The Method of Frobenius

Most of the special functions of physics — Bessel, Legendre, Laguerre, Hermite, the whole hypergeometric family — are born at a single kind of trouble spot in a differential equation: the regular singular point. There the coefficients of $y''+p(x)y'+q(x)y=0$ blow up, but only gently, so that $(x-x_0)p(x)$ and $(x-x_0)^2q(x)$ stay analytic. The ordinary Taylor ansatz $y=\sum a_n(x-x_0)^n$ fails at such a point, yet Fuchs's theorem promises that at least one solution survives if we allow a leading fractional power. The Method of Frobenius is the systematic machine for building it. The one modification is decisive: multiply the power series by $(x-x_0)^s$ with the exponent $s$ left free. Substitute, demand that the most singular term cancel, and $s$ is forced to satisfy a simple quadratic — the indicial equation $s(s-1)+p_0s+q_0=0$ . Its two roots $s_1,s_2$ are the exponents of the singularity: they say whether a solution vanishes, diverges, or branches at $x_0$ , and physics reads its boundary conditions straight off them. Everything downstream — the recurrence that grinds out the coefficients, the radius of convergence, whether the second solution hides a logarithm — is governed by those two numbers and, above all, by their difference $s_1-s_2$ . This article develops the method as a working tool. We derive the indicial equation and recurrence from scratch, build the Bessel function $J_\nu$ as the flagship example, show why equal roots manufacture a logarithm through a $\partial/\partial s$ trick, and work two fully numerical cases — one with non-integer roots, one where an integer root gap still gives two clean series. By the end the method should feel less like a recipe and more like what it is: the gateway from elementary ODE theory to Sturm-Liouville theory and the orthogonal-polynomial machinery of mathematical physics.

Key takeaways

At a regular singular point of $y''+p\,y'+q\,y=0$ , seek $y=(x-x_0)^s\sum_{n\ge0}a_n(x-x_0)^n$ with $a_0\ne0$ ; the fractional power $(x-x_0)^s$ is the whole idea.
The exponent $s$ is fixed by the indicial equation $s(s-1)+p_0s+q_0=0$ , whose roots $s_1,s_2$ are the exponents of the singularity and set the local shape of every solution.
A recurrence relation then fixes all $a_n$ from $a_0$ , and Fuchs's theorem guarantees convergence out to the nearest other singular point in the complex plane.
The difference $s_1-s_2$ decides the structure: non-integer gives two pure series; equal roots always force a logarithm in the second solution; a positive-integer gap may or may not.
Frobenius's method constructs $J_\nu$ , $Y_\nu$ , $P_\ell$ , $Q_\ell$ and the $_2F_1$ hypergeometric function — the engine behind separation of variables in cylindrical and spherical geometry.
Key idea: at a regular singular point postulate $y=(x-x_0)^s\sum a_n(x-x_0)^n$ with $a_0\ne0$ ; the indicial equation fixes $s$ , the recurrence fixes the $a_n$ , and the difference of indicial roots tells you whether a logarithm is forced into the second solution.

The core insight

At a regular singular point, postulate $y=(x-x_0)^s\sum a_n(x-x_0)^n$ with $a_0\ne0$ ; the indicial equation fixes the exponent $s$ , the recurrence fixes the coefficients $a_n$ , and the difference of the two indicial roots tells you whether a logarithm is forced into the second solution.

Where this lives today

Frobenius's method is not a museum piece; it is the working engine behind several very modern computations. In gravitational physics, the quasinormal ringing of a perturbed black hole is governed by the Regge-Wheeler and Teukolsky equations, which have a regular singular point at the event horizon; Leaver's celebrated 1985 continued-fraction method expands the solution as a Frobenius series about that horizon and reads off the complex ringdown frequencies now being measured in LIGO-Virgo-KAGRA gravitational-wave signals. In quantum field theory, Feynman integrals satisfy Fuchsian (Picard-Fuchs) differential equations whose singularities sit at physical thresholds, and modern multi-loop amplitude calculations are organized around the Frobenius exponents at each threshold. In computer algebra, every serious system — Maple's `DEtools`, Mathematica's `DSolve`, the D-finite and holonomic-function libraries — implements Frobenius as its default constructor for solutions near a regular singular point, and the theory of connection problems (relating Frobenius bases at different singularities, the numerical analogue of analytic continuation) is an active area linking special-function theory, the Riemann-Hilbert problem, and high-precision evaluation of the functions of mathematical physics.

