2.1 General Theory of Linear Equations
Linear equations of order $n$: global existence and uniqueness, the $n$-dimensional solution space, Wronskians and Abel's formula, reduction of order, constant-coefficient and Euler–Cauchy equations, and the interlacing of zeros (Sturm's separation theorem).
A spring with damping, an electrical circuit, a beam under load, the radial part of a vibrating drum, the quantum harmonic oscillator — all are governed by linear differential equations $a_n(x)y^{(n)}+\dots+a_1(x)y'+a_0(x)y=r(x)$. Linearity is what makes them tractable: solutions can be added and scaled, so finitely many of them generate all the others. For nonlinear equations we had to ask, in Chapter 1, whether a solution even survives for long; for a linear equation with continuous coefficients the answer is always "yes, on the whole interval", and the real questions become algebraic and geometric: how many independent solutions are there, how can we tell whether given solutions are independent, how do we find them, and what do they look like?
This section answers these questions for the homogeneous equation $L[y]=0$. The solution space has dimension exactly $n$; the Wronskian detects independence and obeys Abel's formula, so it vanishes either nowhere or everywhere; one known solution of a second-order equation yields a second by reduction of order; constant-coefficient and Euler–Cauchy equations are solved completely through a polynomial; and the zeros of solutions of second-order equations interlace (Sturm). CSIR NET tests every one of these facts, usually in the form "can these functions be solutions of such an equation?", "find $W(2)$ given $W(1)$", "how many zeros can the other solution have?" — and the answer is always one of the theorems below applied to a well-chosen counterexample. We write $x$ for the independent variable and $D=d/dx$; the systems theory of §1.3 (where the variable was $t$) is used with $t$ renamed $x$.
Linear Equations of Order $n$: Existence, Uniqueness and Superposition
A linear differential equation of order $n$ on an interval $J$ is $$a_n(x)y^{(n)}+a_{n-1}(x)y^{(n-1)}+\dots+a_1(x)y'+a_0(x)y=r(x),$$ with coefficients $a_k$ and right-hand side $r$ defined on $J$. It is homogeneous if $r\equiv0$ and non-homogeneous otherwise. Where $a_n(x)\ne0$ we may divide by $a_n$ and obtain the normal form
$$L[y]:=y^{(n)}+p_{n-1}(x)y^{(n-1)}+\dots+p_1(x)y'+p_0(x)y=f(x).$$Throughout this section the standing hypothesis is: $p_0,\dots,p_{n-1}$ and $f$ are continuous on the open interval $J$. A point where $a_n$ vanishes (or where some $p_k$ fails to be continuous) is a singular point of the equation; the theory below applies on each interval free of singular points. A solution on $J$ is an $n$ times differentiable function satisfying the equation at every point of $J$. The map $y\mapsto L[y]$ is the linear differential operator of the equation.
Proof. Put $u=(y,y',\dots,y^{(n-1)})^T$. By the reduction theorem of §1.3, $y$ solves the IVP if and only if $u$ solves $u'=C(x)u+f(x)e_n$, $u(x_0)=(\eta_0,\dots,\eta_{n-1})^T$, where $C(x)$ is the companion matrix with last row $(-p_0,\dots,-p_{n-1})$ and $e_n=(0,\dots,0,1)^T$. The entries of $C$ and of $f e_n$ are continuous on $J$, so the global existence theorem for linear systems (§1.3, existence and uniqueness for systems, part (4)) gives exactly one solution $u$, defined on all of $J$. Its first component is the required $y$, and $y^{(n)}=f-\sum_kp_ky^{(k)}$ is continuous, so $y\in C^n(J)$. The last sentence is the case $f\equiv0$: both $y$ and the zero function solve the same IVP. ∎
Two features distinguish this theorem from the Picard–Lindelöf theorem of §1.1. The solution lives on the whole interval — linear equations never blow up while their coefficients stay continuous — and the interval is determined by the coefficients alone, not by the initial data. The guaranteed interval is therefore the largest open interval containing $x_0$ on which all the $p_k$ and $f$ of the normal form are continuous. (A solution may happen to extend across a singular point, but nothing guarantees it.)
- Linearity. $L[c_1y_1+c_2y_2]=c_1L[y_1]+c_2L[y_2]$ for all $n$ times differentiable $y_1,y_2$ and constants $c_1,c_2$.
- Dimension. The set $S$ of solutions of $L[y]=0$ on $J$ is a vector space of dimension exactly $n$. For each $x_0\in J$ the map $$\Lambda_{x_0}:S\to\mathbb{R}^n,\qquad y\mapsto\big(y(x_0),y'(x_0),\dots,y^{(n-1)}(x_0)\big)$$ is a linear isomorphism.
- Affine structure. If $y_p$ is one solution of $L[y]=f$, then the solutions of $L[y]=f$ are exactly the functions $y_p+y_h$ with $y_h\in S$.
