Power Series Solutions

Given a linear second-order ODE with variable coefficients — Legendre's, Hermite's, Bessel's — there is usually no elementary closed form to write down. The power-series method answers the obvious next question: if you cannot guess the solution, can you build it, coefficient by coefficient? The strategy is disarmingly direct. Posit a Taylor expansion $y=\sum a_n(x-x_0)^n$ about a chosen base point, feed it into the equation, and let linearity do the work: differentiating term by term and collecting equal powers of $(x-x_0)$ turns a single differential equation into infinitely many algebraic equations for the coefficients. Because the monomials are linearly independent, each power's coefficient must vanish on its own, and that condition is a recurrence relation — a rule that manufactures every $a_n$ from the two free constants $a_0$ and $a_1$ . Three facts make this more than bookkeeping. First, the method works cleanly exactly at an ordinary point, where the normalized coefficients $P$ and $Q$ are analytic; there the two free constants are precisely the Cauchy data $y(x_0)$ and $y'(x_0)$ , so the recurrence constructs an explicit basis for the two-dimensional solution space. Second, convergence is guaranteed: Fuchs's theorem says the series converges at least out to the nearest singularity of the coefficients in the complex plane, which is why Legendre's series stops at $|x|=1$ . Third, and most physically, when the recurrence truncates the infinite series collapses to a polynomial — and demanding that truncation is exactly what quantizes angular momentum and energy.

Key takeaways

At an ordinary point $x_0$ (where $P$ and $Q$ are analytic) every solution of $y''+P(x)y'+Q(x)y=0$ is itself analytic and equals a convergent Taylor series $y=\sum_{n\ge0}a_n(x-x_0)^n$ .
Substituting the series and demanding that each power $(x-x_0)^k$ vanish gives a recurrence relation fixing all $a_n$ from the two free constants $a_0=y(x_0)$ and $a_1=y'(x_0)$ .
Fuchs's theorem guarantees the radius of convergence is at least the distance from $x_0$ to the nearest singularity of $P$ or $Q$ in the complex plane.
The two solutions seeded by $(a_0,a_1)=(1,0)$ and $(0,1)$ have Wronskian $W(x_0)=1\ne0$ , so they form a fundamental set.
When the recurrence truncates, the series becomes a polynomial — the origin of the Legendre and Hermite polynomials and of physical quantization.
At a singular point the bare series fails; one shifts to a Frobenius series $(x-x_0)^s\sum a_n(x-x_0)^n$ with $s$ from an indicial equation.
Key idea: at an ordinary point, write $y=\sum a_n(x-x_0)^n$ , force the coefficient of each power to vanish, and read off a recurrence that builds every $a_n$ from $a_0,a_1$ ; Fuchs guarantees the series converges up to the nearest complex singularity.

The core insight

At an ordinary point, write $y=\sum a_n(x-x_0)^n$ , substitute into the ODE, and force the coefficient of each power $(x-x_0)^k$ to vanish; this yields a recurrence that determines every $a_n$ from the free data $a_0,a_1$ , and Fuchs's theorem guarantees convergence out to the nearest singularity of the coefficients.

Where this lives today

The power-series method is not a museum piece; it is the computational bedrock under the special functions that pervade physics. Every scientific library that returns a Bessel, Airy, Legendre, or hypergeometric value — SciPy, the GSL, Boost, the routines behind the NIST DLMF — stitches together exactly three tools: a convergent power series near an ordinary point, a Frobenius series near a regular singular point, and an asymptotic expansion far away, switching between them where they overlap. In numerical analysis the "Taylor method" integrates ODEs by generating the local series coefficients on the fly through automatic differentiation, and high-order implementations (used in validated, interval-arithmetic integration of the solar system) routinely beat Runge–Kutta for smooth problems because the recurrence delivers arbitrarily many terms cheaply. At the research frontier the same idea reaches into general relativity: black-hole perturbation theory solves the Regge–Wheeler and Teukolsky equations — Fuchsian equations with regular singular points at the horizon and infinity — by matched series, the Mano–Suzuki–Takasugi method summing hypergeometric series to compute the gravitational-wave templates that LIGO and Virgo match against. And in computer algebra, the theory of holonomic (D-finite) functions treats a function as its recurrence — the very relation this article derives.