The physical picture

Think of a regular singular point as a place where the equation's coefficients diverge just gently enough that a solution still has a definite local shape: a leading power $(x-x_0)^s$ multiplying something smooth. The indicial equation is nothing more than the demand that this leading power balance the most singular terms of the ODE — dominant balance made exact. Because the equation is second order there is a two-dimensional space of solutions, and so there are two exponents, two possible local shapes. The first figure shows those two shapes for Bessel's equation of order one, where the indicial roots are $s=\pm1$ : one solution rises like $x$ and vanishes at the origin, the other falls like $1/x$ and diverges — and quantum mechanics keeps whichever is normalizable.

Figure

Two indicial exponents, two local shapes. For Bessel order ν = 1 the indicial equation s² − ν² = 0 gives roots s = +1 and s = −1. The regular branch behaves as y ∼ x and vanishes at the singular point x₀ = 0, while the singular branch behaves as y ∼ 1/x and diverges. Physical boundary conditions pick one exponent.

The second question is one of reach: how far from $x_0$ can we trust the series? Fuchs's theorem answers it geometrically. The Frobenius series converges inside the largest disc centred at $x_0$ that contains no other singular point of the equation — the radius equals the distance, in the complex plane, to the nearest other singularity. Legendre's equation is the clean illustration: it is singular at $x=\pm1$ , so a Frobenius (or ordinary) expansion about $x_0=0$ converges for $|x|<1$ and no further, which is exactly why the Legendre series must be truncated to a polynomial to survive at the endpoints $x=\pm1$ . The second figure shows that disc.

Figure

Fuchs's convergence disc in the complex x-plane. A Frobenius series about x₀ converges out to the nearest other singular point. Legendre's equation is singular at x = −1 and x = +1, so an expansion about x₀ = 0 converges inside the disc of radius R = 1 and diverges beyond it.

The logarithm that haunts the degenerate cases is not a mysterious accident either. It is the $\partial/\partial s$ derivative of the Frobenius series evaluated at a repeated root — the very same trick that turns $x^s$ into $\ln x$ as $s\to0$ , since $\partial_s x^s=x^s\ln x$ . Differentiating a power with respect to its exponent manufactures a logarithm, and that is all the mystery there is.

The core equations

We work throughout with the standard form of a linear second-order homogeneous ODE. The coefficient functions $p$ and $q$ may misbehave at the expansion point $x_0$ ; the entire classification of that point rests on how badly they misbehave.

y''+p(x)\,y'+q(x)\,y=0 \qquad (1)

where: $y(x)$ — the unknown function; $p(x),\,q(x)$ — coefficient functions, possibly singular at $x_0$ ; $x_0$ — the expansion point under study. If $p$ and $q$ are both analytic at $x_0$ the point is ordinary and ordinary power series suffice; Frobenius is needed only when they are not. The dividing line between the tractable and the hopeless is set by two limits. A point is a regular singular point when the coefficients diverge no worse than $p\sim(x-x_0)^{-1}$ and $q\sim(x-x_0)^{-2}$ , so that the products below stay finite and analytic.

\lim_{x\to x_0}(x-x_0)\,p(x)=p_0,\qquad \lim_{x\to x_0}(x-x_0)^2 q(x)=q_0 \qquad (2)

where: $p_0$ — leading coefficient of the $1/(x-x_0)$ part of $p$ ; $q_0$ — leading coefficient of the $1/(x-x_0)^2$ part of $q$ . Reading: these two numbers are the only pieces of $p$ and $q$ that reach the most singular order, so they alone determine the exponents. The one structural change that makes the method work is to attach a free power $(x-x_0)^s$ to the series. The requirement $a_0\ne0$ is what makes $s$ genuinely the leading exponent rather than a mislabelled ordinary series.

y(x)=(x-x_0)^{s}\sum_{n=0}^{\infty}a_n\,(x-x_0)^{n},\qquad a_0\ne0 \qquad (3)

where: $s$ — the indicial exponent, generally non-integer; $a_n$ — series coefficients with $a_0\ne0$ by convention. Reading: the ansatz is an ordinary Taylor series bodily lifted by a fractional power that captures the singular behaviour. Substituting (3) into (1) and collecting the lowest power $(x-x_0)^s$ gives a condition on $s$ alone — the indicial equation. It is a quadratic, so it has exactly two roots, matching the two-dimensional solution space.