Proof. (1) Differentiation is linear and $(c_1y_1+c_2y_2)^{(k)}$ is multiplied by the same function $p_k(x)$ as $y_1^{(k)}$ and $y_2^{(k)}$. (2) By (1), $S$ is closed under linear combinations, and $\Lambda_{x_0}$ is linear. It is injective by the last sentence of Theorem 2.1 (if $\Lambda_{x_0}y=0$ then $y\equiv0$) and surjective because every IVP has a solution. A linear bijection onto $\mathbb{R}^n$ shows $\dim S=n$. (3) If $L[y]=f$ then $L[y-y_p]=f-f=0$, so $y-y_p\in S$; conversely $L[y_p+y_h]=f+0=f$. ∎
Part (2) contains the most useful single remark of the chapter: a solution of a homogeneous linear equation of order $n$ cannot vanish together with its first $n-1$ derivatives at a point unless it is identically zero. For $n=2$: a nontrivial solution has no double zero, and two distinct solutions cannot have tangent graphs at a point.
Linear Independence, the Wronskian and Abel's Formula
Functions $f_1,\dots,f_m$ defined on an interval $I$ are linearly dependent on $I$ if there are constants $c_1,\dots,c_m$, not all zero, with $c_1f_1(x)+\dots+c_mf_m(x)=0$ for every $x\in I$; otherwise they are linearly independent on $I$. If $f_1,\dots,f_n$ are $n-1$ times differentiable on $I$, their Wronskian is
$$W(f_1,\dots,f_n)(x)=\det\begin{pmatrix}f_1&f_2&\cdots&f_n\\f_1'&f_2'&\cdots&f_n'\\\vdots&&&\vdots\\f_1^{(n-1)}&f_2^{(n-1)}&\cdots&f_n^{(n-1)}\end{pmatrix}(x).$$For two functions $W(f_1,f_2)=f_1f_2'-f_1'f_2$. Dependence depends on the interval: $x$ and $|x|$ are dependent on $(0,1)$ but independent on $(-1,1)$.
Proof. Let $\sum c_jf_j\equiv0$ on $I$ with $c=(c_1,\dots,c_n)^T\ne0$. Differentiating the identity $k$ times gives $\sum_jc_jf_j^{(k)}(x)=0$ for $k=0,\dots,n-1$ and every $x\in I$. So for every $x$ the Wronskian matrix at $x$ sends the nonzero vector $c$ to $0$; a square matrix with a nontrivial kernel has determinant $0$. For the counterexample: $f_2=x|x|$ is differentiable with $f_2'=2|x|$, so $W=x^2\cdot2|x|-2x\cdot x|x|=0$ for all $x$. If $c_1x^2+c_2x|x|\equiv0$ on $(-1,1)$, then $x=\tfrac12$ gives $c_1+c_2=0$ and $x=-\tfrac12$ gives $c_1-c_2=0$, so $c_1=c_2=0$. ∎
The counterexample works because the two functions are dependent on $(0,1)$ and on $(-1,0)$ with different constants. For solutions of one linear equation this cannot happen, and the Wronskian becomes a perfect test:
- (Abel's formula.) For all $x,x_0\in J$, $$W(x)=W(x_0)\exp\!\left(-\int_{x_0}^{x}p_{n-1}(s)\,ds\right).$$ For $y''+p(x)y'+q(x)y=0$ this reads $W(x)=W(x_0)\,e^{-\int_{x_0}^xp}$.
- (Test.) The following are equivalent: (a) $y_1,\dots,y_n$ are linearly independent on $J$; (b) $W(x_0)\ne0$ for some $x_0\in J$; (c) $W(x)\ne0$ for every $x\in J$.
Proof. (1) The determinant is a multilinear function of its rows, so by the product rule $W'$ is the sum of the $n$ determinants obtained by differentiating one row at a time. Differentiating row $k$ (the row of $(k-1)$-th derivatives) for $k\le n-1$ produces a copy of row $k+1$, giving a determinant with two equal rows, which is $0$. Only the last row survives: it becomes $\big(y_1^{(n)},\dots,y_n^{(n)}\big)$, and since each $y_j$ solves the equation, $$y_j^{(n)}=-p_{n-1}y_j^{(n-1)}-\dots-p_0y_j.$$ By linearity in the last row, the resulting determinant splits into $n$ terms; the term with $-p_ky_j^{(k)}$ ($k\le n-2$) contains the row of $k$-th derivatives twice and vanishes, and the term with $-p_{n-1}y_j^{(n-1)}$ is $-p_{n-1}W$. Hence $W'=-p_{n-1}(x)W$, a first-order linear equation whose solutions are exactly $W(x)=W(x_0)\exp\left(-\int_{x_0}^xp_{n-1}\right)$. (This is Liouville's formula of §1.3 for the companion system, whose trace is $-p_{n-1}$.) For $n=2$ the computation is one line: $W'=(y_1y_2'-y_1'y_2)'=y_1y_2''-y_1''y_2$, and substituting $y_i''=-py_i'-qy_i$ gives $$W'=y_1(-py_2'-qy_2)-(-py_1'-qy_1)y_2=-pW.$$
(2) (c)⇒(b) is trivial and (b)⇒(a) is Theorem 2.3. For (a)⇒(c), suppose $W(x_1)=0$ for some $x_1\in J$. Then the columns $\Lambda_{x_1}y_j$ of the Wronskian matrix at $x_1$ are dependent: $\sum c_j\Lambda_{x_1}y_j=0$ with $c\ne0$. The solution $y=\sum c_jy_j$ then satisfies $\Lambda_{x_1}y=0$, so $y\equiv0$ by Theorem 2.1, and the $y_j$ are dependent. The final sentence follows from (1): the exponential never vanishes. ∎
- Fundamental sets exist: for any $x_0\in J$, the solutions $y_j$ with $\Lambda_{x_0}y_j=e_j$ (the $j$-th standard basis vector) form one, with $W(x_0)=1$.
