Series Solutions Near Ordinary Points

Can a differential equation with no elementary closed-form solution still be solved exactly? Consider $y'' - 2xy' + \lambda y = 0$ , the Hermite equation. It has no solution expressible in terms of $\sin$ , $\cos$ , $e^x$ , or any finite combination of familiar functions. Yet this equation, and thousands like it arising in quantum mechanics, heat flow, and vibration problems, can be solved completely. The method: assume the solution is a power series and grind out its coefficients term by term. This is not a trick reserved for special cases: it is a fully general method. It is guaranteed to work whenever the equation's coefficients are themselves well-behaved (analytic) at the point of interest. The big idea is deceptively simple. Write $y = \sum_{n=0}^\infty a_n (x-x_0)^n$ , substitute into the ODE, and collect like powers of $(x - x_0)$ . Then force each coefficient to vanish. This turns a differential equation into an algebraic recurrence relation among the $a_n$ . It can be solved recursively, starting from $a_0$ and $a_1$ (for a second-order equation) as free parameters. The entire content of the theory is in showing when this formal procedure actually converges to a genuine solution. It also shows how far out the resulting series is trustworthy. This article develops the theory of series solutions at ordinary points — points where the equation's coefficient functions are analytic and the leading coefficient does not vanish. We state precisely what an ordinary point is. We prove the fundamental existence theorem guaranteeing a convergent power-series solution there. We also work through the mechanics of building recurrence relations by hand. Along the way we address a natural question. What happens at points where the leading coefficient does vanish? These points require the more delicate Frobenius machinery developed elsewhere. By the end, you will be able to recognize an ordinary point on sight. You will be able to set up and solve the recurrence relation for any linear ODE with polynomial or analytic coefficients. You will also determine a guaranteed radius of convergence without summing a single series. Finally, you will understand exactly why the method can fail if its hypotheses are dropped.

Key takeaways

An ordinary point $x_0$ of $y'' + p(x)y' + q(x)y = 0$ is one where $p,q$ are both analytic at $x_0$ .
At an ordinary point, every solution is analytic there: it equals a convergent power series $\sum a_n(x-x_0)^n$ .
Substituting the series into the ODE and matching powers of $(x-x_0)$ produces a recurrence relation for the coefficients $a_n$ , $n \geq 2$ , in terms of earlier ones.
The two free parameters $a_0 = y(x_0)$ and $a_1 = y'(x_0)$ generate two linearly independent solutions, matching the dimension guaranteed by Higher-Order Linear Equations theory.
The guaranteed radius of convergence of the series solution is at least the distance from $x_0$ to the nearest singular point of $p, q$ in the complex plane.
Points where the leading coefficient of the ODE vanishes are generally not ordinary; they require the Frobenius extension of this method.
Key idea: analytic coefficients force analytic solutions, and matching power-series coefficients converts a differential equation into a solvable algebraic recurrence.

Where this lives today

Series solutions at ordinary points are the entry point to the vast family of special functions of mathematical physics: Hermite polynomials (quantum harmonic oscillator), Legendre polynomials (angular momentum, potential theory, spherical harmonics), and Chebyshev polynomials (numerical approximation and spectral methods) all arise as polynomial solutions produced by exactly this recurrence-relation method, often when the recursion terminates for special parameter values. In numerical analysis, truncated power-series solutions provide high-order Taylor-method integrators for ODEs, and the guaranteed radius of convergence from the theorem gives a rigorous a priori error bound for such schemes. In control theory and engineering, series solutions supply local approximations to the response of linear time-varying systems whose coefficients are analytic in time. The method also underlies symbolic computation: computer algebra systems generate solutions to linear ODEs with polynomial coefficients by implementing exactly the recurrence relations derived here, and the distinction between ordinary and singular points (foreshadowing Regular Singular Points and Frobenius) determines which algorithm the software must invoke.