s(s-1)+p_0\,s+q_0=0 \qquad (4)

where: $s$ — its roots are the exponents $s_1,s_2$ ; $p_0,q_0$ — the limiting coefficients from (2). Reading: this is the Euler-equation "characteristic polynomial" of the leading behaviour, and it involves only $p_0$ and $q_0$ . Solving the quadratic gives the two exponents explicitly. We always label them so that $\operatorname{Re}s_1\ge\operatorname{Re}s_2$ ; the larger root $s_1$ always yields a genuine Frobenius series.

s_{1,2}=\tfrac{1}{2}\Big[(1-p_0)\pm\sqrt{(1-p_0)^2-4q_0}\,\Big] \qquad (5)

where: $s_1$ — the larger root; $s_2$ — the smaller root. Reading: their difference $s_1-s_2=\sqrt{(1-p_0)^2-4q_0}$ is the single quantity that decides whether a logarithm appears. Once $s$ is fixed, the higher powers give a recurrence that fixes every coefficient in terms of earlier ones. Here $P$ is the indicial polynomial itself, and the crucial fact is that $P(s+n)$ can vanish — precisely when $s+n$ lands on the other indicial root.

a_n\,P(s+n)=-\sum_{k=0}^{n-1}a_k\big[(s+k)\,p_{n-k}+q_{n-k}\big],\qquad P(r)\equiv r(r-1)+p_0 r+q_0 \qquad (6)

where: $P(r)$ — the indicial polynomial; $p_m,q_m$ — Taylor coefficients of $(x-x_0)p$ and $(x-x_0)^2q$ . Reading: if $P(s+n)\ne0$ the coefficient $a_n$ is determined; if $P(s+n)=0$ the recurrence either is consistent (a free coefficient, no logarithm) or fails (a logarithm is forced). When the roots coincide or differ by an integer, the second independent solution generically carries an explicit logarithm built on top of the first.

y_2=k\,y_1\ln(x-x_0)+(x-x_0)^{s_2}\sum_{n=0}^{\infty}c_n(x-x_0)^{n} \qquad (7)

where: $y_1$ — the first Frobenius solution (larger root); $c_n$ — new coefficients fixed by substitution; $k$ — a constant, always nonzero for equal roots, possibly zero for an integer gap. Reading: the log term is the price the second solution pays when the two exponents can no longer keep two independent pure series apart. Whatever the case, two universal tools finish the job. Reduction of order manufactures the missing second solution from the known first one, reproducing any logarithm automatically, and Abel's identity certifies independence through a Wronskian that never vanishes.

y_2(x)=y_1(x)\int^{x}\frac{\exp\!\big[-\int p\,\mathrm{d}x'\big]}{y_1(x')^{2}}\,\mathrm{d}x',\qquad W(y_1,y_2)=C\exp\!\Big[-\int p\,\mathrm{d}x\Big] \qquad (8)

where: $y_1$ — the known solution; $y_2$ — the second independent solution; $W$ — the Wronskian; $C$ — a nonzero constant. Reading: a nonzero $W$ confirms $y_1,y_2$ are linearly independent everywhere the exponential factor is finite.

Derivation AWhere the indicial equation and recurrence come from

The goal is to show that the two central objects of the method — the quadratic indicial equation and the coefficient recurrence — both drop out of a single substitution, once the ODE is put into its Euler form. We take $x_0=0$ without loss of generality, expand the well-behaved products $xp(x)$ and $x^2q(x)$ as ordinary Taylor series, insert the Frobenius ansatz, and read off the coefficient of each power of $x$ . The lowest power delivers the indicial equation; every higher power delivers one step of the recurrence.

Step 1. Multiply the standard form (1) by $x^2$ to reach the Euler form $x^2y''+x\,(xp)\,y'+(x^2q)\,y=0$ , and expand the regular products as $xp(x)=p_0+p_1x+p_2x^2+\cdots$ and $x^2q(x)=q_0+q_1x+q_2x^2+\cdots$ . Why: at a regular singular point these two products are analytic by definition, so they have honest Taylor series — this is the entire payoff of the regularity condition (2).

Step 2. Insert $y=\sum_n a_n x^{s+n}$ , so that $y''=\sum_n a_n(s+n)(s+n-1)x^{s+n-2}$ and $x^2y''=\sum_n a_n(s+n)(s+n-1)x^{s+n}$ . Every term of the Euler-form equation now scales as $x^{s+n}$ , so like powers can be collected. Why: the factor $x^2$ was chosen precisely so that $y''$ , $xp\,y'$ and $x^2q\,y$ all land on the same ladder of powers $x^{s+n}$ , making the bookkeeping finite at each order.