- If $y_1,\dots,y_n$ is a fundamental set, every solution is $\sum c_jy_j$ for unique constants; those of the IVP with data $\Lambda_{x_0}y=\eta$ solve the linear system $\sum_jc_j\Lambda_{x_0}y_j=\eta$, whose matrix is the Wronskian matrix at $x_0$.
- (The equation determined by two functions.) Let $y_1,y_2\in C^2(J)$ with $W(y_1,y_2)(x)\ne0$ for every $x\in J$. Then there is exactly one equation $y''+p(x)y'+q(x)y=0$ with $p,q$ continuous on $J$ having $y_1,y_2$ as solutions, namely $$\det\begin{pmatrix}y&y_1&y_2\\y'&y_1'&y_2'\\y''&y_1''&y_2''\end{pmatrix}=0,\qquad\text{i.e.}\qquad p=-\frac{W'}{W},\quad q=\frac{y_1'y_2''-y_1''y_2'}{W}.$$
Proof. (1) and (2) restate Theorem 2.2(2): $\Lambda_{x_0}$ is an isomorphism, so it carries bases to bases, and the coordinates of a solution are found by inverting the Wronskian matrix at $x_0$. (3) Expand the determinant along its first column, using $W'=(y_1y_2'-y_1'y_2)'=y_1y_2''-y_1''y_2$: $$y\,(y_1'y_2''-y_1''y_2')-y'\,W'+y''\,W.$$ Dividing by $W$ (which never vanishes) gives the stated $p$ and $q$, continuous because $y_1,y_2\in C^2$. Substituting $y=y_1$ or $y=y_2$ gives a determinant with two equal columns, so both are solutions, and they are a fundamental set since $W\ne0$. Uniqueness: if $y''+\tilde py'+\tilde qy=0$ is another such equation, subtracting gives $(p-\tilde p)y_i'+(q-\tilde q)y_i=0$ for $i=1,2$, a homogeneous linear system for $(p-\tilde p,q-\tilde q)$ at each $x$ whose determinant $y_1'y_2-y_1y_2'=-W$ is nonzero; so $\tilde p=p$, $\tilde q=q$. ∎
Reduction of Order
There is no general formula for the solutions of $y''+p(x)y'+q(x)y=0$. But Abel's formula tells us the Wronskian of a second solution in advance, and $W(y_1,y_2)=y_1^2\,(y_2/y_1)'$ wherever $y_1\ne0$; so one solution determines a second by a single integration.
Proof. Try $y=vy_1$ with $v\in C^2(I)$. Then $y'=v'y_1+vy_1'$ and $y''=v''y_1+2v'y_1'+vy_1''$, so $$y''+py'+qy=v\,(y_1''+py_1'+qy_1)+v''y_1+v'(2y_1'+py_1)=v''y_1+v'(2y_1'+py_1).$$ Thus $y$ is a solution iff $w=v'$ satisfies the first-order linear equation $w'=-\left(\tfrac{2y_1'}{y_1}+p\right)w$ on $I$, whose solutions are $w=C\,y_1^{-2}e^{-\int_{x_0}^xp}$ (integrating factor $y_1^2e^{\int p}$). Taking $C=1$ and $v=\int_{x_0}^xw$ gives $y_2$. Its Wronskian with $y_1$ is $$W(y_1,vy_1)=y_1(v'y_1+vy_1')-y_1'vy_1=v'y_1^2=e^{-\int_{x_0}^xp}\ne0,$$ so $y_1,y_2$ are independent (Theorem 2.4). For order $n$, the same substitution makes every term without a derivative of $v$ add up to $vL[y_1]=0$, leaving an equation in $v',\dots,v^{(n)}$, i.e. of order $n-1$ in $w$. ∎
Constant Coefficients and Euler–Cauchy Equations
When the coefficients are constants the operator is a polynomial in $D=d/dx$, and polynomial algebra solves the equation completely.
