Second-Order Linear Equations

Suppose a mass on a spring is pushed by an external force while friction slows it down. Can we describe every possible motion using just two building-block solutions? This question is at the heart of the theory of second-order linear equations. Specifically, it asks how the solution set of a linear differential equation is organized. The equation $y'' + p(t) y' + q(t) y = g(t)$ governs an enormous range of physical systems. These range from oscillating circuits to vibrating beams. Its structure is the prototype for all of linear differential equations theory. The big picture is that linearity imposes an algebraic skeleton on the solution set. Solutions to the homogeneous equation form a vector space of dimension two. The general solution to the full equation is that space shifted by a single particular solution. This article builds that structure rigorously from existence and uniqueness. It proves the superposition principle. It also shows exactly how much of the solution set two initial conditions can pin down. It sets the stage for the concrete methods that make this structure computable. These methods are Homogeneous Equations with Constant Coefficients and The Wronskian and Linear Independence.

Key takeaways

A second-order linear ODE has the form $y'' + p(t)y' + q(t)y = g(t)$ with $p, q, g$ continuous.
Homogeneous solutions ( $g \equiv 0$ ) form a real vector space; superposition holds because the map $y \mapsto y'' + py' + qy$ is linear.
The general solution of the nonhomogeneous equation is (particular solution) + (general homogeneous solution).
Existence-uniqueness (from Picard–Lindelöf) forces the homogeneous solution space to be exactly 2-dimensional.
Two solutions form a basis exactly when their Wronskian is nonzero at one point.
Key idea: linearity turns solving a differential equation into linear algebra — find a spanning set of two homogeneous solutions and add one particular solution.

Where this lives today

The mass-spring-damper equation $m y'' + c y' + k y = F(t)$ and the series RLC circuit equation are both instances of $y''+py'+qy=g$ ; the structure theorem is why engineers can solve "find the response to a step plus a sinusoid" by superposing two simpler particular solutions, exactly as in Example 2. In quantum mechanics, the time-independent Schrödinger equation in one dimension is a second-order linear ODE in the spatial variable, and the vector-space structure of its solution set underlies the classification of bound states. In control theory and signal processing, transfer-function methods for linear time-invariant systems rest on the same superposition principle, now packaged via Laplace transforms. In numerical analysis, boundary value problems built on this equation (e.g. shooting methods) exploit the two-dimensional solution space to reduce a BVP to a root-finding problem in two unknowns $(c_1,c_2)$ . The theorem is also the base case for the general theory of linear $n$ -th order and linear systems of ODEs, where the same proof (via a first-order system and Picard–Lindelöf) yields an $n$ -dimensional homogeneous solution space.

:::quiz[Which property of the operator $L[y]=y''+p(t)y'+q(t)y$ is the direct cause of the superposition principle?]

$L$ is invertible
$L$ is linear ✓
$L$ has constant coefficients
$L$ has continuous coefficients

:::quiz[If $y_1, y_2$ solve $L[y]=0$ but their Wronskian is identically zero, what can we conclude about the general solution formula in Part (4)?]

It still applies without changes
It fails to capture all solutions, since $y_1,y_2$ are not independent and hence not a basis of $V$ ✓
It applies only for $t>0$
It implies $g$ must be zero

Ask the tutor: How do we find explicit independent homogeneous solutions when $p(t)$ and $q(t)$ are constants, and why does the characteristic equation always produce a valid basis?

I \subseteq \mathbb{R}

Intuition

Think of $L$ as a machine: feed it a twice-differentiable function, and it outputs another function. The crucial structural fact is that $L$ is linear: it respects addition and scalar multiplication of functions, exactly as a matrix respects addition and scaling of vectors. This is not an approximation or a special property of nice equations — it follows immediately from the fact that differentiation is linear and that $p, q$ multiply $y', y$ without touching them nonlinearly (no $y^2$ , no $\sin(y)$ , no $y \cdot y'$ ). Because $L$ is linear, the homogeneous solution set $\{y : L[y] = 0\}$ is the kernel of $L$ , hence automatically a vector space: sums and scalar multiples of homogeneous solutions are again homogeneous solutions. The question the theory must answer is: how big is this kernel? Existence-uniqueness for the associated first-order system answers this precisely — the kernel is exactly two-dimensional, matching the two initial conditions $(y(t_0), y'(t_0))$ needed to pin down a unique solution. Once we have any two independent homogeneous solutions $y_1, y_2$ , every homogeneous solution is a combination $c_1 y_1 + c_2 y_2$ , and if $y_p$ is one particular solution of $L[y]=g$ , then every solution of $L[y] = g$ is $y_p + c_1 y_1 + c_2 y_2$ . This is the superposition principle in its full form, and the picture below shows it as an affine shift of a plane inside the infinite-dimensional space of functions.