Step 3. Collect the lowest power, $x^s$ (the $n=0$ term). Only the leading Taylor pieces $p_0,q_0$ contribute, giving $a_0\big[s(s-1)+p_0s+q_0\big]x^s=0$ . Since $a_0\ne0$ by the ansatz, the bracket must vanish — this is the indicial equation (4). Why: requiring $a_0\ne0$ is exactly what forces the bracket, not $a_0$ , to be zero; drop that requirement and $s$ would be undetermined.

Step 4. Collect a general power $x^{s+n}$ with $n\ge1$ . The diagonal term gives $a_n\,P(s+n)$ with $P(r)=r(r-1)+p_0r+q_0$ , and the off-diagonal Taylor coefficients feed in earlier $a_k$ , yielding the recurrence (6): $a_n\,P(s+n)=-\sum_{k=0}^{n-1}a_k[(s+k)p_{n-k}+q_{n-k}]$ . Why: because $P(s+n)$ is the same indicial polynomial evaluated at the shifted exponent, it vanishes exactly when $s+n$ equals the other root — the algebraic origin of every logarithmic complication downstream. *Verified: the $x^s$ coefficient reproduces (4); the $x^{s+n}$ coefficient reproduces (6); consistent with Arfken & Weber ch. 7 and Stone & Goldbart ch. 16.*

Go deeper · Derivation B · 1 minBessel's equation and the Bessel function of the first kind

Derivation BBessel's equation and the Bessel function of the first kind

The goal is to run the full machine on the single most important example in physics, Bessel's equation $x^2y''+xy'+(x^2-\nu^2)y=0$ , obtaining the indicial roots $\pm\nu$ , a clean two-term recurrence, and the closed-form series for $J_\nu$ . The origin is a regular singular point: here $xp(x)=1$ so $p_0=1$ , and $x^2q(x)=x^2-\nu^2$ so $q_0=-\nu^2$ (with $q_2=1$ the only other nonzero Taylor coefficient). Every classical cylinder function follows from what this derivation produces.

Step 1. The indicial equation is $s(s-1)+1\cdot s-\nu^2=s^2-\nu^2=0$ , so $s_1=\nu$ and $s_2=-\nu$ , differing by $2\nu$ . Why: the $-\nu^2$ from $q_0$ turns the generic quadratic into a difference of squares, which is why Bessel's exponents are simply $\pm\nu$ — the order of the equation read directly off the singular point.

Step 2. Substituting $y=\sum_n a_n x^{s+n}$ and using $x^2y''+xy'=\sum_n a_n(s+n)^2x^{s+n}$ gives $\sum_n a_n\big[(s+n)^2-\nu^2\big]x^{s+n}+\sum_n a_n x^{s+n+2}=0$ . The lone $x^2y$ term shifts the index by two. Why: because the perturbation $x^2q$ enters two orders above the leading balance, the recurrence connects $a_n$ to $a_{n-2}$ only — a two-term recurrence, not a three-term one.

Step 3. Matching powers gives $a_n\big[(s+n)^2-\nu^2\big]=-a_{n-2}$ , i.e. for the larger root $s=\nu$ , $a_n=-a_{n-2}/\big[(s+n)^2-\nu^2\big]=-a_{n-2}/\big[n(n+2\nu)\big]$ . At $n=1$ the bracket $(1+2\nu)\ne0$ forces $a_1=0$ , so every odd coefficient vanishes. Why: with only even indices surviving, the solution is even in $x$ up to the overall $x^\nu$ , exactly the parity of $J_\nu$ .

Step 4. Iterating over even indices $n=2m$ gives $a_{2m}=\dfrac{(-1)^m a_0}{2^{2m}\,m!\,(\nu+1)(\nu+2)\cdots(\nu+m)}$ . Why: each step contributes a factor $-1/[2m\cdot(2m+2\nu)]=-1/[4m(m+\nu)]$ , and the accumulated product is the rising factorial that the Gamma function will absorb.