The physical picture

Think of the unknown solution not as a curve but as an infinite list of Taylor coefficients $a_0,a_1,a_2,\dots$ . The differential equation is a constraint that links these numbers: it says the local curvature $a_2$ is fixed once you know the value $a_0$ and slope $a_1$ , the next coefficient is fixed once you know those, and so on down the line. The recurrence is nothing but this "differentiate the equation and read off the next coefficient" process made explicit and iterative: two numbers to start, and the entire function unrolls. The first figure shows that unrolling for the Airy equation, where the coefficients couple in a three-step chain and one whole family is forced to zero.

Figure

The recurrence as an engine: for the Airy equation y″ = xy each coefficient is fixed three steps back, aₙ₊₃ = aₙ /((n+3)(n+2)). The free data a₀ and a₁ seed two independent chains — a₃ = a₀/6, a₆ = a₀/180 and a₄ = a₁/12, a₇ = a₁/504 — while a₂ = 0 kills the third family, so a₂ = a₅ = a₈ = ⋯ = 0

Why does the machine ever break down? At a singular point one of $P$ or $Q$ has a pole, so $y''=-Py'-Qy$ demands infinite curvature unless the solution behaves like $(x-x_0)^s$ or carries a $\ln(x-x_0)$ — behaviour no ordinary Taylor series can represent. Equally important, the reach of a convergent series is set not by anything you can see on the real axis but by the nearest singularity in the complex plane. Legendre's operator looks smooth at $x=0$ , yet once normalized its coefficients carry poles at $x=\pm1$ , so the series converges only for $|x|<1$ . The second figure makes this the whole story: a disc, centred on the expansion point, reaching to the closest obstruction.

Figure

Fuchs's theorem draws the disc of convergence: the series about x₀ = 0 converges at least out to the nearest singularity of the normalized coefficients P, Q in the complex plane. For the Legendre equation those sit at x = ±1; for coefficients built from 1/(1 + x²) they sit at x = ±i — both a distance 1 from the origin, so the radius R ≥ 1 in each case

The core equations

Everything begins with the equation put in standard form — divided through by the leading coefficient so the $y''$ term stands alone. The analytic character of the two remaining coefficient functions is what decides whether the method applies.

y''+P(x)\,y'+Q(x)\,y=0 \qquad (1)

where: $y(x)$ — the unknown solution; $P(x)=a_1(x)/a_2(x)$ — coefficient of $y'$ ; $Q(x)=a_0(x)/a_2(x)$ — coefficient of $y$ . A point $x_0$ is ordinary when both $P$ and $Q$ are analytic there, and singular otherwise. At an ordinary point the existence–uniqueness theory promises an analytic solution, so we are entitled to posit a Taylor series — the power-series ansatz. The whole method is the disciplined consequence of substituting this one line.

y(x)=\sum_{n=0}^{\infty}a_n\,(x-x_0)^n \qquad (2)

where: $a_n$ — the Taylor coefficients to be determined; $x_0$ — the ordinary point about which we expand. The solution is its list of coefficients, and the ODE will constrain them. Differentiating the series term by term is legitimate inside the radius of convergence and produces the derivative series that get substituted for $y'$ and $y''$ .

y'=\sum_{n=1}^{\infty} n\,a_n\,(x-x_0)^{n-1},\qquad y''=\sum_{n=2}^{\infty} n(n-1)\,a_n\,(x-x_0)^{n-2} \qquad (3)

where: $n(n-1)$ — the factor picked up by differentiating $x^n$ twice. An index shift $n\to n+2$ on the $y''$ sum realigns every series to a common power $(x-x_0)^n$ so coefficients can be compared. Two constants always survive the recurrence, and they are not arbitrary labels: they are exactly the Cauchy data that a second-order initial-value problem requires.

a_0=y(x_0),\qquad a_1=y'(x_0) \qquad (4)

where: $a_0$ — the initial value; $a_1$ — the initial slope. Reading it: choosing $(a_0,a_1)=(1,0)$ and $(0,1)$ generates two solutions, and their independence makes them a basis. That independence is certified once and for all by the Wronskian evaluated at the base point, where the two series and their derivatives take the simplest possible values.