Proof. $D(e^{\lambda x}u)=e^{\lambda x}(u'+\lambda u)=e^{\lambda x}(D+\lambda)u$. Applying this $k$ times gives $D^k(e^{\lambda x}u)=e^{\lambda x}(D+\lambda)^ku$ by induction, and the rule follows by linearity. With $u=1$, $(D+\lambda)^k1=\lambda^k$. ∎
Proof. They are solutions. Write $P(\mu+\lambda_j)=\mu^{m_j}H_j(\mu)$ for a polynomial $H_j$ (the root $\lambda_j$ of $P$ becomes the root $0$, of the same multiplicity). By the shift rule, $$P(D)\big[x^ke^{\lambda_jx}\big]=e^{\lambda_jx}P(D+\lambda_j)x^k=e^{\lambda_jx}H_j(D)D^{m_j}x^k=0\qquad(k\lt m_j).$$
They are independent. We prove the lemma: if $Q_1,\dots,Q_s$ are polynomials and $\sum_jQ_j(x)e^{\lambda_jx}=0$ on an interval, with $\lambda_j$ distinct, then every $Q_j\equiv0$. Induct on $s$; for $s=1$ divide by $e^{\lambda_1x}\ne0$. Assume the lemma for $s-1$ exponentials and multiply the relation by $e^{-\lambda_sx}$: $$\sum_{j\lt s}Q_j(x)e^{\mu_jx}+Q_s(x)=0,\qquad\mu_j=\lambda_j-\lambda_s\ne0.$$ Differentiate $d+1$ times, where $d=\deg Q_s$ (if $Q_s\equiv0$ there is nothing to do). The polynomial $Q_s$ disappears, and $D(Qe^{\mu x})=(Q'+\mu Q)e^{\mu x}$ shows that each term becomes $\tilde Q_je^{\mu_jx}$ with $\deg\tilde Q_j=\deg Q_j$ (the leading coefficient is multiplied by $\mu_j^{d+1}\ne0$); in particular $\tilde Q_j\equiv0$ only if $Q_j\equiv0$. By the induction hypothesis $\tilde Q_j\equiv0$, so $Q_j\equiv0$ for $j\lt s$, and then $Q_s\equiv0$ from the relation itself. A linear relation among the functions $x^ke^{\lambda_jx}$ is exactly a relation $\sum Q_je^{\lambda_jx}=0$ with $\deg Q_j\lt m_j$, so all coefficients vanish. Having $n$ independent solutions in an $n$-dimensional space (Theorem 2.2), we have a basis.
Real form. If the $a_k$ are real, $\overline{P(\lambda)}=P(\bar\lambda)$, so $\bar\lambda_j$ is a root of the same multiplicity, and $P(D)$ maps real functions to real functions; hence the real and imaginary parts of a complex solution are solutions. With $u=x^ke^{(\alpha+i\beta)x}$ we have $\operatorname{Re}u=\tfrac12(u+\bar u)$, $\operatorname{Im}u=\tfrac1{2i}(u-\bar u)$; this invertible change of basis inside the span of $u,\bar u$ keeps a basis a basis. ∎
- Constant coefficients: factor $P(\lambda)$; each real root $\lambda$ of multiplicity $m$ contributes $e^{\lambda x},xe^{\lambda x},\dots,x^{m-1}e^{\lambda x}$; each pair $\alpha\pm i\beta$ of multiplicity $m$ contributes $x^ke^{\alpha x}\cos\beta x$, $x^ke^{\alpha x}\sin\beta x$, $k\lt m$. Count: you must have $n$ functions.
- Euler–Cauchy ($x^ky^{(k)}$ terms only): try $y=x^r$, or substitute $x=e^t$; $\log x$ plays the role of $x$ above.
- One solution known (guess $e^{ax}$, $x^r$, a polynomial): reduction of order (Theorem 2.6).
- Questions about behaviour (bounded? $\to0$? oscillatory?): read them off the roots — $\operatorname{Re}\lambda\lt0$ decays, $\operatorname{Re}\lambda=0$ with $m=1$ is bounded, $\operatorname{Re}\lambda=0$ with $m\ge2$ grows like $x^{m-1}$, non-real roots oscillate.
- Questions about $W$ or zeros without solving: Abel (Theorem 2.4) and Sturm (Theorems 2.11–2.13).
Proof. By the chain rule, $\dfrac{dY}{dt}=y'(e^t)e^t=xy'$, so $\theta=x\,D$ on functions of $x$. We show $x^{k+1}D^{k+1}=(\theta-k)\,x^kD^k$ for $k\ge0$: indeed $$\theta\big(x^ky^{(k)}\big)=x\big(kx^{k-1}y^{(k)}+x^ky^{(k+1)}\big)=k\,x^ky^{(k)}+x^{k+1}y^{(k+1)}.$$ Induction on $k$ then gives $x^kD^k=(\theta-k+1)\cdots(\theta-1)\theta$, and the factors commute because they are polynomials in $\theta$. Substituting, the equation becomes $Q(\theta)Y=0$, and Theorem 2.8 gives the fundamental set $t^ke^{rt}$ (or $t^ke^{\alpha t}\cos\beta t$, $t^ke^{\alpha t}\sin\beta t$); with $t=\log x$, $e^{rt}=x^r$ and $t^k=(\log x)^k$. The map $y\mapsto Y$ is a linear bijection between solutions on $(0,\infty)$ and solutions on $\mathbb{R}$, so bases correspond. For $x\lt0$ put $x=-s$: then $\dfrac{d^k}{dx^k}=(-1)^k\dfrac{d^k}{ds^k}$ and $x^k=(-1)^ks^k$, so each term $x^ky^{(k)}$ keeps its form and the equation in $s\gt0$ is the same Euler–Cauchy equation. ∎
Zeros of Solutions and the Sturm Separation Theorem
For second-order equations the zeros of solutions carry most of the qualitative information — whether solutions oscillate, how fast, and where boundary conditions can be met (§2.3). We work with $y''+p(x)y'+q(x)y=0$, $p,q$ continuous on $J$.