Step 5. Choosing the standard normalization $a_0=1/[2^\nu\,\Gamma(\nu+1)]$ folds the product $(\nu+1)\cdots(\nu+m)\,\Gamma(\nu+1)=\Gamma(m+\nu+1)$ , giving the boxed result

J_\nu(x)=\sum_{m=0}^{\infty}\frac{(-1)^m}{m!\,\Gamma(m+\nu+1)}\Big(\frac{x}{2}\Big)^{2m+\nu}

with leading behaviour $J_\nu(x)\sim (x/2)^\nu/\Gamma(\nu+1)$ . Why: for non-integer $\nu$ the smaller root $s_2=-\nu$ generates an independent $J_{-\nu}$ , but for integer $\nu$ one finds $J_{-n}=(-1)^nJ_n$ — the two series collapse to one, and the genuinely independent second solution is the logarithmic Neumann function $Y_n$ . *Verified: coefficients and the leading term $(x/2)^\nu/\Gamma(\nu+1)$ match Arfken & Weber ch. 14; the collapse $J_{-n}=(-1)^nJ_n$ is standard.*

Go deeper · Derivation C · 1 minWhy equal indicial roots force a logarithm

Derivation CWhy equal indicial roots force a logarithm

The goal is to explain the most surprising feature of the method — that a double indicial root produces a second solution containing $\ln(x-x_0)$ — and to show it is inevitable, not a patch. The trick is to keep the exponent $s$ as a free continuous parameter, let the coefficients $a_n(s)$ solve the recurrence with $s$ undetermined, and differentiate the resulting family of near-solutions with respect to $s$ . Take $x_0=0$ , $a_0=1$ , and write $y(x,s)=x^s\sum_n a_n(s)x^n$ .

Step 1. Because the $a_n(s)$ are chosen to kill every higher power by construction, applying the differential operator $L$ leaves only the leading term: $L\,y(x,s)=a_0\,P(s)\,x^s=P(s)\,x^s$ , where $P$ is the indicial polynomial. Why: the recurrence is exactly the statement "all coefficients above the leading one cancel", so the entire residual of $Ly$ is the single indicial term $P(s)x^s$ .

Step 2. Equal roots means $P$ has a double zero, $P(s)=(s-s_1)^2$ , so $L\,y(x,s)=(s-s_1)^2x^s$ . Why: a repeated root is a second-order zero, and it is that extra power of $(s-s_1)$ that will survive one differentiation and still vanish at $s_1$ .

Step 3. Differentiate with respect to $s$ . Since $L$ carries no $s$ -dependence, $L\,\dfrac{\partial y}{\partial s}=\dfrac{\partial}{\partial s}\big[(s-s_1)^2x^s\big]=\big[2(s-s_1)+(s-s_1)^2\ln x\big]x^s$ . Why: the derivative passes straight through the linear operator, so $\partial y/\partial s$ is a solution of $Ly=0$ wherever the right-hand side vanishes.

Step 4. Evaluate at $s=s_1$ : both terms carry a factor $(s-s_1)$ , so the right-hand side is zero. Hence $y_2\equiv\partial_s y|_{s_1}$ satisfies $Ly_2=0$ and is a genuine second solution. Writing it out, $\partial_s\big[x^s\big]=x^s\ln x$ gives

y_2(x)=\left.\frac{\partial y(x,s)}{\partial s}\right|_{s=s_1}=y_1(x)\ln x+x^{s_1}\sum_{n\ge1}a_n'(s_1)\,x^{n}.

Why: differentiating the explicit factor $x^s$ is what conjures $y_1\ln x$ , while differentiating the coefficients supplies the accompanying regular series — the logarithm is forced, with coefficient exactly one. *Verified: $Ly_2=0$ follows from the double root; the form $y_2=y_1\ln x+(\text{series})$ matches Riley-Hobson-Bence ch. 16 and Bender & Orszag ch. 3.*

Limiting cases

The entire casework of the method is controlled by one number, the difference of the indicial roots $s_1-s_2$ (with $\operatorname{Re}s_1\ge\operatorname{Re}s_2$ ).

Case	Result	Meaning
Roots differ by a non-integer	$s_1-s_2\notin\mathbb{Z}$	Two independent pure Frobenius series, $y_1=(x-x_0)^{s_1}\sum a_n(x-x_0)^n$ and $y_2=(x-x_0)^{s_2}\sum b_n(x-x_0)^n$ . No logarithm — the simplest case.
Equal roots	$s_1=s_2$	Only one Frobenius series exists. The second is $y_2=y_1\ln(x-x_0)+(x-x_0)^{s_1}\sum_{n\ge1}c_n(x-x_0)^n$ ; a logarithm is always present.
Roots differ by a positive integer	$s_1-s_2=N\in\mathbb{Z}^+$	The larger root always gives a Frobenius series. The smaller root may give a second pure series, or the recurrence may fail at $n=N$ , forcing $y_2=k\,y_1\ln(x-x_0)+(x-x_0)^{s_2}\sum d_n(x-x_0)^n$ with possibly $k=0$ . Must be checked explicitly.
Irregular singular point	$(x-x_0)p\text{ or }(x-x_0)^2q\text{ non-analytic}$	Frobenius fails: no finite power balances the leading singularity. Solutions may carry essential singularities such as $e^{1/(x-x_0)}$ ; asymptotic (WKB) methods are needed instead.

