p_0x+q_0, comparing constant terms of FF: expanding (x−r)2=x2−2rx+r2(x-r)^2 = x^2 - 2rx + r^2 against x(x−1)+p0x+q0=x2+(p0−1)x+q0x(x-1)+p_0 x+q_0 = x^2+(p_0-1)x+q_0 gives p0−1=−2rp_0 - 1 = -2r, i.e. p0=1−2rp_0 = 1-2r. Hence

v′(x)=x−p0h(x)x2rk(x)=x−(1−2r)−2r⋅h(x)k(x)=x−1⋅h(x)k(x).v'(x) = \frac{x^{-p_0} h(x)}{x^{2r}k(x)} = x^{-(1-2r)-2r}\cdot\frac{h(x)}{k(x)} = x^{-1}\cdot \frac{h(x)}{k(x)}.

Since h/kh/k is analytic and nonzero at 00, write h/k=∑n≥0dnxnh/k = \sum_{n\ge0} d_n x^n with d0≠0d_0 \ne 0; then

v′(x)=d0x−1+d1+d2x+⋯  ⟹  v(x)=d0ln⁡x+∑n≥1dnnxn.v'(x) = d_0 x^{-1} + d_1 + d_2 x + \cdots \implies v(x) = d_0 \ln x + \sum_{n\ge1}\frac{d_n}{n} x^n.

Therefore

y2=y1v=d0 y1(x)ln⁡x+xr∑n≥1dnna0xn⋅(⋯ ),y_2 = y_1 v = d_0\, y_1(x)\ln x + x^r\sum_{n\ge1}\frac{d_n}{n} a_0 x^n \cdot(\cdots),

which, after absorbing constants, has exactly the claimed form y1ln⁡x+xr∑n≥1cnxny_1 \ln x + x^r \sum_{n\ge1} c_n x^n (the coefficient d0≠0d_0 \ne 0 can be normalized to 11 by rescaling y2y_2). The convergence of ∑cnxn\sum c_n x^n on 0<x<ρ0<x<\rho follows because h/kh/k, as a ratio/product of functions analytic on ∣x∣<ρ|x|<\rho with k(0)≠0k(0)\ne0, is itself analytic on ∣x∣<ρ|x| < \rho (shrinking ρ\rho if kk has zeros, but k(0)=1k(0) = 1 and kk continuous keeps it nonzero near 00, and one takes ρ\rho as the radius up to the nearest zero of kk or singularity of hh). Linear independence of y1,y2y_1, y_2 is immediate since y2/y1y_2/y_1 contains ln⁡x\ln x, which is non-constant, so y2y_2 is not a scalar multiple of y1y_1. This proves part (3).

Step 6: Case r1−r2=N∈Z>0r_1 - r_2 = N \in \mathbb{Z}_{>0} — proving part (4). Again y1=xr1∑anxny_1 = x^{r_1}\sum a_n x^n exists by Step 2. Attempt the recursion (R) with r=r2r = r_2: by the computation in Step 4, F(n+r2)=n(n−N)F(n+r_2) = n(n-N), which vanishes at n=Nn = N. At n=Nn=N, equation (R) becomes

0⋅bN=−∑k=1N[pk(N−k+r2)+qk]bN−k=:S.0 \cdot b_N = -\sum_{k=1}^{N}\big[p_k(N-k+r_2)+q_k\big]b_{N-k} =: S.

Two sub-cases:

  • If S=0S = 0, then bNb_N is a free parameter, and one can consistently choose bN=0b_N = 0 and continue the recursion for n>Nn > N (each subsequent F(n+r2)=n(n−N)≠0F(n+r_2) = n(n-N)\ne 0 for n>Nn>N), producing a genuine second series solution y2=xr2∑bnxny_2 = x^{r_2}\sum b_n x^n with no logarithm — this is the case C=0C=0 in the theorem statement.
  • If S≠0S \ne 0, no choice of bNb_N satisfies the equation, so a pure Frobenius series with exponent r2r_2 does not exist, and a logarithmic term is unavoidable. One constructs y2y_2 via reduction of order exactly as in Step 5: set y2=y1(x)v(x)y_2 = y_1(x)v(x), and the same computation gives
v′(x)=x−1−N⋅(analytic, nonzero at 0),v'(x) = x^{-1-N}\cdot(\text{analytic, nonzero at }0),