:::svg[The solution set of $L[y]=g$ is a 2-dimensional plane through $y_p$ , parallel to the homogeneous plane through $0$ ]

The central theorem

I \subseteq \mathbb{R}

x_1 = y

t_0\in I

\Phi(y_1) = (1,0)

Worked examples

Example 1 — Verifying the structure theorem on an equation with variable coefficients. Consider on $I = (0,\infty)$ the Euler equation

t^2 y'' - 2t y' + 2y = t^3, \qquad t>0.

Divide by $t^2$ (legitimate since $t>0$ ) to put it in standard form:

y'' - \frac{2}{t} y' + \frac{2}{t^2} y = t, \qquad p(t) = -\frac{2}{t},\ q(t)=\frac{2}{t^2},\ g(t)=t.

All three are continuous on $(0,\infty)$ , so the Structure Theorem applies on this interval (note: $t=0$ must be excluded because $p,q$ blow up there). Homogeneous solutions. Try $y=t^r$ : then $y'=rt^{r-1}$ , $y''=r(r-1)t^{r-2}$ , and

t^2\cdot r(r-1)t^{r-2} - 2t\cdot rt^{r-1} + 2t^r = \big[r(r-1) - 2r + 2\big]t^r = (r^2-3r+2)t^r = (r-1)(r-2)t^r.

This vanishes for all $t>0$ iff $r=1$ or $r=2$ . So $y_1(t)=t$ and $y_2(t)=t^2$ are homogeneous solutions. Independence check (a preview of the Wronskian test): the Wronskian is

W(t) = y_1y_2' - y_1'y_2 = t\cdot 2t - 1\cdot t^2 = t^2,

which is nonzero for every $t>0$ ; hence $y_1,y_2$ are linearly independent on $(0,\infty)$ and, by the Structure Theorem, form a basis of the 2-dimensional homogeneous solution space. Particular solution. Try $y_p = At^3$ . Then $y_p' = 3At^2$ , $y_p''=6At$ , and

t^2(6At) - 2t(3At^2) + 2(At^3) = 6At^3 - 6At^3 + 2At^3 = 2At^3.

We need this to equal $t^3$ , so $A = \tfrac12$ . Thus $y_p(t) = \tfrac12 t^3$ solves the nonhomogeneous equation. General solution. By Part (4) of the Structure Theorem,

y(t) = \frac{1}{2}t^3 + c_1 t + c_2 t^2, \qquad t > 0.

Solving an IVP. Impose $y(1) = 2$ , $y'(1) = 1$ . Compute $y'(t) = \tfrac32 t^2 + c_1 + 2c_2 t$ . At $t=1$ :

y(1) = \tfrac12 + c_1 + c_2 = 2 \ \Rightarrow\ c_1+c_2 = \tfrac32,

y'(1) = \tfrac32 + c_1 + 2c_2 = 1 \ \Rightarrow\ c_1 + 2c_2 = -\tfrac12.

Subtracting, $c_2 = -\tfrac12 - \tfrac32 = -2$ , then $c_1 = \tfrac32 - c_2 = \tfrac32+2 = \tfrac72$ . So the unique solution of the IVP on $(0,\infty)$ is

y(t) = \frac{1}{2}t^3 + \frac{7}{2}t - 2t^2.

Uniqueness is guaranteed by Part (1) since $p,q,g$ are continuous on $(0,\infty) \ni t_0=1$ .

Example 2 — Superposition with a non-obvious particular solution and a forcing term split in two. Consider the constant-coefficient equation

y'' + y = 4\cos t + 6t, \qquad t \in \mathbb{R} = I.

Here $p(t)=0$ , $q(t)=1$ , $g(t) = 4\cos t + 6t$ , all continuous on all of $\mathbb{R}$ , so the Structure Theorem applies globally. Homogeneous solutions. $y_1(t) = \cos t$ , $y_2(t) = \sin t$ satisfy $y_i'' + y_i = 0$ (direct check: $(\cos t)''=-\cos t$ , so $-\cos t + \cos t = 0$ ; similarly for $\sin t$ ). Their Wronskian is $W = \cos t \cdot \cos t - (-\sin t)\sin t = \cos^2 t + \sin^2 t = 1 \ne 0$ , so they are independent. Splitting the forcing (superposition as a solving technique, not just a structural fact). Write $g = g_1+g_2$ with $g_1 = 4\cos t$ , $g_2 = 6t$ . Solve $L[y]=g_1$ and $L[y]=g_2$ separately, then add — this works because of linearity (Part 2 of the theorem), and is the mechanism behind the method of undetermined coefficients. For $g_2 = 6t$ : try $y_{p2} = At$ (no resonance with the homogeneous solutions, which are trigonometric, not polynomial). Then $y_{p2}''=0$ , so $L[y_{p2}] = 0 + At = At$ ; we need $At = 6t$ , so $A=6$ : $y_{p2}(t) = 6t$ . For $g_1 = 4\cos t$ : naive guess $y=A\cos t$ fails because $\cos t$ solves the homogeneous equation (resonance) — this is exactly the kind of trap the independence structure warns about. Instead try $y_{p1} = t(A\cos t + B\sin t)$ . Compute:

y_{p1}' = A\cos t + B\sin t + t(-A\sin t + B\cos t),

y_{p1}'' = -A\sin t + B\cos t + (-A\sin t+B\cos t) + t(-A\cos t - B\sin t) = 2(-A\sin t+B\cos t) - t(A\cos t+B\sin t).

Then

y_{p1}'' + y_{p1} = 2(-A\sin t + B\cos t) - t(A\cos t+B\sin t) + t(A\cos t+B\sin t) = -2A\sin t + 2B\cos t.

Set this equal to $4\cos t$ : matching coefficients, $-2A = 0 \Rightarrow A=0$ , and $2B=4 \Rightarrow B=2$ . So $y_{p1}(t) = 2t\sin t$ . Assemble by superposition. $y_p = y_{p1}+y_{p2} = 2t\sin t + 6t$ satisfies $L[y_p] = L[y_{p1}]+L[y_{p2}] = 4\cos t + 6t = g$ , exactly by Part (2) of the Structure Theorem. By Part (4), the general solution is

y(t) = 2t\sin t + 6t + c_1\cos t + c_2 \sin t.

Solving an IVP. Impose $y(0)=1$ , $y'(0)=0$ . At $t=0$ : $y(0) = 0 + 0 + c_1 + 0 = c_1 = 1$ . Differentiate:

y'(t) = 2\sin t + 2t\cos t + 6 - c_1\sin t + c_2\cos t.

At $t=0$ : $y'(0) = 0+0+6-0+c_2 = 6+c_2 = 0 \Rightarrow c_2=-6$ . So

y(t) = 2t\sin t + 6t + \cos t - 6\sin t

is the unique solution of the IVP, guaranteed by Part (1) since $p,q,g$ are continuous on all of $\mathbb{R} \ni t_0=0$ .

p

Scope, edge cases & common traps

Situation	What holds	What can fail
$p,q,g$ continuous on open interval $I$ , $t_0 \in I$	Unique solution on all of $I$ ; $\dim V = 2$	—
Leading coefficient $a(t)$ vanishes at some $t_0$ (Euler / singular equations)	Theorem applies away from zeros of $a$	Existence/uniqueness at $t_0$ itself can fail; solutions may blow up or fail to extend
$p,q$ continuous but $g$ merely piecewise continuous	Existence/uniqueness theory extends with minor technical care (solutions $C^1$ , not $C^2$ , at jumps)	Formula $y=y_p+c_1y_1+c_2y_2$ still organizes solutions on each smooth piece
Nonlinear equation (e.g. $y''+y^2=0$ )	Local existence–uniqueness may still hold (Picard–Lindelöf)	Superposition fails entirely: solution set is not a vector space or an affine plane
Two homogeneous solutions with Wronskian $\equiv 0$	They may still both solve the equation	They are not a basis; general solution formula in Part (4) is invalid unless independence holds

The mathematician behind it

Origins · 1 minCharles Émile Picard and how this was discovered

CÉ

Charles Émile Picard

1856–1941 · French analysis, existence theory for differential equations

Picard developed the method of successive approximations that underlies the existence-uniqueness theorem this article leans on, working in the broader context of late-19th-century rigorization of analysis. His name is permanently attached, alongside Ernst Lindelöf, to the theorem that guarantees the two-dimensional structure proved here is not an assumption but a consequence of a deeper existence principle.

History

- 1820s: Cauchy gives the first rigorous existence proofs for differential equations using majorant/dominant-function techniques.

- 1890: Picard introduces the method of successive approximations (Picard iteration) to prove existence and uniqueness constructively.

- 1894: Lindelöf sharpens and generalizes Picard's approach, giving the modern Lipschitz-condition formulation.

- Early 1900s: The vector-space viewpoint on linear ODEs — solution sets as kernels of linear operators — becomes standard, unifying the theory of

n

-th order linear equations.