W[y_1,y_2](x_0)=\begin{vmatrix} 1 & 0\\[2pt] 0 & 1\end{vmatrix}=1\neq 0 \qquad (5)

where: $y_1$ — the solution with $(a_0,a_1)=(1,0)$ ; $y_2$ — the solution with $(a_0,a_1)=(0,1)$ ; $W$ — their Wronskian. Reading it: $W\ne0$ at one point (hence everywhere, by Abel's identity) means $y_1,y_2$ span the solution space. How far does the constructed series reach? Fuchs's convergence bound answers this with a single geometric statement about the complex plane.

R_{\text{conv}}\ \ge\ \operatorname{dist}\big(x_0,\ \{\text{singularities of }P,Q\text{ in }\mathbb{C}\}\big) \qquad (6)

where: $R_{\text{conv}}$ — the radius of convergence of both series; $\operatorname{dist}$ — the complex-plane distance from $x_0$ to the nearest singular point of the coefficients. Reading it: even purely real problems inherit their radius from singularities that may lie off the real axis, such as $x=\pm i$ . Finally, the recurrences of the three workhorse equations of mathematical physics share the same DNA — each expresses one coefficient in terms of an earlier one — but differ in step size and truncation condition.

\underbrace{(n{+}2)(n{+}1)a_{n+2}=[n(n{+}1)-\ell(\ell{+}1)]a_n}_{\text{Legendre}},\quad \underbrace{a_{n+2}=\tfrac{2n-2\lambda}{(n+2)(n+1)}a_n}_{\text{Hermite}},\quad \underbrace{a_{n+3}=\tfrac{a_n}{(n+3)(n+2)}}_{\text{Airy}} \qquad (7)

where: $\ell$ — the Legendre parameter (truncates the series when $n=\ell$ ); $\lambda$ — the Hermite eigenvalue parameter (truncates when $\lambda=n$ ); the Airy recurrence has step 3 and never truncates. Reading it: two-term recurrences with a gap of 2 (Legendre, Hermite) decouple into an even and an odd series; the Airy gap of 3 decouples into three residue classes.

Derivation AThe general recurrence at an ordinary point — the engine

The goal is to strip away the particular equation and expose the machine that all of them run on: for any ODE whose normalized coefficients are analytic at the origin, we want a single closed formula that delivers the next coefficient $a_{n+2}$ from all the lower ones. The tool is the Cauchy product — the rule for multiplying two power series — applied to the products $P(x)y'(x)$ and $Q(x)y(x)$ . Once this general recurrence is in hand, every specific case (Legendre, Hermite, Airy) is just a matter of reading off the coefficients $p_k$ and $q_k$ of $P$ and $Q$ , and the two free constants $a_0,a_1$ visibly parametrize a two-dimensional space.

Step 1. Expand the coefficients as $P=\sum_k p_k x^k$ and $Q=\sum_k q_k x^k$ (this is legitimate precisely because $x_0=0$ is ordinary), write $y=\sum_n a_n x^n$ , and shift the index on $y''$ so it starts at $x^0$ : $y''=\sum_{n\ge0}(n+2)(n+1)a_{n+2}x^n$ . Why: realigning every series to the common power $x^n$ is what lets us equate coefficients power by power.

Step 2. Form $Py'$ as a Cauchy product. With $y'=\sum_m(m+1)a_{m+1}x^m$ , the coefficient of $x^n$ in the product $P y'$ is $\sum_{k=0}^{n}p_{n-k}(k+1)a_{k+1}$ . Why: the Cauchy product is exactly the statement that the $x^n$ term of a product collects every way of splitting $n$ between the two factors — this is where a variable coefficient couples many $a$ 's at once.