(the exponent shift by NN arises because p0=1−r1−r2p_0 = 1 - r_1 - r_2 now, from F(x)=(x−r1)(x−r2)F(x)=(x-r_1)(x-r_2), and y12∼x2r1y_1^2 \sim x^{2r_1}, giving v′∼xr2−r1−1(⋯ )=x−N−1(⋯ )v' \sim x^{r_2-r_1-1}(\cdots) = x^{-N-1}(\cdots)). Expanding the analytic factor as ∑dnxn\sum d_n x^n and integrating term by term, the term dNx−1d_N x^{-1} integrates to dNln⁡xd_N \ln x, giving

v(x)=dNln⁡x+(Laurent terms x−N,…,x−1 absent or resolved)+analytic remainder,v(x) = d_N \ln x + \Big(\text{Laurent terms } x^{-N},\dots,x^{-1}\text{ absent or resolved}\Big) + \text{analytic remainder},

so that y2=y1vy_2 = y_1 v again takes the form Cy1ln⁡x+xr2∑n≥0cnxnC y_1 \ln x + x^{r_2}\sum_{n\ge0} c_n x^n, c0≠0c_0 \ne 0 after normalization, with C=dNC = d_N possibly zero (recovering the first sub-case) or nonzero. Convergence of all series follows by the same majorant/analyticity argument as before. This establishes part (4), and linear independence of y1,y2y_1,y_2 follows as in Step 5 when C≠0C\ne0 (via the ln⁡x\ln x term), or as in Step 4 (via distinct leading exponents r1≠r2r_1\ne r_2) when C=0C=0.

Conclusion. In every case, y1y_1 and y2y_2 are solutions of (∗)(\ast), convergent on 0<x<ρ0<x<\rho, and linearly independent, hence a fundamental set. ■\blacksquare

∎

Worked examples

Example 1 (distinct roots not differing by an integer — an asymmetric Bessel-type equation). Solve near x=0x=0:

4x2y′′+4x(1+12x)y′+(4x2−1)y=0.4x^2 y'' + 4x\Big(1+\tfrac12 x\Big) y' + (4x^2 - 1)y = 0.

Dividing by 4x24x^2: p(x)=1+x/2xp(x) = \dfrac{1+x/2}{x}, q(x)=4x2−14x2=1−14x2q(x) = \dfrac{4x^2-1}{4x^2} = 1 - \dfrac{1}{4x^2}. Check regular singularity: xp(x)=1+x/2xp(x) = 1 + x/2 (analytic, entire), x2q(x)=x2−1/4x^2q(x) = x^2 - 1/4 (analytic, entire). So x=0x=0 is a regular singular point with

p0=1, p1=12, pk=0 (k≥2);q0=−14, q2=1, qk=0 otherwise.p_0 = 1,\ p_1 = \tfrac12,\ p_k = 0\ (k\ge2); \qquad q_0 = -\tfrac14,\ q_2 = 1,\ q_k=0\text{ otherwise}.

Indicial equation: F(r)=r(r−1)+1⋅r−14=r2−14=0  ⟹  r=±12F(r) = r(r-1) + 1\cdot r - \tfrac14 = r^2 - \tfrac14 = 0 \implies r = \pm\tfrac12. So r1=1/2r_1 = 1/2, r2=−1/2r_2 = -1/2, and r1−r2=1r_1 - r_2 = 1 — wait, check: this differs by an integer (=1=1), so this example actually falls into case (4), not case (2). Let's instead use the genuinely asymmetric equation designed to land in case (2); adjust the middle coefficient's constant term. Take instead:

4x2y′′+4x(1+12x)y′+(4x2−94)y=0.4x^2 y'' + 4x\Big(1+\tfrac12 x\Big) y' + (4x^2 - \tfrac{9}{4})y = 0.

Now q(x)=1−916x2q(x) = 1 - \dfrac{9}{16x^2}, so x2q(x)=x2−9/16x^2 q(x) = x^2 - 9/16, giving q0=−9/16q_0 = -9/16.

Indicial equation: F(r)=r(r−1)+r−916=r2−916=0  ⟹  r=±34F(r) = r(r-1) + r - \tfrac{9}{16} = r^2 - \tfrac{9}{16}=0 \implies r = \pm \tfrac34. Now r1=3/4r_1 = 3/4, r2=−3/4r_2 = -3/4, and r1−r2=3/2r_1 - r_2 = 3/2, not an integer. Good — this is genuinely case (2). Write y=xr∑n≥0anxny = x^r\sum_{n\ge0} a_n x^n. Plug into the recursion (R) with p1=1/2p_1 = 1/2 (the only nonzero pkp_k, k≥1k\ge1), q2=1q_2 = 1 (the only nonzero qkq_k, k≥1k\ge1), all other pk,qk=0p_k, q_k = 0:

F(n+r) an=−∑k=1n[pk(n−k+r)+qk]an−k.F(n+r)\,a_n = -\sum_{k=1}^n\big[p_k(n-k+r) + q_k\big]a_{n-k}.