Proof. (a) If $y(x_0)=y'(x_0)=0$ then $y\equiv0$ by Theorem 2.1. (b) Suppose $[a,b]$ contains infinitely many zeros. By the Bolzano–Weierstrass theorem there are distinct zeros $x_k\to x^*\in[a,b]$. By continuity $y(x^*)=0$, and then $y'(x^*)=\lim_k\dfrac{y(x_k)-y(x^*)}{x_k-x^*}=0$, contradicting (a). (c) At a common zero or common critical point $x_0$, $W(x_0)=y_1(x_0)y_2'(x_0)-y_1'(x_0)y_2(x_0)=0$, so the solutions are dependent by Theorem 2.4(2). ∎
Proof. No common zero: Theorem 2.10(c). Let $a\lt b$ be consecutive zeros of $y_1$; replacing $y_1$ by $-y_1$ if necessary, $y_1\gt0$ on $(a,b)$. Then $y_1'(a)=\lim_{x\to a^+}\frac{y_1(x)}{x-a}\ge0$ and $y_1'(b)\le0$, and both are nonzero by Theorem 2.10(a): $y_1'(a)\gt0\gt y_1'(b)$. The Wronskian $W=y_1y_2'-y_1'y_2$ is continuous and never zero (Theorem 2.4), so it has one sign on $[a,b]$. At the zeros of $y_1$, $$W(a)=-y_1'(a)\,y_2(a),\qquad W(b)=-y_1'(b)\,y_2(b).$$ Since $W(a)$, $W(b)$ have the same sign while $y_1'(a)$, $y_1'(b)$ have opposite signs, $y_2(a)$ and $y_2(b)$ are nonzero and of opposite signs, and by the intermediate value theorem $y_2$ vanishes somewhere in $(a,b)$. If $y_2$ had two zeros $c\lt d$ in $(a,b)$, we may take them consecutive, and the argument with the roles of $y_1,y_2$ exchanged would give a zero of $y_1$ in $(c,d)\subset(a,b)$, contradicting the choice of $a,b$. The counting statement follows: if $y_1$ has zeros $a_1\lt\dots\lt a_m$ in $[\alpha,\beta]$, then $y_2$ has exactly one zero in each $(a_i,a_{i+1})$ and at most one in each of $[\alpha,a_1)$ and $(a_m,\beta]$. ∎
Sturm's theorem compares two solutions of the same equation. Comparing solutions of two different equations $y''+q_1y=0$ and $y''+q_2y=0$ with $q_2\ge q_1$ is the Sturm comparison theorem, proved in §2.3; it says that larger $q$ means faster oscillation. To apply either theorem we first remove the first-derivative term:
Proof. Let $m=\exp\left(-\tfrac12\int p\right)\gt0$; then $m'=-\tfrac12pm$ and $m''=\left(\tfrac14p^2-\tfrac12p'\right)m$. With $y=mu$: $y'=mu'+m'u$, $y''=mu''+2m'u'+m''u$, and $$y''+py'+qy=m\,u''+(2m'+pm)\,u'+(m''+pm'+qm)\,u.$$ Here $2m'+pm=0$ and $$m''+pm'+qm=m\left(\tfrac14p^2-\tfrac12p'-\tfrac12p^2+q\right)=mQ.$$ Divide by $m$. ∎
Proof. Suppose $u(a)=u(b)=0$ with $a\lt b$; by Theorem 2.10(b) we may take them consecutive, and replacing $u$ by $-u$, $u\gt0$ on $(a,b)$. Integrating by parts, $$\int_a^bu\,u''\,dx=\big[u\,u'\big]_a^b-\int_a^b(u')^2\,dx=-\int_a^b(u')^2\,dx\lt0,$$ the inequality being strict because $u'\equiv0$ on $(a,b)$ would make $u$ constant, hence $0$. But $u\,u''=-Qu^2\ge0$ on $[a,b]$, so the left side is $\ge0$. Contradiction. The second statement follows from Theorem 2.12, which preserves zeros. ∎
The sign of $q$ alone decides nothing: $y''+4y'+y=0$ has $q=1\gt0$ but roots $-2\pm\sqrt3$ and no oscillation, matching $Q=1-4=-3\lt0$; conversely, when $Q\ge m^2\gt0$ the comparison theorem of §2.3 shows that consecutive zeros are at most $\pi/m$ apart.
- $W\equiv0$ does not imply dependence for arbitrary functions: $x^2$ and $x|x|$ (or $x^3$ and $|x|^3$) on $(-1,1)$. It does for solutions of one normal linear equation (Theorem 2.4), and also for analytic functions.
- Independence depends on the interval. $x$ and $|x|$ are dependent on $(0,1)$, independent on $(-1,1)$.
- Abel's formula needs the normal form. For $a_2y''+a_1y'+a_0y=0$ use $p=a_1/a_2$, not $a_1$; and it says nothing across a zero of $a_2$: $x$ and $x^2$ solve $x^2y''-2xy'+2y=0$ with $W=x^2$, which vanishes at the singular point $0$ only.
- A Wronskian of solutions cannot have an isolated zero. So $\{x,e^x\}$ on $\mathbb{R}$, $\{x,\sin x\}$ on $(0,5)$, $\{1+x,e^x\}$ on $(-1,1)$ are not fundamental sets of any normal second-order equation with continuous coefficients there.