Step 3. Do the same for $Qy$ : the coefficient of $x^n$ is $\sum_{k=0}^{n}q_{n-k}a_k$ . Add the three contributions from $y''$ , $Py'$ , and $Qy$ , and set the total coefficient of $x^n$ to zero because the monomials are linearly independent. Why: forcing each power to vanish separately is the single logical move that turns one differential equation into an infinite tower of algebraic ones.

Step 4. Solving for the highest coefficient gives the engine, $(n+2)(n+1)a_{n+2}=-\sum_{k=0}^{n}\big[(k+1)p_{n-k}a_{k+1}+q_{n-k}a_k\big]$ . The prefactor $(n+2)(n+1)$ never vanishes, so $a_{n+2}$ is always determined; only $a_0$ and $a_1$ are left free. Why: the non-vanishing prefactor is the algebraic fingerprint of an ordinary point — at a singular point it would degenerate and the construction would stall.

Step 5. Specialize as a check. For Legendre, working from the un-normalized $(1-x^2)y''-2xy'+\ell(\ell+1)y=0$ reproduces $(n+2)(n+1)a_{n+2}=[n(n+1)-\ell(\ell+1)]a_n$ after using $n(n-1)+2n=n(n+1)$ ; for Hermite it gives $a_{n+2}=(2n-2\lambda)a_n/[(n+2)(n+1)]$ . Why: a general formula earns trust only by collapsing to the known special cases, which it does term by term. *Why it works: reading $p_k,q_k$ from the Legendre and Hermite equations reproduces exactly the recurrences of equation (7), and the prefactor $(n+2)(n+1)>0$ for all $n\ge0$ confirms that every higher coefficient is uniquely fixed by $a_0$ and $a_1$ .*

Go deeper · Derivation B · 2 minSolving the Airy equation about the origin

Derivation BSolving the Airy equation about the origin

The goal is to run the engine on the simplest non-trivial case — the Airy equation $y''=xy$ , the textbook model for the transition between oscillatory and exponential behaviour and the wavefunction near a linear potential's turning point — and to extract both independent solutions explicitly. Along the way we will see the characteristic three-step coupling, the forced vanishing of one whole family of coefficients, and the fact that because the coefficients $P=0$ and $Q=-x$ have no singularities anywhere, Fuchs's bound gives an infinite radius: the Airy functions are entire.

Step 1. Substitute $y=\sum a_n x^n$ into $y''-xy=0$ . The first term is $\sum_{n\ge0}(n+2)(n+1)a_{n+2}x^n$ ; the term $xy=\sum_n a_n x^{n+1}$ raises the power by one, so relabelling gives $\sum_{n\ge1}a_{n-1}x^n$ . Why: multiplication by $x$ is the whole source of the step-3 structure — it offsets the two sums by exactly one power.

Step 2. Read off the coefficient of $x^0$ . Only $y''$ contributes there, giving $2\cdot1\cdot a_2=0$ , hence $a_2=0$ . Why: nothing on the right feeds the $x^0$ slot, so the coefficient two steps up from $a_{-1}$ (which does not exist) is forced to zero — this single fact will annihilate an entire residue class.

Step 3. For $n\ge1$ the coefficient of $x^n$ gives $(n+2)(n+1)a_{n+2}=a_{n-1}$ , i.e. $a_{n+2}=a_{n-1}/[(n+2)(n+1)]$ . Relabelling $n\to n+1$ casts it in the clean form $a_{n+3}=a_n/[(n+3)(n+2)]$ , so coefficients couple only within residue classes modulo 3. Why: a gap of three between coupled coefficients splits the solution into three independent arithmetic threads.

Step 4. Seed the first thread with $a_0=1,\,a_1=0$ : then $a_3=a_0/(3\cdot2)=1/6$ , $a_6=a_3/(6\cdot5)=1/180$ , and so on — the even-of-mod-3 solution. Why: choosing $a_0=1,a_1=0$ isolates one basis solution; its coefficients shrink factorially, guaranteeing convergence for all $x$ .