Ask the tutor: How do I know in advance whether the recurrence relation for a given equation will "telescope" into a closed-form function like $e^{x^2/2}$ , or whether it defines a genuinely new special function?

y'' + p(x)y' + q(x)y = 0

Intuition

Analyticity is a strong form of "niceness": a function is analytic at $x_0$ exactly when it locally coincides with its own Taylor series, not just approximately but exactly, on some interval of positive radius. Polynomials, $e^x$ , $\sin x$ , $\cos x$ , and $1/(1-x)$ (away from $x=1$ ) are all analytic wherever they are defined and smooth. Analyticity fails at genuine singularities — $\tan x$ at $x = \pi/2$ , or $1/x$ at $x=0$ — where the function blows up or otherwise cannot be captured by any convergent power series centered there. The heart of the theorem below is a conservation-of-niceness principle: if the "control functions" $p$ and $q$ in the ODE are analytic at $x_0$ , that niceness is inherited by the solution $y$ . This is not obvious a priori — differentiation and solving equations can in general destroy smoothness — but linear ODEs with analytic coefficients are special: they propagate analyticity rather than destroy it. This is the same phenomenon that makes analytic coefficients so much more tractable than merely continuous ones, and it is what licenses the entire "guess a power series and match coefficients" strategy. Geometrically, picture the real line with the singular points of $p$ and $q$ marked as forbidden posts. An ordinary point $x_0$ sits away from these posts. The theorem guarantees that the power-series solution centered at $x_0$ converges at least out to the nearest post — think of a circle in the complex plane, centered at $x_0$ , expanding until it first grazes a singularity.

Figure

Radius of convergence is governed by the nearest singularity in the complex plane

For example, for the equation $(1+x^2) y'' + xy' + y = 0$ , dividing by $1+x^2$ shows $p, q$ are analytic everywhere except where $1+x^2 = 0$ , i.e. at $x = \pm i$ . The origin $x_0 = 0$ is an ordinary point, and the guaranteed radius of convergence of the series solution about $0$ is $1$ , the distance from $0$ to $\pm i$ — even though the equation and its solutions are entirely real.

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

x_0 = 0

(a_0,a_1) = (1,0)

Worked examples

Example 1 (asymmetric linear-coefficient equation). Solve $y'' - xy' - y = 0$ near the ordinary point $x_0 = 0$ (here $p(x) = -x$ , $q(x) = -1$ , both entire, so $\rho = \infty$ ), finding both fundamental solutions explicitly and testing initial data $y(0) = 2$ , $y'(0) = -1$ . Write $y = \sum_{n=0}^\infty a_n x^n$ . Then

y'' = \sum_{n=0}^\infty (n+2)(n+1)a_{n+2}x^n, \qquad xy' = \sum_{n=1}^\infty n a_n x^n = \sum_{n=0}^\infty n a_n x^n \ \ (\text{the }n=0\text{ term is }0).

The ODE $y'' - xy' - y = 0$ gives, coefficient by coefficient,

(n+2)(n+1)a_{n+2} - n a_n - a_n = 0 \quad\Longrightarrow\quad a_{n+2} = \frac{(n+1)a_n}{(n+2)(n+1)} = \frac{a_n}{n+2}, \qquad n \geq 0.

This recurrence separates into even and odd chains. Even chain ( $a_0$ free): $a_2 = a_0/2$ , $a_4 = a_2/4 = a_0/8$ , $a_6 = a_4/6 = a_0/48$ . In general $a_{2k} = a_0 / (2^k k!)$ , verified by induction: $a_{2(k+1)} = a_{2k}/(2k+2) = \dfrac{a_0}{2^k k!\,(2k+2)} = \dfrac{a_0}{2^{k+1}(k+1)!}$ . So

y_1(x) = a_0 \sum_{k=0}^\infty \frac{x^{2k}}{2^k k!} = a_0 \sum_{k=0}^\infty \frac{(x^2/2)^k}{k!} = a_0\, e^{x^2/2}.