- Double zeros are forbidden only up to order $n-1$. $x^2$ cannot solve a normal second-order linear equation on an interval containing $0$, but it solves $y'''=0$; a solution of an order-$n$ equation cannot vanish with its first $n-1$ derivatives.
- Superposition is for homogeneous linear equations. The sum of two solutions of $L[y]=f$ solves $L[y]=2f$; the difference solves $L[y]=0$. For nonlinear equations ($y'=y^2$: $\tfrac{1}{1-x}$ and $0$ are solutions, $\tfrac{2}{1-x}$ is not) nothing survives.
- Repeated roots need powers of $x$; repeated indicial roots need powers of $\log x$. For $(\lambda-2)^2$: $e^{2x},xe^{2x}$; for $(r-2)^2$ in an Euler equation: $x^2,x^2\log x$ — not $x^3$.
- Zeros are isolated only away from singular points. $\sin(2\log x)$ solves $x^2y''+xy'+4y=0$ and has infinitely many zeros in $(0,1)$, accumulating at the singular point $0$.
- The sign of $q$ does not decide oscillation; $Q=q-\tfrac14p^2-\tfrac12p'$ does. $y''+4y'+y=0$ ($q\gt0$) never oscillates; $y''+2xy'+(x^2+2)y=0$ oscillates with zeros exactly $\pi$ apart although $q$ looks unrelated to that spacing.
- Sturm's theorem needs independence and one equation. Dependent solutions share all their zeros; solutions of different equations need the comparison theorem (§2.3). Interlacing alone is not sufficient for two functions to be solutions of one equation — see Example 6(d).
- Constant-coefficient operators commute; variable ones do not. $(D-1)(D-2)=(D-2)(D-1)$, but $(D-x)(D+x)y=y''-x^2y+y$ while $(D+x)(D-x)y=y''-x^2y-y$.
Exercises
- 1 Find the largest open interval on which Theorem 2.1 guarantees a unique solution of $(x^2-1)y''+(\cot x)\,y'+y=\log x$ with data given at $x_0=2$, and at $x_0=\tfrac12$.
- 2 Compute $W(e^x,e^{2x},e^{3x})$, and more generally show that $W(e^{\lambda_1x},\dots,e^{\lambda_nx})=V(\lambda_1,\dots,\lambda_n)\,e^{(\lambda_1+\dots+\lambda_n)x}$, where $V$ is the Vandermonde determinant. Deduce that exponentials with distinct exponents are independent.
- 3 Show that $W(x^2,x|x|)\equiv0$ on $\mathbb{R}$ although the functions are independent on $\mathbb{R}$; that they are dependent on $(0,\infty)$; and that they cannot both be solutions of $y''+py'+qy=0$ with $p,q$ continuous on an interval containing $0$.
- 4 Find the equation $y''+p(x)y'+q(x)y=0$ on $(0,\infty)$ having $x$ and $x^3$ as a fundamental set.
- 5 Let $y_1,y_2$ be solutions of $xy''+2y'+xe^xy=0$ on $(0,\infty)$ with $W(y_1,y_2)(1)=2$. Find $W(y_1,y_2)(x)$.
- 6 Let $\{y_1,y_2\}$ be a fundamental set of $y''+py'+qy=0$ on $J$ and $u=ay_1+by_2$, $v=cy_1+dy_2$. Prove $W(u,v)=(ad-bc)\,W(y_1,y_2)$ and decide when $\{u,v\}$ is a fundamental set.
- 7 Let $y$ solve $y'''+a_2(x)y''+a_1(x)y'+a_0(x)y=0$ (continuous coefficients on $J$) with $y(x_0)=y'(x_0)=y''(x_0)=0$. Show $y\equiv0$. Give a third-order equation with continuous coefficients having the nonzero solution $x^2$, which vanishes doubly at $0$.
- 8 Let $y_1,y_2$ solve $L[y]=f$ with $f\not\equiv0$. True or false: (a) $y_1+y_2$ is a solution; (b) $y_1-y_2$ solves $L[y]=0$; (c) $\tfrac12(y_1+y_2)$ is a solution; (d) $3y_1-2y_2$ is a solution.
- 1 $y_1=e^x$ solves $xy''-(x+1)y'+y=0$ on $(0,\infty)$. Find a second solution by reduction of order.
- 2 $y_1=x^{-1/2}\cos x$ solves $x^2y''+xy'+\left(x^2-\tfrac14\right)y=0$ on $(0,\infty)$. Find a second solution.
- 3 Solve $y''-4y'+4y=0$, $y(0)=1$, $y'(0)=1$.
- 4 Find the general real solution of $y'''+y=0$.
- 5 Find the general solution of $y^{(4)}-16y=0$, and all solutions bounded on $\mathbb{R}$.
- 6 Find a homogeneous linear equation with constant real coefficients, of least possible order, having $xe^{-x}\cos2x$ as a solution.
- 7 For which real $k$ do all solutions of $y''+ky'+9y=0$ tend to $0$ as $x\to\infty$? For which are all solutions bounded on $[0,\infty)$? For which do nontrivial solutions have infinitely many zeros?