Step 5. Seed the second thread with $a_0=0,\,a_1=1$ : then $a_4=a_1/(4\cdot3)=1/12$ , $a_7=a_4/(7\cdot6)=1/504$ , giving the second, independent solution. The third thread starts at $a_2=0$ and is therefore identically zero. Two independent series, both entire — exactly what Fuchs predicts when the coefficients have no finite singularity. Why: the two seeds produce a fundamental pair whose combinations are the Airy functions $\mathrm{Ai}$ and $\mathrm{Bi}$ . *Why it works: substituting either series back into $y''-xy=0$ satisfies it term by term; the explicit values $a_3=1/6,\ a_6=1/180$ and $a_4=1/12,\ a_7=1/504$ match the chains derived above, and since $P=0$ and $Q=-x$ are singularity-free everywhere, Fuchs's bound gives $R_{\text{conv}}=\infty$ .*

Go deeper · Derivation C · 2 minHermite truncation and the quantization of energy

Derivation CHermite truncation and the quantization of energy

The goal is to show that a convergence demand — an analytic requirement, seemingly far from physics — is the mechanism that quantizes the energy of the harmonic oscillator. We peel the Gaussian off the Schrödinger equation, land on Hermite's equation, and discover that its power-series solution blows up like $e^{x^2}$ for almost every energy. Only when the recurrence terminates into a polynomial does the wavefunction remain normalizable — and that termination is precisely the quantization rule $E_n=\hbar\omega(n+\tfrac12)$ .

Step 1. In the oscillator problem $-\tfrac{\hbar^2}{2m}\psi''+\tfrac12 m\omega^2 x^2\psi=E\psi$ , rescale to the dimensionless $\xi=\sqrt{m\omega/\hbar}\,x$ and write $\psi=H(\xi)e^{-\xi^2/2}$ to remove the known Gaussian tail. The residual factor $H$ obeys Hermite's equation $H''-2\xi H'+(\varepsilon-1)H=0$ with $\varepsilon=2E/\hbar\omega$ . Why: stripping the Gaussian isolates the part of the wavefunction that the power series must handle, converting a physics problem into the standard Hermite ODE.

Step 2. Matching to the form $H''-2\xi H'+2\lambda H=0$ identifies $2\lambda=\varepsilon-1$ , i.e. $\lambda=\dfrac{E}{\hbar\omega}-\dfrac12$ . The recurrence (equation 7) reads $a_{n+2}=\dfrac{2n-2\lambda}{(n+2)(n+1)}a_n$ . Why: the eigenvalue parameter $\lambda$ is just the energy in disguise, so whatever quantizes $\lambda$ quantizes $E$ .

Step 3. For large $n$ , $a_{n+2}/a_n\to 2/n$ , which is the coefficient ratio of $e^{\xi^2}$ . So unless the series stops, $H\sim e^{\xi^2}$ and $\psi=H e^{-\xi^2/2}\sim e^{+\xi^2/2}$ diverges — not normalizable. Why: the asymptotic ratio test exposes that a generic (non-terminating) solution is physically forbidden, forcing us to demand truncation.

Step 4. The series truncates exactly when the numerator vanishes at some $n$ , i.e. $\lambda=n$ for a non-negative integer $n$ . Then $H$ is the Hermite polynomial of degree $n$ and $\psi$ decays properly. Why: truncation is the only escape from the exponential blow-up, and it is available only at discrete values of $\lambda$ .

Step 5. Translate back: $\lambda=n$ means $E/\hbar\omega-\tfrac12=n$ , so $E_n=\hbar\omega\!\left(n+\tfrac12\right)$ , $n=0,1,2,\dots$ — the equally spaced spectrum with its zero-point energy $\tfrac12\hbar\omega$ . Why: the quantization of energy, one of the founding results of quantum mechanics, is here nothing but the condition that a power series be finite. *Why it works: $\lambda=(E/\hbar\omega)-\tfrac12$ with the truncation condition $\lambda=n$ gives $E_n=\hbar\omega(n+\tfrac12)$ ; the large- $n$ ratio $a_{n+2}/a_n\to2/n$ matches the Taylor coefficients of $e^{\xi^2}$ , confirming that the generic non-terminating solution is not normalizable.*

Limiting cases

The entire behaviour of the method is dictated by the analytic structure of $P$ and $Q$ at the base point, and by whether the recurrence happens to terminate.