This series is recognized in closed form — a pleasant bonus, though not guaranteed in general. Odd chain ( $a_1$ free): $a_3 = a_1/3$ , $a_5 = a_3/5 = a_1/15$ , $a_7 = a_5/7 = a_1/105$ . In general $a_{2k+1} = \dfrac{a_1}{(2k+1)!!} = \dfrac{a_1 \, 2^k k!}{(2k+1)!}$ , so

y_2(x) = a_1 \sum_{k=0}^\infty \frac{x^{2k+1}}{(2k+1)!!} = a_1\, x\sum_{k=0}^\infty \frac{2^k k!}{(2k+1)!} x^{2k}.

This one does not simplify to an elementary closed form, but the theorem guarantees it converges for all $x$ (since $\rho = \infty$ ) — a genuine new special function generated purely by the recurrence. For the initial data $y(0)=2$ , $y'(0) = -1$ : since $a_0 = y(0) = 2$ and $a_1 = y'(0) = -1$ , the solution is

y(x) = 2e^{x^2/2} - \sum_{k=0}^\infty \frac{x^{2k+1}}{(2k+1)!!}.

As a check at low order: $a_0 = 2$ , $a_1=-1$ , $a_2 = a_0/2 = 1$ , $a_3 = a_1/3 = -1/3$ , $a_4 = a_2/4=1/4$ . Direct substitution of $y \approx 2 - x + x^2 - \tfrac13 x^3 + \tfrac14 x^4$ into $y'' - xy' - y$ and checking the $x^0, x^1, x^2$ coefficients vanish confirms the arithmetic: $y'' = 2 - 2x + 3x^2+\cdots$ , $xy' = -x+2x^2 - x^3 + \cdots$ , $y'' - xy' - y = (2-2)+(-2+1+1)x + (2-2-1)x^2 + \cdots$ ; the $x^0$ term is $0$ , and the recurrence guarantees all higher terms vanish by construction.

Example 2 (non-constant, asymmetric coefficients requiring a genuine ordinary-point check). Consider the Airy-type variant $y'' + (2 + x)y' + xy = 0$ centered at $x_0 = 1$ (chosen off-center deliberately to avoid symmetry). Here $P(x) = 1$ , so $x_0 = 1$ is an ordinary point for any finite $x_0$ (the coefficients $2+x$ and $x$ are polynomials, hence entire), and $\rho = \infty$ . Let $t = x - 1$ , so $x = t+1$ , and set $y = \sum_{n=0}^\infty a_n t^n$ . Substituting $x = t+1$ :

2 + x = 3 + t, \qquad x = 1 + t.

The ODE becomes

y'' + (3+t) y' + (1+t) y = 0.

Compute:

y'' = \sum_{n\geq 0} (n+2)(n+1) a_{n+2} t^n,

(3+t)y' = 3\sum_{n\geq 0}(n+1)a_{n+1} t^n + \sum_{n\geq 1} n a_n t^n,

(1+t) y = \sum_{n \geq 0} a_n t^n + \sum_{n\geq 1} a_{n-1} t^n.

Matching the coefficient of $t^n$ for $n \geq 1$ :

(n+2)(n+1)a_{n+2} + 3(n+1)a_{n+1} + n a_n + a_n + a_{n-1} = 0,

a_{n+2} = -\frac{3(n+1) a_{n+1} + (n+1) a_n + a_{n-1}}{(n+1)(n+2)}, \qquad n \geq 1,

and separately for $n=0$ (where the $a_{n-1}$ term is absent):

2 a_2 + 3 a_1 + a_0 = 0 \quad \Longrightarrow \quad a_2 = -\frac{3a_1 + a_0}{2}.

Take, asymmetrically, $a_0 = 1$ , $a_1 = 0$ (this is the solution with $y(1) = 1$ , $y'(1) = 0$ ). Then:

a_2 = -\frac{0 + 1}{2} = -\frac12.