Worked examples

1 · Distinct non-integer roots, worked to four coefficients (confidence builder)

Solve $2x^2y''-xy'+(1+x)y=0$ near its regular singular point $x_0=0$ : find the indicial roots and the first coefficients of both independent solutions.

Step 1. Put it in standard form: $p=-1/(2x)$ , $q=(1+x)/(2x^2)$ , so $xp=-\tfrac12$ gives $p_0=-\tfrac12$ and $x^2q=\tfrac12(1+x)$ gives $q_0=\tfrac12$ (with $q_1=\tfrac12$ ). The indicial equation is $s(s-1)-\tfrac12 s+\tfrac12=0$ , i.e. $2s^2-3s+1=(2s-1)(s-1)=0$ , so $\mathbf{s_1=1}$ and $\mathbf{s_2=\tfrac12}$ . Their difference is $\tfrac12$ — non-integer, so both solutions are pure Frobenius series with no logarithm.

Step 2. Substituting $y=\sum_n a_n x^{s+n}$ into the equation, the coefficient of $x^{s+n}$ is $a_n\,P(s+n)+a_{n-1}=0$ with $P(r)=2r^2-3r+1=(2r-1)(r-1)$ . Hence the one-term recurrence $a_n=-a_{n-1}/P(s+n)$ .

Step 3. For the larger root $s=1$ : $P(1+n)=(2n+1)(n)$ , so $a_n=-a_{n-1}/[n(2n+1)]$ . With $a_0=1$ this gives $a_1=-\tfrac13$ , $a_2=+\tfrac1{30}$ , $a_3=-\tfrac1{630}$ , so

y_1=x\Big(1-\tfrac13x+\tfrac1{30}x^2-\tfrac1{630}x^3+\cdots\Big).

Step 4. For the smaller root $s=\tfrac12$ : $P(\tfrac12+n)=(2n)(n-\tfrac12)=n(2n-1)$ , so $a_n=-a_{n-1}/[n(2n-1)]$ . With $a_0=1$ this gives $a_1=-1$ , $a_2=+\tfrac16$ , $a_3=-\tfrac1{90}$ , so

y_2=x^{1/2}\Big(1-x+\tfrac16x^2-\tfrac1{90}x^3+\cdots\Big).

The general solution is $y=Ay_1+By_2$ . Two clean series, exponents 1 and 1/2, no logarithm — the non-integer gap did all the work.

2 · An integer root gap that needs no logarithm

Solve $xy''+2y'+xy=0$ near $x_0=0$ . This is the $\ell=0$ spherical Bessel equation; show the indicial roots differ by an integer yet both give elementary closed forms.

Step 1. Here $xp=2$ so $p_0=2$ , and $x^2q=x^2$ so $q_0=0$ . The indicial equation is $s(s-1)+2s=s(s+1)=0$ , giving $\mathbf{s_1=0}$ and $\mathbf{s_2=-1}$ , a difference of $N=1$ — an integer, so a logarithm is possible and must be checked.

Step 2. Substituting the ansatz, the coefficient of $x^{s+n-1}$ gives $a_n(s+n)(s+n+1)+a_{n-2}=0$ , i.e. $a_n=-a_{n-2}/[(s+n)(s+n+1)]$ , a two-term recurrence.

Step 3. Larger root $s=0$ : $a_n=-a_{n-2}/[n(n+1)]$ . The $n=1$ relation forces $a_1=0$ , so odds vanish; $a_0=1$ gives $a_2=-\tfrac16$ , $a_4=+\tfrac1{120}$ , and

y_1=1-\tfrac{x^2}{6}+\tfrac{x^4}{120}-\cdots=\frac{\sin x}{x}.