Case	Result	Meaning
Ordinary point	$P,Q$ analytic at $x_0$	Plain Taylor series works; two independent analytic solutions; radius $\ge$ distance to nearest complex singularity.
Regular singular point	$(x-x_0)P,\ (x-x_0)^2Q$ analytic	Plain series fails; use a Frobenius series $(x-x_0)^s\sum a_n(x-x_0)^n$ with $s$ from the indicial equation.
Irregular singular point	worse than regular	Neither method applies directly; expect essential singularities and asymptotic (often divergent) expansions.
Truncating recurrence	$a_{n+2}=0$ from some $n$ onward	For special parameter values the series becomes a polynomial — the origin of Legendre and Hermite polynomials and of quantization.

Common trapThe bare Taylor series works everywhere, so long as I write down enough terms.

It does not. The ansatz $\sum a_n(x-x_0)^n$ is licensed only at an ordinary point. Push it onto a singular point and the leading algebra degenerates: the prefactor of the highest coefficient can vanish, forcing an inconsistent recurrence or silently dropping one of the two independent solutions (Bessel's equation at the origin is the classic case — the naive series recovers only $J_\nu$ and misses the second solution entirely). The cure is the Frobenius series $(x-x_0)^s\sum a_n(x-x_0)^n$ with the exponent $s$ fixed by an indicial equation. Two smaller traps travel with the method even at ordinary points: re-index every sum carefully when you shift $n\to n+2$ , and never read the radius of convergence off the real axis — the obstruction may sit at a complex point such as $x=\pm i$ that your real problem never visits.

Worked examples

1 · Legendre with ℓ = 2 — the polynomial falls out of the recurrence

Solve Legendre's equation $(1-x^2)y''-2xy'+\ell(\ell+1)y=0$ about the ordinary point $x_0=0$ for $\ell=2$ , and show that the even solution truncates to the Legendre polynomial $P_2(x)$ .

Step 1. The recurrence is $(n+2)(n+1)a_{n+2}=[n(n+1)-\ell(\ell+1)]a_n$ with $\ell=2$ , so $\ell(\ell+1)=6$ . Start the even chain from $a_0$ . Take $n=0$ : $2\cdot1\cdot a_2=[0-6]a_0$ , hence $a_2=-3a_0$ .

Step 2. Continue with $n=2$ : $4\cdot3\cdot a_4=[2\cdot3-6]a_2=[6-6]a_2=0$ , so $a_4=0$ — and therefore $a_6=a_8=\cdots=0$ . The even series has truncated after two terms: $y_{\text{even}}=a_0(1-3x^2)$ .

Step 3. Fix the normalization by the standard convention $P_\ell(1)=1$ . Here $y_{\text{even}}(1)=a_0(1-3)=-2a_0$ , so setting $-2a_0=1$ gives $a_0=-\tfrac12$ and $y_{\text{even}}=-\tfrac12+\tfrac32x^2=\dfrac{3x^2-1}{2}$ . This is exactly $\boxed{P_2(x)=\tfrac12(3x^2-1)}$ .

Step 4. The odd chain seeded by $a_1$ has numerator $n(n+1)-6$ , which never vanishes for odd $n$ , so that series does not truncate; it converges only on $|x|<1$ and diverges at $x=\pm1$ — the second Legendre function $Q_2(x)$ , singular at the poles. The integer $\ell$ selects which of the two independent series becomes a well-behaved polynomial — the mathematical seed of angular-momentum quantization.

2 · Radius of convergence from complex singularities

Where is the power-series solution of $(1+x^2)y''+2x\,y'-2y=0$ about $x_0=0$ guaranteed to converge, even though the coefficients are perfectly smooth for all real $x$ ?