For $n=1$ : $a_3 = -\dfrac{3\cdot 2\, a_2 + 2 a_1 + a_0}{2\cdot 3} = -\dfrac{6(-\tfrac12) + 0 + 1}{6} = -\dfrac{-3+1}{6} = \dfrac{2}{6} = \dfrac13.$ For $n=2$ : $a_4 = -\dfrac{3\cdot 3\, a_3 + 3a_2 + a_1}{3\cdot 4} = -\dfrac{9(\tfrac13) + 3(-\tfrac12) + 0}{12} = -\dfrac{3 - 1.5}{12} = -\dfrac{1.5}{12} = -\dfrac{1}{8}.$ So near $x_0=1$ ,

y(x) = 1 - \frac12 (x-1)^2 + \frac13 (x-1)^3 - \frac18(x-1)^4 + \cdots,

valid (converging) for all real $x$ , since $\rho = \infty$ . This example shows the recurrence handling three-term dependencies ( $a_{n+2}$ depending on $a_{n+1}, a_n, a_{n-1}$ ) and a genuinely asymmetric initial condition, exactly as the general theorem anticipates — no elementary closed form is expected or found here.

x^2 y'' - x y' + y = 0

Scope, edge cases & common traps

Situation	What to check	Typical pitfall
$P(x)y''+Q(x)y'+R(x)y=0$ , want ordinary point at $x_0$	$P,Q,R$ analytic at $x_0$ (e.g. polynomials always are) and $P(x_0)\neq 0$	Forgetting to check $P(x_0) \neq 0$ even when $Q,R$ look harmless
Determining guaranteed radius of convergence	Find distance in $\mathbb{C}$ , not just $\mathbb{R}$ , from $x_0$ to nearest zero of $P$	Ignoring complex zeros (e.g. $1+x^2$ has none real, but $\pm i$ still bound $\rho$ )
Setting up the recurrence	Shift indices carefully so all sums start at the same power of $(x-x_0)$	Off-by-one errors in reindexing $y'$ , $y''$ ; mismatched lower limits
Two free constants $a_0,a_1$	These correspond exactly to $y(x_0), y'(x_0)$	Treating $a_0,a_1$ as arbitrary series coefficients unconnected to initial data
Coefficients not polynomial (e.g. $\sin x$ , $e^x$ )	Still fine, as long as analytic; must also expand $p,q$ in series and use Cauchy products	Trying to get an exact finite-term recurrence when one does not exist — a longer, still valid, convolution recurrence should be used
Point where $P(x_0)=0$	Not an ordinary point in general	Blindly applying the ordinary-point recurrence, silently losing a solution (see counterexample)

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered

Augustin-Louis Cauchy

1789–1857 · Early 19th-century France, foundations of complex analysis and rigorous ODE theory

Cauchy developed the majorant method used in the proof above — bounding an unknown series by an explicitly solvable comparison series — as part of his broader program to place calculus and differential equations on a rigorous footing. His existence theorems for differential equations, together with his foundational work on convergence of series and complex analysis, directly underlie the modern theorem that analytic coefficients force analytic solutions.

History

- 1821: Cauchy publishes rigorous foundations of limits and convergence in Cours d'Analyse, setting the stage for rigorous treatment of power series.

- 1830s: Cauchy develops existence theorems for ODEs with analytic coefficients, introducing majorant/dominant-function comparison arguments.

- 1840s: Cauchy and later Kovalevskaya extend analytic existence theory to broader classes of differential equations (the Cauchy–Kovalevskaya framework).

- 1866–1873: Georg Frobenius and Lazarus Fuchs develop the extension of series methods to regular singular points, handling cases like the Euler equation counterexample above.

- Late 19th century: Hermite, Legendre, Bessel, and Chebyshev equations become standard textbook applications of ordinary-point series solutions in mathematical physics.

What exactly distinguishes an "ordinary point" from a "regular point of the domain"?