Only k=1k=1 (from p1p_1) and k=2k=2 (from q2q_2) contribute:

F(n+r) an=−12(n−1+r) an−1−an−2(a−1:=0).F(n+r)\,a_n = -\tfrac12(n-1+r)\,a_{n-1} - a_{n-2} \quad (a_{-1}:=0).

Root r1=3/4r_1 = 3/4: F(n+3/4)=(n+3/4)2−9/16=n2+32n=n(n+32)F(n + 3/4) = (n+3/4)^2 - 9/16 = n^2 + \tfrac32 n = n(n+\tfrac32).

  • n=1n=1: F(1+34)=1⋅52=52F(1+\tfrac34)=1\cdot\tfrac52=\tfrac52. RHS =−12(34)a0=−38a0=-\tfrac12(\tfrac34)a_0 = -\tfrac38 a_0. With a0=1a_0=1: a1=−3/85/2=−320a_1 = -\dfrac{3/8}{5/2} = -\dfrac{3}{20}.
  • n=2n=2: F(2+34)=2⋅72=7F(2+\tfrac34) = 2\cdot\tfrac72=7. RHS =−12(1+34)a1−a0=−78a1−1= -\tfrac12(1+\tfrac34)a_1 - a_0 = -\tfrac78 a_1 - 1. Compute −78⋅(−320)=21160-\tfrac78\cdot(-\tfrac3{20}) = \tfrac{21}{160}, so RHS =21160−1=−139160= \tfrac{21}{160}-1 = -\tfrac{139}{160}. Then a2=−139160⋅7=−1391120a_2 = -\dfrac{139}{160\cdot 7} = -\dfrac{139}{1120}.

So

y1(x)=x3/4(1−320x−1391120x2−⋯).y_1(x) = x^{3/4}\Big(1 - \tfrac{3}{20}x - \tfrac{139}{1120}x^2 - \cdots\Big).

Root r2=−3/4r_2 = -3/4: F(n−34)=(n−34)2−916=n2−32n=n(n−32)F(n-\tfrac34) = (n-\tfrac34)^2-\tfrac9{16} = n^2 - \tfrac32 n = n(n-\tfrac32).

  • n=1n=1: F(1−34)=1⋅(−12)=−12F(1-\tfrac34) = 1\cdot(-\tfrac12) = -\tfrac12. RHS =−12(1−34−1)b0= -\tfrac12(1-\tfrac34-1)b_0 — careful: RHS uses (n−1+r)=(0−34)=−34(n-1+r) = (0 - \tfrac34) = -\tfrac34 at n=1n=1: −12(−34)b0=38b0=38-\tfrac12(-\tfrac34)b_0 = \tfrac38 b_0 = \tfrac38. So b1=3/8−1/2=−34b_1 = \dfrac{3/8}{-1/2} = -\dfrac34.
  • n=2n=2: F(2−34)=2⋅12=1F(2-\tfrac34) = 2\cdot\tfrac12 = 1. RHS =−12(1−34)b1−b0=−18⋅(−34)−1=332−1=−2932=-\tfrac12(1-\tfrac34)b_1 - b_0 = -\tfrac18\cdot(-\tfrac34) - 1 = \tfrac3{32}-1 = -\tfrac{29}{32}. So b2=−2932b_2 = -\tfrac{29}{32}.

So

y2(x)=x−3/4(1−34x−2932x2−⋯).y_2(x) = x^{-3/4}\Big(1 - \tfrac34 x - \tfrac{29}{32}x^2 - \cdots\Big).

The general solution near x=0+x=0^+ is y=C1y1(x)+C2y2(x)y = C_1 y_1(x) + C_2 y_2(x), both series convergent for all xx (since p0,q0p_0,q_0 here are polynomials, entire functions, ρ=∞\rho=\infty).

Example 2 (roots differing by a positive integer, forcing a logarithm — a genuinely asymmetric case). Solve near x=0x=0:

x2y′′+x(1−x)y′−(x+1)y=0.x^2 y'' + x(1-x)y' - (x + 1)y = 0.