- 8 Show that if $y''+py'+qy=0$ has constant $p,q$ and a solution $y_1=e^{\lambda x}$ with $\lambda$ a double root, reduction of order gives $y_2=xe^{\lambda x}$ — an independent derivation of the repeated-root rule for $n=2$.
- 1 Solve $x^2y''+2xy'-6y=0$ on $(0,\infty)$.
- 2 Solve $x^2y''-xy'+y=0$, $y(1)=1$, $y'(1)=0$.
- 3 Find a real fundamental set of $x^2y''+3xy'+5y=0$ on $(0,\infty)$.
- 4 Solve the third-order equation $x^3y'''+xy'-y=0$ on $(0,\infty)$.
- 5 Solve $(x-1)^2y''-3(x-1)y'+4y=0$ on $(1,\infty)$.
- 6 For which real $b$ does every nontrivial solution of $x^2y''+by=0$ have infinitely many zeros in $(0,\infty)$? Show that for $b\le\tfrac14$ every nontrivial solution has at most one zero there.
- 1 Show that no nontrivial solution of $y''-(1+x^2)y=0$ has two zeros, and deduce that the boundary value problem $y''-(1+x^2)y=0$, $y(0)=y(1)=0$ has only the trivial solution.
- 2 Transform Bessel's equation $x^2y''+xy'+(x^2-\nu^2)y=0$ into normal form and show $Q(x)=1+\dfrac{1-4\nu^2}{4x^2}$.
- 3 Let $q\gt0$ be continuous on $J$ and $y\not\equiv0$ solve $y''+q(x)y=0$. Prove that the zeros of $y$ and of $y'$ strictly interlace.
- 4 Can $x^2-1$ and $x$ form a fundamental set on $\mathbb{R}$ of some $y''+p(x)y'+q(x)y=0$ with continuous $p,q$? If so, find the equation.
- 5 Prove that two linearly independent solutions of $y''+py'+qy=0$ cannot attain a local maximum at the same point.
- 6 Let $y_1,y_2$ be independent solutions of $y''+py'+qy=0$ on $\mathbb{R}$, and suppose $y_1$ has exactly $7$ zeros in $[0,10]$. What are the possible numbers of zeros of $y_2$ in $[0,10]$?
- 7 Let $p,q$ be continuous on $J$ with $q\lt0$. Prove that a nontrivial solution of $y''+py'+qy=0$ has no positive local maximum and no negative local minimum, and hence at most one zero in $J$.
- 8 Show that $W(y_1,y_2)=y_1^2\left(\dfrac{y_2}{y_1}\right)'$ where $y_1\ne0$, and use the quotient $y_1/y_2$ and Rolle's theorem to give a second proof of the existence part of Sturm's separation theorem.
E1. CSIR Part B The largest open interval on which the initial value problem $(x^2-4)\,y''+(\sec x)\,y'+xy=\dfrac{1}{x-3}$, $y(1)=0$, $y'(1)=1$, is guaranteed to have a unique solution is (A) $(-2,2)$; (B) $\left(-\tfrac{\pi}{2},\tfrac{\pi}{2}\right)$; (C) $\left(-\tfrac{\pi}{2},2\right)$; (D) $\left(0,\tfrac{\pi}{2}\right)$.
E2. CSIR Part B GATE MA Let $y_1,y_2$ be solutions of $x^2y''-x(x+2)y'+(x+2)y=0$ on $(0,\infty)$ with $W(y_1,y_2)(1)=e$. Then $W(y_1,y_2)(2)$ equals (A) $4e^2$; (B) $2e^2$; (C) $\tfrac14e^2$; (D) $4e$.
E3. CSIR Part B GATE MA If $y$ solves $x^2y''-3xy'+4y=0$ on $(0,\infty)$ with $y(1)=0$, $y'(1)=2$, then $y(e)$ equals (A) $e^2$; (B) $2e^2$; (C) $2e$; (D) $4e^2$.
E4. CSIR Part C Let $p,q$ be continuous on $\mathbb{R}$ and let $y_1,y_2$ be solutions of $y''+p(x)y'+q(x)y=0$. Which of the following statements are true? (One or more options may be correct.) (A) If $y_1(0)=y_2(0)=0$, then $y_1,y_2$ are linearly dependent. (B) If $y_1,y_2$ are linearly independent, then between two consecutive zeros of $y_2$ there is exactly one zero of $y_1$. (C) The function $\sin^2x$ can be a solution of such an equation. (D) If $W(y_1,y_2)(0)=1$, then $W(y_1,y_2)(x)\le1$ for all $x\ge0$.
E5. CSIR Part C Consider $y^{(4)}+2y''+y=0$ on $\mathbb{R}$. Which of the following are true? (One or more options may be correct.) (A) Every solution is bounded on $\mathbb{R}$. (B) Every solution is periodic. (C) Every nontrivial solution has infinitely many zeros in $(0,\infty)$. (D) The solutions that are bounded on $\mathbb{R}$ form a two-dimensional subspace of the solution space.
E6. CSIR Part C Which of the following pairs can be a fundamental set of solutions on $(-1,1)$ of an equation $y''+p(x)y'+q(x)y=0$ with $p,q$ continuous on $(-1,1)$? (One or more options may be correct.) (A) $\{x,x^2\}$; (B) $\{1+x,e^x\}$; (C) $\{\cos x,\ x+\sin x\}$; (D) $\{e^x,x\}$.