Every point in the domain where an ODE is defined is not automatically ordinary. "Ordinary point" is a technical condition specifically about the coefficient functions $p(x) = Q(x)/P(x)$ and $q(x) = R(x)/P(x)$ in the standard form $y'' + p y' + q y = 0$ : they must both be analytic there. This typically fails exactly where the leading coefficient $P(x)$ vanishes, since division by zero (or by something vanishing to some order) tends to introduce poles into $p$ or $q$ . A point can look perfectly innocuous — the original equation $P,Q,R$ might all be polynomials with no obvious problem — yet still be singular if $P(x_0)=0$ , as in the Euler equation counterexample.

Why must we check the complex plane for singularities, not just the real line, when finding the radius of convergence?

Power series convergence is fundamentally a complex-analytic phenomenon: the radius of convergence of $\sum c_n(x-x_0)^n$ equals the distance from $x_0$ to the nearest point in the complex plane where the function being expanded fails to be analytic. For $1/(1+x^2)$ , there is no real singularity, but at $x = \pm i$ the function blows up. Any power series centered at a real point $x_0$ expanding this function is blind to the fact that the singularity is complex — but the convergence radius still shrinks to reflect it. This is a manifestation of a deep fact from complex analysis: real-variable intuition about smoothness is insufficient to predict convergence radii.

Can the method produce all solutions, or just some?

At a genuine ordinary point, the method produces all solutions: the corollary to the theorem shows the two series solutions built from $(a_0,a_1)=(1,0)$ and $(0,1)$ form a fundamental set, so every solution near $x_0$ is a linear combination of these two. This is a complete solution of the equation locally, in sharp contrast to singular points, where power series in $(x-x_0)$ alone may capture only part (sometimes none) of the solution space, as the Euler-equation counterexample demonstrates.

Do I have to expand

p(x)

and

q(x)

as full power series if they are not polynomials?

Yes, in principle: if $p, q$ are transcendental but analytic (like $\sin x$ or $e^x$ ), you expand them in their own Taylor series and use the Cauchy product formula to combine them with the series for $y', y''$ , exactly as in the proof's Step 1. This produces a recurrence for $a_{n+2}$ in terms of all previous coefficients (a convolution sum), rather than a short finite recurrence. It is more laborious than the polynomial-coefficient case but no different in principle, and the existence theorem still guarantees convergence out to the nearest complex singularity of $p, q$ .

What is the difference between this and the Frobenius method?

This article's method applies exactly at ordinary points, producing solutions that are honestly analytic (power series with only nonnegative integer exponents) with no logarithms. The Frobenius method extends the strategy to regular singular points, where $p(x)$ may have a simple pole and $q(x)$ at worst a double pole; there, solutions generally take the form $(x-x_0)^r$ times a power series, possibly with logarithmic terms, as illustrated starkly by the $x\ln x$ solution in the counterexample above. The ordinary-point theory is the foundation on top of which Frobenius theory is built, since the "nice" part of the Frobenius solution is still constructed by matching power-series coefficients.

a_0=a_1=0

, does the method just give the trivial solution?

Yes — and this is expected, not a defect. The map from initial data $(a_0,a_1) = (y(x_0), y'(x_0))$ to the solution is linear and, by the existence-uniqueness theory for linear ODEs, injective in the sense that the only solution with $y(x_0)=y'(x_0)=0$ is $y\equiv 0$ . The recurrence relation, when started from $a_0=a_1=0$ , forces every subsequent $a_n=0$ as well, consistent with uniqueness.

Is it always possible to sum the resulting series in closed form, like

e^{x^2/2}

in Example 1?

No, and this should not be expected in general. Example 1's odd-indexed solution and Example 2's solution do not reduce to elementary functions; this is the normal situation, not an exception. Many of the most important special functions in mathematical physics (Hermite, Legendre, Bessel, Airy functions) are literally defined as the power series produced by this method, precisely because they cannot be written in terms of elementary functions. The power-series representation itself, with its guaranteed convergence, is the final, complete answer.