Here p(x)=1−xxp(x) = \dfrac{1-x}{x}, q(x)=−x+1x2q(x) = -\dfrac{x+1}{x^2}. Check: xp(x)=1−xxp(x) = 1-x (entire), x2q(x)=−x−1x^2q(x) = -x-1 (entire): regular singular point, with p0=1,p1=−1p_0=1, p_1=-1; q0=−1,q1=−1q_0=-1, q_1=-1 (all other pk,qk=0p_k,q_k=0).

Indicial equation: F(r)=r(r−1)+r−1=r2−1=0  ⟹  r=±1F(r) = r(r-1)+r-1 = r^2-1=0 \implies r=\pm1. So r1=1r_1=1, r2=−1r_2=-1, N=r1−r2=2N=r_1-r_2=2: case (4).

Recursion. From (R): F(n+r)an=−∑k=1n[pk(n−k+r)+qk]an−kF(n+r)a_n = -\sum_{k=1}^n[p_k(n-k+r)+q_k]a_{n-k}. Nonzero pk,qkp_k,q_k only at k=1k=1: p1=−1,q1=−1p_1=-1,q_1=-1, so

F(n+r)an=−[−(n−1+r)−1]an−1=(n+r) an−1.F(n+r)a_n = -\big[-(n-1+r) -1\big]a_{n-1} = (n+r)\,a_{n-1}.

Root r1=1r_1 = 1: F(n+1)=(n+1)2−1=n2+2n=n(n+2)F(n+1) = (n+1)^2-1 = n^2+2n = n(n+2).

  • n=1n=1: F(2)=3F(2) = 3. RHS =(1+1)a0=2=(1+1)a_0=2. So a1=2/3a_1 = 2/3.
  • n=2n=2: F(3)=8F(3)=8. RHS=(2+1)a1=3⋅23=2=(2+1)a_1 = 3\cdot\tfrac23=2. So a2=2/8=1/4a_2 = 2/8=1/4.
  • n=3n=3: F(4)=15F(4)=15. RHS=(3+1)a2=4⋅14=1=(3+1)a_2=4\cdot\tfrac14=1. So a3=1/15a_3=1/15.

So y1(x)=x(1+23x+14x2+115x3+⋯)y_1(x) = x\Big(1+\tfrac23x+\tfrac14x^2+\tfrac1{15}x^3+\cdots\Big), an entire function (radius ∞\infty, since p0,q0p_0,q_0 are polynomials). Attempted root r2=−1r_2=-1: F(n−1)=(n−1)2−1=n2−2n=n(n−2)F(n-1)=(n-1)^2-1 = n^2-2n=n(n-2), vanishing at n=2=Nn=2=N, consistent with the theorem. Recursion for bnb_n: F(n−1)bn=(n−1)bn−1F(n-1)b_n = (n-1)b_{n-1}.

  • n=1n=1: F(0)=−1F(0)=-1. RHS =0⋅b0=0=0\cdot b_0=0. So b1=0b_1=0.
  • n=2n=2: F(1)=0F(1)=0. RHS=1⋅b1=0=1\cdot b_1=0. This is 0⋅b2=00\cdot b_2 = 0: consistent, and b2b_2 is free — but we must check whether a log term is actually needed by continuing carefully, since consistency at n=Nn=N only rules out an obstruction, not the possibility that the general theory still requires re-deriving y2y_2 from reduction of order to see if C=0C=0 genuinely.

Since the compatibility condition holds (S=0S=0 at n=2n=2), Frobenius' theorem guarantees a solution with no logarithm, C=0C=0. Set b2=0b_2 = 0 (a free choice; any other choice would just add a multiple of y1y_1, since with b2b_2 free the recursion for n>2n>2 generates an extra multiple of the y1y_1-tail). Continue:

  • n=3n=3: F(2)=3F(2)=3. RHS=(2)b2=0=(2)b_2=0. So b3=0b_3=0; indeed all subsequent bn=0b_n=0 by induction, since each recursion is homogeneous once b1=b2=0b_1=b_2=0 (there is no external "source" term because q1q_1's contribution already folded into the (n+r)an−1(n+r)a_{n-1} pattern, and here it produces zero once bn−1=0b_{n-1}=0).

So y2(x)=x−1(1+0⋅x+0⋅x2+⋯)=x−1=1xy_2(x) = x^{-1}\big(1 + 0\cdot x + 0\cdot x^2+\cdots\big) = x^{-1} = \dfrac1x.