Step 1. Put the equation in standard form by dividing by $(1+x^2)$ : $P(x)=\dfrac{2x}{1+x^2}$ and $Q(x)=\dfrac{-2}{1+x^2}$ . On the real line both are smooth — nothing warns of a limited radius.

Step 2. Find the singularities in the complex plane. Both $P$ and $Q$ are analytic except where the denominator vanishes: $1+x^2=0\Rightarrow x=\pm i$ . These are the only obstructions, and each lies a distance $|{\pm i}|=1$ from the origin.

Step 3. Apply Fuchs's bound (equation 6): $R_{\text{conv}}\ge\operatorname{dist}(0,\{\pm i\})=1$ . So the two series solutions are guaranteed to converge at least on the disc $|x|<1$ , even though a real-axis inspection reveals no reason for any limit at all — the lesson the amber points in the second figure encode.

Physics of the portrait

The convergence guarantee that turns "write down a series" into a theorem is Fuchs's, and with it the whole subject of linear ODEs in the complex domain — where an equation's global behaviour is read off from the arrangement of its singular points.

Origins · 1 minLazarus Fuchs and how this was discovered

Lazarus Fuchs

1833–1902 · theory of linear ODEs with analytic coefficients, the Fuchsian class

Lazarus Immanuel Fuchs was a German mathematician and a leading student of Karl Weierstrass in the Berlin school of complex analysis. Where earlier workers — Newton, Euler, Gauss — had used power series to solve differential equations case by case, Fuchs asked the structural question: for a linear ODE with analytic coefficients, what can you say about its solutions purely from the location and type of the coefficients' singular points in the complex plane? His foundational papers of 1866 established that near an ordinary point every solution is analytic, with a radius of convergence at least the distance to the nearest singularity — the theorem quoted in every derivation above — and classified singular points into the regular and irregular kinds. Equations all of whose singular points are regular are called Fuchsian in his honour, and they include almost every equation of mathematical physics: Legendre, hypergeometric, Bessel, Riemann's. His framework was the launch-pad for Frobenius's method (1873) and for Poincaré's theory of automorphic "Fuchsian functions," recasting the solution of differential equations as a chapter of complex function theory. Fuchs held chairs at Greifswald, Heidelberg, and finally Berlin, where he succeeded Kummer and edited Crelle's Journal. His work is why the power-series method comes with a warranty rather than a hope.

History

1669

— Newton solves equations by infinite series in his tract De analysi

1785

— Legendre introduces his polynomials in the study of gravitational attraction

1812

— Gauss analyses the hypergeometric series and its convergence

1866

— Fuchs proves the series converges out to the nearest singularity of the coefficients

1873

— Frobenius extends the method to regular singular points via the indicial equation

Going further

Texts. Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., ch. 7–8 (series solutions and Frobenius) · Hassani, Mathematical Physics, ch. on differential equations in the complex plane · Stone and Goldbart, Mathematics for Physics, ch. on ODEs and special functions · Byron and Fuller, Mathematics of Classical and Quantum Physics, ch. 4 · Riley, Hobson and Bence, Mathematical Methods for Physics and Engineering, series-solutions chapter · Boyce and DiPrima, Elementary Differential Equations, ch. 5. The general recurrence collapses to the Legendre recurrence $(n+2)(n+1)a_{n+2}=[n(n+1)-\ell(\ell+1)]a_n$ and the Hermite recurrence $a_{n+2}=(2n-2\lambda)a_n/[(n+2)(n+1)]$ ; the Airy chains give $a_3=a_0/6,\ a_6=a_0/180,\ a_4=a_1/12,\ a_7=a_1/504$ ; the $\ell=2$ truncation yields $P_2(x)=\tfrac12(3x^2-1)$ ; and the oscillator truncation condition $\lambda=n$ gives $E_n=\hbar\omega(n+\tfrac12)$ .

Prerequisites. Linear Second-Order ODEs · Ordinary and Singular Points · Existence and Uniqueness Theorems · The Wronskian and Linear Independence. Leads to. The Method of Frobenius · The Indicial Equation and Exponents · Sturm–Liouville Theory.

