4.1 Classification, Canonical Forms and Equations with Constant Coefficients
Second-order linear equations in two variables: the discriminant $B^2-4AC$ and the hyperbolic, parabolic and elliptic types, characteristic curves, reduction to canonical form, and the general solution of equations $F(D,D')z=f$ with constant coefficients.
The three classical equations of mathematical physics — the wave equation $u_{tt}=c^2u_{xx}$, the heat equation $u_t=k\,u_{xx}$ and Laplace's equation $u_{xx}+u_{yy}=0$ — look almost alike, yet their solutions could hardly be more different. Waves travel at finite speed and keep their corners; heat spreads instantly and smooths every corner; harmonic functions have no "time" at all and are fixed by their boundary values. Which of these behaviours a given equation shows is decided by a single number computed from its second-order coefficients, the discriminant $B^2-4AC$, exactly as the discriminant of $Ax^2+Bxy+Cy^2$ decides whether a conic is a hyperbola, a parabola or an ellipse. The names are borrowed from the conics for precisely this reason. An equation with variable coefficients can be of different types in different parts of the plane: the Tricomi equation $y\,u_{xx}+u_{yy}=0$, which models air flow around a wing near the speed of sound, is elliptic where the flow is subsonic and hyperbolic where it is supersonic.
This section develops the classification with complete proofs: the type is unchanged by every smooth invertible change of variables; the characteristic curves — two families, one or none — are exactly the curves along which the second derivatives of a solution are not determined by its values and slopes, so they carry the "signals" of hyperbolic equations; and choosing characteristics as new coordinates brings every equation to one of three canonical forms. The second half solves linear equations with constant coefficients, $F(D,D')z=f(x,y)$ with $D=\partial/\partial x$ and $D'=\partial/\partial y$, by factoring the operator — the direct analogue of the complementary function and particular integral of the ODE theory of §2.2. In CSIR NET this material is examined constantly: "in which region is this equation hyperbolic?", "which substitution reduces it to canonical form?", "what is the general solution of $u_{xx}-5u_{xy}+6u_{yy}=0$?", "find a particular integral when $F(a,b)=0$", and Part C statements about invariance of type and the existence of general solutions. Throughout, $u=u(x,y)$ (or $z=z(x,y)$ in the operator notation of the last part) is a real function of class $C^2$ on a domain $\Omega\subseteq\mathbb{R}^2$, and subscripts denote partial derivatives.
Second-Order Equations and Their Type
In Chapter 3 the slopes of a solution surface were called $p=z_x$, $q=z_y$; for second derivatives the traditional (Monge) notation is $r=z_{xx}$, $s=z_{xy}$, $t=z_{yy}$. We shall mostly write the subscripts explicitly.
A second-order partial differential equation in two independent variables is a relation $F(x,y,u,u_x,u_y,u_{xx},u_{xy},u_{yy})=0$ that genuinely involves at least one second derivative. It is
- linear if it has the form $$A u_{xx}+B u_{xy}+C u_{yy}+a_1u_x+a_2u_y+a_0u=g,$$ where $A,B,C,a_1,a_2,a_0,g$ are functions of $(x,y)$ only and $A,B,C$ do not vanish simultaneously; it is homogeneous if $g\equiv0$;
- semilinear if it has the form $Au_{xx}+Bu_{xy}+Cu_{yy}=\Phi(x,y,u,u_x,u_y)$ with $A,B,C$ functions of $(x,y)$ only and $\Phi$ arbitrary;
- quasi-linear if it is linear in the second derivatives, $Au_{xx}+Bu_{xy}+Cu_{yy}=\Phi$, but $A,B,C$ may depend on $u,u_x,u_y$ as well.
The terms $Au_{xx}+Bu_{xy}+Cu_{yy}$ form the principal part of the equation. Every linear equation is semilinear and every semilinear equation is quasi-linear.