Check directly: does y=1/xy=1/x solve the ODE? y′=−x−2y'=-x^{-2}, y′′=2x−3y''=2x^{-3}. Substitute:

x2(2x−3)+x(1−x)(−x−2)−(x+1)(x−1)=2x−1−(1−x)x−1−(1+x−1)=2x−1−x−1+1−1−x−1=0.x^2(2x^{-3}) + x(1-x)(-x^{-2}) - (x+1)(x^{-1}) = 2x^{-1} - (1-x)x^{-1} - (1+x^{-1}) = 2x^{-1} - x^{-1}+1 - 1 - x^{-1} = 0.

It checks out exactly. So here the "positive integer difference" case resolves without a logarithm (C=0C=0), and the general solution is

y(x)=C1 x(1+23x+14x2+115x3+⋯)+C2x.y(x) = C_1\, x\Big(1+\tfrac23 x+\tfrac14x^2+\tfrac1{15}x^3+\cdots\Big) + \frac{C_2}{x}.
CounterexampleDropping regularity: an irregular singular point where the method fails

Consider

x3y′′+y=0x^3 y'' + y = 0

near x=0x=0. Here p(x)=0p(x)=0 and q(x)=1/x3q(x) = 1/x^3. Check the regular singular point condition: x⋅p(x)=0x\cdot p(x)=0 is analytic, but

x2q(x)=x2⋅1x3=1x,x^2 q(x) = x^2\cdot\frac{1}{x^3} = \frac1x,

which is not analytic at x=0x=0 (it blows up). So x=0x=0 is an irregular singular point — the hypothesis of Frobenius' theorem fails. Try the Frobenius ansatz anyway: y=xr∑n≥0anxny = x^r\sum_{n\ge0}a_nx^n, a0≠0a_0\ne0. Then

x3y′′=∑n(n+r)(n+r−1)anxn+r+1,y=∑nanxn+r.x^3y'' = \sum_n (n+r)(n+r-1)a_n x^{n+r+1}, \qquad y = \sum_n a_n x^{n+r}.

Matching the lowest power of xx: the yy-term contributes xrx^r at order n=0n=0, but the x3y′′x^3y'' term contributes its lowest power at xr+1x^{r+1} (from n=0n=0), one power higher. These cannot cancel against each other at the same order unless we also bring in n=1n=1 from yy, giving an equation that reads, order by order,

xr:a0=0,x^r:\quad a_0 = 0,

forcing a0=0a_0=0 — contradicting the requirement a0≠0a_0\ne0. So no Frobenius series solution exists at all: the method breaks down completely, exactly because the regular singular point hypothesis fails. (In fact, solutions of x3y′′+y=0x^3y''+y=0 near 00 exhibit essential-singularity type behavior, e.g. involving exp⁡(±2i/x)\exp(\pm 2i/\sqrt{x})-type asymptotics, not expressible as xr×x^r\times convergent power series — a genuinely different, harder regime beyond Frobenius' theorem.)

Scope, edge cases & common traps

SituationWhat to do
x0≠0x_0 \ne 0Translate: set t=x−x0t = x - x_0, redo everything in tt.
Roots complex conjugateTheory still applies (case "not differing by an integer" holds since difference is purely imaginary, never a positive integer unless zero); solutions are complex, take real/imaginary parts.
r1−r2=r_1 - r_2 = positive integerMust check the compatibility sum SS at n=Nn=N: log term appears iff S≠0S \ne 0. Never assume the log is automatic.
Only need one solutionThe larger root r1r_1 always yields a genuine series solution — start there.
Radius of convergenceAt least the distance from x0x_0 to the nearest other singular point of p,qp,q (same rule as ordinary points).

The mathematician behind it

Origins · 1 minFerdinand Georg Frobenius and how this was discovered
FG
Ferdinand Georg Frobenius
1849–1917 · German mathematician, University of Berlin

Frobenius is best known for foundational work in group representation theory (Frobenius groups, Frobenius reciprocity) and matrix theory (the Frobenius normal form, the Perron–Frobenius theorem's namesake half), but his 1873 paper on differential equations with singularities laid out the systematic method now bearing his name for constructing series solutions near what we now call regular singular points.