- Mid-20th century: The structure theorem here is absorbed as a special case of general linear operator theory in functional analysis and control theory.

Why does the theorem require the interval

I

to be where

p,q,g

are simultaneously continuous?

The proof converts the scalar equation into a first-order vector system $\mathbf{x}'=F(t,\mathbf{x})$ where $F$ depends on $p, q, g$ all at once. Picard–Lindelöf needs $F$ continuous in $t$ and Lipschitz in $\mathbf{x}$ on the domain of interest; this requires all three coefficient functions to behave well together on the same interval. If even one of them is discontinuous or undefined at some point, $F$ fails to be continuous there, the existence-uniqueness argument breaks down at that point, and neither existence nor the two-dimensionality of the homogeneous solution space is guaranteed to extend across it, as the Euler-equation counterexample shows.

Does the general solution formula

y=y_p+c_1y_1+c_2y_2

depend on which particular solution

y_p

we pick?

No — the set of solutions described is identical no matter which particular solution is chosen, though the specific values of $c_1,c_2$ needed for a given IVP will change. If $y_p$ and $\tilde y_p$ are two different particular solutions, their difference $y_p - \tilde y_p$ solves the homogeneous equation (by superposition), hence already lies in $\mathrm{span}\{y_1,y_2\}$ . So shifting from $y_p$ to $\tilde y_p$ just relabels the constants $c_1,c_2$ ; the affine plane of solutions in function space stays exactly the same.

Is linear independence of

y_1,y_2

the same as them being "different functions"?

No, and this is a frequent source of errors. Two functions can be different (not scalar multiples of each other pointwise-looking) yet still be linearly dependent in the relevant sense only if one actually is a constant multiple of the other everywhere on $I$ . Conversely, subtler dependence is impossible for exactly two functions: $y_1,y_2$ are dependent iff one is a scalar multiple of the other. The clean test is the Wronskian: if $W(y_1,y_2)(t_0)\ne 0$ for some $t_0\in I$ , they are independent; a full theorem (covered separately) shows this single-point check suffices for homogeneous solutions, though not for arbitrary functions.

Why can't we just use Part (1) (existence-uniqueness) directly to solve equations, skipping the vector space structure?

Part (1) only tells us a solution to a specific IVP exists and is unique — it does not, by itself, give a formula. The vector space structure in Parts (2)–(4) is what turns this abstract guarantee into a concrete algorithm: find any two independent homogeneous solutions (often by inspired guessing or, for constant coefficients, via characteristic roots), find one particular solution, and combine them. Without the structure theorem, "solving" the ODE would mean re-deriving a Picard iteration for every new IVP, which is not computationally practical.

Does superposition work for nonlinear equations too?

No, and this is exactly why "linear" is in the name. For a nonlinear equation such as $y'' + y^2 = 0$ , if $y_1,y_2$ are both solutions, $L[y_1+y_2] = (y_1+y_2)'' + (y_1+y_2)^2 = y_1''+y_2'' + y_1^2+2y_1y_2+y_2^2$ , which is not $L[y_1]+L[y_2] = y_1''+y_1^2+y_2''+y_2^2$ unless the cross term $2y_1y_2$ vanishes. The proof of Part (2) used only that multiplication by $p(t), q(t)$ is linear in $y,y'$ ; any nonlinear dependence on $y$ or $y'$ destroys this and the entire vector-space picture (Parts 3–4) collapses with it.

How does this article connect to solving equations with constant coefficients?

The Structure Theorem tells us the shape of the answer — two independent homogeneous solutions plus one particular solution — but says nothing about how to find them for a specific $p,q$ . When $p,q$ are constants, the guess $y=e^{rt}$ reduces $L[y]=0$ to a quadratic (the characteristic equation) in $r$ , giving an explicit recipe for $y_1,y_2$ ; this is the content of Homogeneous Equations with Constant Coefficients. The independence of the resulting solutions, and why real, repeated, or complex roots each demand different bases, is checked using exactly the Wronskian machinery introduced informally in the worked examples here.

What happens at points where two candidate homogeneous solutions coincide, as with resonance in Example 2?

Resonance — when the forcing function $g$ "looks like" a homogeneous solution, such as $\cos t$ forcing $y''+y$ — does not violate the Structure Theorem; it only means the naive undetermined-coefficients guess for $y_p$ must be modified (multiplied by $t$ , as shown in Example 2) to avoid accidentally guessing a homogeneous solution, which would make $L[y_p]=0\ne g$ . The theorem itself, guaranteeing existence of some particular solution and the two-dimensional homogeneous space, remains fully valid; only the search strategy for $y_p$ needs adjustment.