Step 4. Smaller root $s=-1$ : the danger point is $n=N=1$ , where the recurrence reads $a_1\,(0)(1)=-a_{-1}=0$ . This is the consistent case $0=0$ : $a_1$ is simply free (its content is absorbed into $y_1$ ), so no logarithm is forced. Taking $a_0=1$ , $a_1=0$ and $a_n=-a_{n-2}/[(n-1)n]$ gives $a_2=-\tfrac12$ , $a_4=+\tfrac1{24}$ , and

y_2=x^{-1}\Big(1-\tfrac{x^2}{2}+\tfrac{x^4}{24}-\cdots\Big)=\frac{\cos x}{x}.

The general solution is $y=A\,\dfrac{\sin x}{x}+B\,\dfrac{\cos x}{x}$ . The integer gap $N=1$ did not force a logarithm — the check at $n=N$ came out consistent, and both solutions are elementary. Contrast this with Bessel of integer order, where the analogous check fails and the Neumann logarithm is unavoidable.

Physics of the portrait

The method carries the name of the mathematician who systematized it, but its roots run back through Fuchs's classification of singular points and forward into nearly every special function a physicist uses.

Origins · 1 minGeorg Frobenius and how this was discovered

Georg Frobenius

1849–1917 · series solutions at regular singular points

Ferdinand Georg Frobenius was a German mathematician, a student of Weierstrass in Berlin and later a professor at the ETH Zürich and the University of Berlin. Although he is best remembered today for founding much of representation theory and the theory of characters of finite groups — the Frobenius endomorphism, Frobenius reciprocity, the Frobenius determinant — his 1873 paper "Über die Integration der linearen Differentialgleichungen durch Reihen" gave the definitive treatment of series solutions about a regular singular point, including the delicate logarithmic cases that his predecessors had handled only piecemeal. He built directly on Lazarus Fuchs, who in the 1860s had classified the singular points of linear ODEs and proved when a fundamental system of solutions exists; the class of equations whose singularities are all regular is still called Fuchsian in his honour. Frobenius's contribution was to make the construction algorithmic: assume $y=(x-x_0)^s\sum a_n(x-x_0)^n$ , extract the indicial equation, run the recurrence, and treat the exceptional root differences by differentiating in the exponent. That procedure is what unifies Bessel's, Legendre's, Laguerre's, Hermite's and the hypergeometric equations into one framework, and it is why a nineteenth-century algebraist's name is attached to a computation every physics student now performs when solving the radial Schrödinger equation.

History

1782

— Legendre introduces the polynomials solving his differential equation

1824

— Bessel systematically studies the cylinder functions in perturbation theory

1857

— Riemann encodes the hypergeometric equation by its three singular points

1866

— Fuchs classifies regular versus irregular singular points and proves existence

1873

— Frobenius gives the systematic series method, logarithmic cases included

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. 16 (linear ODEs and the Frobenius method) · Arfken & Weber, Mathematical Methods for Physicists, 7th ed., ch. 7-8 (series solutions; Bessel and Legendre functions) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 16 (series solutions of ODEs) · Bender & Orszag, Advanced Mathematical Methods for Scientists and Engineers, ch. 3 (local analysis of ODEs) · Whittaker & Watson, A Course of Modern Analysis, ch. 10 (linear differential equations).

Provenance. A verifier should recompute: the indicial equation and recurrence from the Euler-form substitution (Derivation A); the Bessel indicial roots $\pm\nu$ , the recurrence $a_n=-a_{n-2}/[n(n+2\nu)]$ , and the $J_\nu$ series against Arfken & Weber ch. 14; the equal-root $\partial/\partial s$ argument against Riley-Hobson-Bence ch. 16 and Bender & Orszag ch. 3; and both worked examples — $2x^2y''-xy'+(1+x)y=0$ with roots $1,\tfrac12$ and coefficients $-\tfrac13,\tfrac1{30},-\tfrac1{630}$ (root $1$ ) and $-1,\tfrac16,-\tfrac1{90}$ (root $\tfrac12$ ); and $xy''+2y'+xy=0$ with roots $0,-1$ giving $\sin x/x$ and $\cos x/x$ . All special-function normalizations ( $J_\nu\sim(x/2)^\nu/\Gamma(\nu+1)$ , Legendre indicial roots both $0$ at $x=\pm1$ ) are standard textbook values.

Prerequisites. Linear Second-Order ODEs · Power Series Solutions · Ordinary and Singular Points · The Indicial Equation and Exponents · The Wronskian and Linear Independence. Leads to. Reduction of Order · Sturm-Liouville Theory · Two-Point Boundary-Value Problems · Existence and Uniqueness Theorems for ODEs.