History
- 1866–1870: Fuchs studies linear ODEs with singular points and identifies the condition (now called the Fuchsian condition) distinguishing "mild" singularities.
- 1873: Frobenius publishes his method for constructing series solutions at these regular singular points, including the indicial equation and the logarithmic case.
- 1878–1880s: The theory is systematically applied to Legendre's, Bessel's, and the hypergeometric equations, unifying much of 19th-century special function theory.
- Early 1900s: The method becomes a standard tool in mathematical physics for boundary value problems in cylindrical and spherical coordinates.
- 20th century: Frobenius' construction is recognized as a special case of the general theory of regular singular points for linear systems, connecting to Fuchsian differential equations and monodromy theory in complex analysis.
Why do we need the specific "regular" condition (x−x0)p(x)(x-x_0)p(x), (x−x0)2q(x)(x-x_0)^2 q(x) analytic, rather than just requiring p,qp,q to have some pole?

The precise powers matter because they are exactly what makes the indicial equation a genuine, finite-degree polynomial (here quadratic) with well-defined roots, and what makes the leading-order balance in the substituted series close up into a solvable recursion. If pp had a pole of order 2 or qq a pole of order 3 or worse (an irregular singular point), the naive substitution either forces a0=0a_0=0, as in the counterexample, or produces an indicial "equation" of infinite degree, and no clean power-times-series solution exists in general. The specific orders (one power of (x−x0)(x-x_0) for pp, two for qq) are precisely calibrated so that after multiplying by x2x^2, the resulting equation is a genuinely analytic (ordinary-point-type) equation for the auxiliary quantity x−ryx^{-r}y, letting the ordinary-point convergence machinery apply.

Do the two indicial roots always give two independent solutions?

Not automatically — only when they don't differ by a nonnegative integer, or in the careful case-by-case analysis given by the theorem. If r1=r2r_1=r_2, the second solution necessarily contains a logarithm (Step 5 above proves this rigorously via reduction of order and cannot be avoided). If r1−r2r_1 - r_2 is a positive integer, the logarithm may or may not appear, depending on whether an explicit compatibility sum vanishes; this must be checked case by case, as in Worked Example 2, rather than assumed either way.

What if the indicial roots are complex?

The theory still applies essentially unchanged: complex conjugate roots r=α±iβr = \alpha \pm i\beta automatically satisfy "not differing by a nonnegative integer" (since their difference 2iβ2i\beta is not a positive integer unless β=0\beta=0), so case (2) of the theorem always applies, producing two independent (complex) Frobenius series. Real solutions are then obtained by taking the real and imaginary parts of either series, exactly as with complex roots of a constant-coefficient characteristic equation.

How is this different from series solutions at an ordinary point?

At an ordinary point, both pp and qq are analytic, and a pure Taylor series ∑an(x−x0)n\sum a_n(x-x_0)^n with integer, nonnegative exponents always works, with two independent solutions coming from the two free initial conditions a0,a1a_0, a_1 (this is the content of Series Solutions Near Ordinary Points). At a regular singular point, we must first solve an auxiliary indicial equation to discover the correct (possibly fractional, negative, or repeated) leading exponent, and the two solutions do not simply correspond to two free constants in a single ansatz — they may require entirely different exponents or a logarithmic correction.

Can the indicial equation have more than two roots?

No — for a second-order linear ODE, the indicial equation is always a quadratic (degree 2) polynomial in rr, since it arises from the leading-order terms r(r−1)+p0r+q0r(r-1) + p_0 r + q_0, which is quadratic by construction. For an nn-th order linear ODE with a regular singular point, the analogous indicial equation is degree nn, with the same trichotomy of cases (roots differing by integers vs. not) generalized accordingly, though the bookkeeping becomes more intricate.

Is the radius of convergence of the Frobenius series the same rule as for ordinary points?

Yes: the Frobenius series (the power series part, after removing xrx^r or xrln⁡xx^r\ln x) converges at least out to the distance from x0x_0 to the nearest other singular point of the equation in the complex plane, exactly as for ordinary-point Taylor series. This follows from the same majorant/Cauchy estimate argument, since the bound on the Taylor coefficients of p0,q0p_0, q_0 used in the proof is valid exactly on the disk of analyticity of p0,q0p_0,q_0, i.e., up to the nearest singularity.

Why must one always start with the larger root r1r_1?

Because the proof shows F(n+r1)≠0F(n+r_1)\ne 0 for every positive integer nn whenever r1r_1 is the root with the larger real part, guaranteeing the recursion never breaks down. Starting with the smaller root r2r_2 risks hitting F(n+r2)=0F(n+r_2)=0 at n=r1−r2n = r_1-r_2 when this difference is a nonnegative integer, which either blocks the construction outright (needs a log term) or, in the degenerate case r1=r2r_1=r_2, means there's only one exponent to try in the first place.

Before this