Suggested plan — 3 lectures × 60 min.
- Structure. Theorem 2.1 from the companion system of §1.3 (ten minutes; the class has seen it); Example 1 on the board — students always forget the coefficient $\tan x$ and the fact that the interval is fixed by the equation, not the data. Theorem 2.2 and its corollary "no double zeros". Wronskians: Theorem 2.3 with the $x^2$, $x|x|$ counterexample drawn, then Theorem 2.4 with the full proof of Abel's formula for $n=2$ and the row-differentiation argument for general $n$. Example 2.
- Finding solutions. Reduction of order (Theorem 2.6, Example 3); the shift rule and Theorem 2.8 — do the independence lemma carefully, it is a good induction; project Interactive 1 and sweep $b$ through $2$ to show the double root; Example 4; Euler–Cauchy via $x=e^t$ (Theorem 2.9), Example 5.
- Zeros. Theorem 2.10; Sturm's separation theorem with the Wronskian-sign picture; Interactive 2 (start with $y''+y=0$, then Airy and Bessel — ask the class to predict before pressing Play whether zeros get closer or further apart). Normal form, Theorem 2.13, Examples 6 and 7. Finish with E4 and E6 live.
Discussion questions. (1) Why can a Wronskian of arbitrary functions vanish at isolated points, but the Wronskian of solutions cannot? What extra information do solutions carry? (2) Is every pair of $C^2$ functions with nonvanishing Wronskian a fundamental set of some equation? (Yes, on an interval — Theorem 2.5(3).) And every single nonzero function without double zeros? (3) The Euler equation and the constant-coefficient equation are "the same" equation in different variables. What property of the zeros is preserved by $x=e^t$, and what is distorted?
Misconceptions to watch for. Applying Abel's formula without dividing by the leading coefficient; treating $W(x_0)=0$ at one point as proof of dependence for arbitrary functions; writing $x^3$ instead of $x^2\log x$ for a double indicial root; believing $q\gt0$ implies oscillation; assuming that interlacing zeros suffice for two functions to solve one equation; forgetting that the interval of existence is determined by the normal form.
Suggested homework. I(1, 4, 5, 7), II(1, 2, 6, 7), III(2, 4, 6), IV(1, 3, 4, 6, 8); Exam corner E2, E5.
Research question. Why does a drum not sound like a guitar string? The vibration frequencies of a circular membrane are proportional to the positive zeros $j_{0,1}\lt j_{0,2}\lt\cdots$ of the Bessel function $J_0$; how are these zeros spaced, and how well does the theory of this section predict them?
Why it matters. A string's overtones are integer multiples of the fundamental, which our ears hear as harmony; a drum's are not, which is why timpani makers in Europe and tabla makers in India load or shape the membrane to push its overtones towards harmonic ratios. The same zeros fix the resonant frequencies of circular waveguides in microwave engineering and the vibration modes of loudspeaker cones.
What to do.
- Compute the first eight positive zeros of $J_0$ (in Python:
scipy.special.jn_zeros(0, 8)) and their differences. Use the normal form $u''+\left(1+\frac{1}{4x^2}\right)u=0$ (Exercise IV.2 with $\nu=0$) to explain why the differences approach $\pi$ — and why from below. - Compare with McMahon's approximation $j_{0,k}\approx\beta_k+\dfrac{1}{8\beta_k}$, $\beta_k=\left(k-\tfrac14\right)\pi$; tabulate the error.
- Compute the first zeros of the second solution $Y_0$ (
scipy.special.yn_zeros(0, 4)) and check Sturm's interlacing with the zeros of $J_0$. - Compute the frequency ratios $j_{0,2}/j_{0,1}$ and $j_{0,3}/j_{0,1}$ for the symmetric modes of a drum and compare with the ratios $2$ and $3$ of a string.
Tools: Python with SciPy, or any computer algebra system; Interactive 2 (Bessel mode) for a visual check.
Hand in: A one-page note with the table of zeros, differences and McMahon errors, the interlacing check, and a two-sentence answer to the research question.
Going further. The non-symmetric modes involve $J_m$, $m\ge1$. Check numerically that the zeros of $J_0$ and $J_1$ interlace, and read about how the zeros of $J_m$ for several $m$ combine into the full spectrum of the drum ("Can one hear the shape of a drum?").
🧪 Self-check quiz
Q1. The functions $x^2$ and $x|x|$ on $(-1,1)$:
Q2. If $y_1,y_2$ solve $y''+3y'+(\sin x)\,y=0$ and $W(y_1,y_2)(0)=2$, then $W(y_1,y_2)(x)=$
Q3. A fundamental set of $y'''-y''=0$ is
Q4. A fundamental set of $x^2y''+xy'-y=0$ on $(0,\infty)$ is
Q5. Let $y_1,y_2$ be linearly independent solutions of $y''+p(x)y'+q(x)y=0$ ($p,q$ continuous on $\mathbb{R}$), and let $1$ and $3$ be consecutive zeros of $y_1$. Then on $(1,3)$, $y_2$ has