Ask the tutor: Hand me any second-order ODE and a point, and I will classify the point, extract the indicial roots, and tell you at once whether the second solution hides a logarithm.

When exactly do I need the Method of Frobenius instead of an ordinary power series?

Use an ordinary Taylor series $\sum a_n(x-x_0)^n$ whenever $x_0$ is an ordinary point — that is, whenever $p(x)$ and $q(x)$ are both analytic there. Switch to Frobenius the moment $x_0$ is a regular singular point, meaning $p$ diverges no worse than $1/(x-x_0)$ and $q$ no worse than $1/(x-x_0)^2$ , so that $(x-x_0)p$ and $(x-x_0)^2q$ stay analytic. The tell is that a naive Taylor substitution produces an inconsistent leading order or demands a negative-power term. If even the Frobenius products are non-analytic, the point is irregular and neither series method works — you are in asymptotic-analysis territory.

Why does the difference of the indicial roots matter so much?

Because it decides whether two independent pure series can coexist. If $s_1-s_2$ is not an integer, the two ladders of exponents $\{s_1+n\}$ and $\{s_2+n\}$ never overlap, so each root generates its own clean series and independence is automatic. If the roots are equal, there is only one ladder and hence only one Frobenius series — the second solution must find room by growing a logarithm. If they differ by a positive integer $N$ , the two ladders collide at $n=N$ , where the recurrence for the smaller root divides by the vanishing $P(s_2+N)$ : sometimes consistently (a free coefficient, no log), sometimes fatally (a forced log). The single number $s_1-s_2$ thus routes you to the correct one of three procedures.

Where does the logarithm actually come from — is it a trick or is it forced?

It is forced, and the cleanest way to see it is to leave the exponent $s$ free, solve the recurrence for $a_n(s)$ , and differentiate the whole family with respect to $s$ . Because the recurrence cancels every coefficient above the leading one, applying the operator $L$ leaves only $P(s)x^s$ ; for a double root $P(s)=(s-s_1)^2$ , and one $s$ -derivative still vanishes at $s_1$ , so $\partial_s y|_{s_1}$ is a genuine second solution. Since $\partial_s x^s=x^s\ln x$ , that derivative automatically contains $y_1\ln(x-x_0)$ . The logarithm is the mathematical signature of a repeated root, exactly as a repeated root of a constant-coefficient equation forces a factor of $x$ .

An integer root gap — do I always get a logarithm?

No, and worked example 2 is the proof: $xy''+2y'+xy=0$ has roots $0$ and $-1$ , a gap of one, yet the solutions are the perfectly ordinary $\sin x/x$ and $\cos x/x$ with no logarithm anywhere. What happens is that at the collision index $n=N$ the recurrence reads $0\cdot a_N=(\text{something})$ ; if that something is also zero the equation is consistent, $a_N$ is free, and you get a second pure series. Only if the right-hand side is nonzero does the coefficient blow up, signalling that a logarithmic term must be inserted. So an integer gap means "check the $n=N$ coefficient explicitly", not "expect a logarithm".

How do I get the second solution when the recurrence genuinely fails?

Two reliable routes. The first is reduction of order: given the known solution $y_1$ , the formula $y_2=y_1\int^x \exp[-\int p\,dx']/y_1^2\,dx$ produces a second independent solution and reproduces any logarithm automatically, because the integrand develops a $1/(x-x_0)$ term that integrates to a log. The second is the substitution ansatz $y_2=k\,y_1\ln(x-x_0)+(x-x_0)^{s_2}\sum d_n(x-x_0)^n$ : plug it into the ODE, and matching powers fixes both the constant $k$ and all the $d_n$ . Reduction of order is more mechanical; the log-ansatz is often faster when you already suspect the form. Either way, Abel's identity $W=C\exp[-\int p\,dx]\ne0$ confirms the result is truly independent.

What guarantees the Frobenius series actually converges?

Fuchs's theorem. At a regular singular point $x_0$ , the constructed series converges at least throughout the largest open disc centred at $x_0$ that contains no other singular point of the equation, and its radius equals the distance to the nearest such singularity in the complex plane. For Bessel's equation the only finite singularity is the origin itself, so $J_\nu$ converges for all $x$ ; for Legendre's equation the singularities at $x=\pm1$ cap an expansion about the origin at $|x|<1$ ; for the hypergeometric equation the singularities at $0,1,\infty$ give unit radius about each finite one. Convergence is never in doubt inside that disc — the only question the method leaves open is the form of the second solution, not whether the series adds up.