Ask the tutor: Run your own equation past me — pick a base point and I will grind out the coefficients, test whether they truncate, and pin down the radius of convergence from its singularities.

Why does forcing every coefficient to zero actually solve the equation?

Because a power series that equals zero on an interval must have every one of its coefficients equal to zero — the monomials $\{(x-x_0)^n\}$ are linearly independent functions. When you substitute the series into the ODE and gather terms, the left side becomes one big power series that must vanish identically. Linear independence then converts that single statement into infinitely many separate statements, one per power: the coefficient of $(x-x_0)^0$ is zero, of $(x-x_0)^1$ is zero, and so on. Each of those is an algebraic equation among the $a_n$ , and together they are the recurrence. So the "trick" is really a theorem about the uniqueness of power-series representations.

What operationally distinguishes an ordinary point from a singular one?

Put the equation in standard form $y''+P y'+Q y=0$ and look at $P$ and $Q$ at that point. If both are analytic there — representable by their own convergent Taylor series — the point is ordinary and the plain power-series method works. If either blows up, the point is singular. Singular points come in two grades: regular, where the milder combinations $(x-x_0)P$ and $(x-x_0)^2Q$ are analytic (Frobenius handles these), and irregular, where even those blow up (asymptotic methods needed). The practical test is fast: form $P$ and $Q$ , check for poles, and if you find one, see whether the right power of $(x-x_0)$ tames it.

Legendre's equation has polynomial coefficients — why doesn't its series converge everywhere?

Because "polynomial coefficients" refers to the un-normalized equation $(1-x^2)y''-2xy'+\ell(\ell+1)y=0$ . Fuchs's theorem is a statement about the standard-form coefficients $P=-2x/(1-x^2)$ and $Q=\ell(\ell+1)/(1-x^2)$ , obtained after dividing by the leading term $(1-x^2)$ . That division introduces poles at $x=\pm1$ , the zeros of $(1-x^2)$ . The nearest of these to the expansion point $x_0=0$ sits a distance 1 away, so the series is guaranteed only on $|x|<1$ and indeed diverges at the poles $x=\pm1$ — unless the recurrence truncates into a polynomial, which sidesteps the issue by having no infinite tail to diverge.

How can a demand for convergence "quantize" a physical quantity?

This is the deepest payoff of the method. Physically acceptable wavefunctions must be normalizable, but the generic power-series solution of Hermite's equation grows like $e^{x^2}$ , which ruins normalizability once the Gaussian factor is restored. The only escape is for the series to stop — to become a polynomial — and the recurrence terminates only when its parameter $\lambda$ equals a non-negative integer. Translating $\lambda=n$ back into energy gives $E_n=\hbar\omega(n+\tfrac12)$ . So the discreteness of the spectrum is not imposed by hand; it is forced by demanding that the series converge to an admissible function.

When exactly must I abandon the plain series for Frobenius?

The moment the base point is a regular singular point rather than an ordinary one. Symptomatically, if you push the plain ansatz onto such a point the leading algebra misbehaves: the coefficient that should determine $a_0$ instead forces $a_0=0$ , or the recurrence becomes inconsistent, or you get only one solution when you know there should be two. The reason is that the true solution behaves like $(x-x_0)^s$ with a non-integer (or repeated) exponent $s$ , or carries a logarithm — neither of which an integer-power Taylor series can represent. Frobenius fixes this by promoting the ansatz to $(x-x_0)^s\sum a_n(x-x_0)^n$ and determining $s$ from a quadratic indicial equation before the recurrence even begins.

Does it matter which point x₀ I expand about?

Yes, in two ways. First, $x_0$ must be an ordinary point for the plain method to apply — expanding about a singular point demands Frobenius instead. Second, among ordinary points the choice sets the radius of convergence, since Fuchs's bound is the distance from your $x_0$ to the nearest singularity: expanding closer to a singularity gives a smaller disc. If you need the solution over a wide interval you may expand about several base points and analytically continue. The coefficients change with $x_0$ , but the underlying function does not.