For example, $u_{xx}+u_{yy}=u^3$ is semilinear but not linear; the minimal surface equation $(1+u_y^2)u_{xx}-2u_xu_yu_{xy}+(1+u_x^2)u_{yy}=0$ is quasi-linear but not semilinear; and $u_{xx}u_{yy}-u_{xy}^2=1$ (the Monge–Ampère equation) is not even quasi-linear. As for first-order equations, the classification below depends only on the principal part. It helps to write the principal part with the symmetric matrix $$M=\begin{pmatrix}A&B/2\\ B/2&C\end{pmatrix},\qquad Au_{xx}+Bu_{xy}+Cu_{yy}=\operatorname{tr}\!\big(M\,\nabla^2u\big),$$ where $\nabla^2u$ is the Hessian matrix of $u$; note that $B^2-4AC=-4\det M$.
Let the principal part $Au_{xx}+Bu_{xy}+Cu_{yy}$ have coefficients that do not all vanish at the point $(x_0,y_0)$, and let $\delta=B^2-4AC$ be its discriminant, evaluated at $(x_0,y_0)$. The equation is
It is hyperbolic (parabolic, elliptic) in a set $S$ if it is so at every point of $S$. For a quasi-linear equation the discriminant depends on the solution and its slopes, so the type is defined along a given solution.
Two conventions are in use. Many books write the principal part as $Au_{xx}+2Bu_{xy}+Cu_{yy}$ and then test the sign of $B^2-AC$; the two tests agree once the factor $2$ is accounted for. In this course the middle coefficient is the full coefficient of $u_{xy}$, and the test is $B^2-4AC$. For the heat equation in the variables $(x,t)$, $u_t=ku_{xx}$ has $A=k$, $B=C=0$: it is parabolic although it is "first order in $t$".
Let $\xi=\xi(x,y)$, $\eta=\eta(x,y)$ be of class $C^2$ near $(x_0,y_0)$ with Jacobian $J=\xi_x\eta_y-\xi_y\eta_x\ne0$ there, so that $(x,y)\mapsto(\xi,\eta)$ is a $C^2$ diffeomorphism of a neighbourhood (inverse function theorem). Write $u(x,y)=w(\xi(x,y),\eta(x,y))$. Then $$Au_{xx}+Bu_{xy}+Cu_{yy}+a_1u_x+a_2u_y+a_0u=\bar A w_{\xi\xi}+\bar B w_{\xi\eta}+\bar C w_{\eta\eta}+\bar a_1w_\xi+\bar a_2w_\eta+a_0w,$$ where $$\bar A=A\xi_x^2+B\xi_x\xi_y+C\xi_y^2,\qquad \bar C=A\eta_x^2+B\eta_x\eta_y+C\eta_y^2,$$ $$\bar B=2A\xi_x\eta_x+B(\xi_x\eta_y+\xi_y\eta_x)+2C\xi_y\eta_y,$$ $$\bar a_1=A\xi_{xx}+B\xi_{xy}+C\xi_{yy}+a_1\xi_x+a_2\xi_y,\qquad \bar a_2=A\eta_{xx}+B\eta_{xy}+C\eta_{yy}+a_1\eta_x+a_2\eta_y .$$ In matrix form $\bar M=P\,M\,P^{\mathsf T}$ with $P=\begin{pmatrix}\xi_x&\xi_y\\ \eta_x&\eta_y\end{pmatrix}$, and consequently $$\bar B^2-4\bar A\bar C=J^2\,\big(B^2-4AC\big).$$ In particular the type of the equation at a point does not depend on the coordinates, a linear (semilinear, quasi-linear) equation stays linear (semilinear, quasi-linear), and the coefficient $a_0$ of $u$ is unchanged.
Proof. By the chain rule $u_x=w_\xi\xi_x+w_\eta\eta_x$ and $u_y=w_\xi\xi_y+w_\eta\eta_y$. Differentiating once more (and using $w_{\xi\eta}=w_{\eta\xi}$, valid for $w\in C^2$), $$u_{xx}=w_{\xi\xi}\xi_x^2+2w_{\xi\eta}\xi_x\eta_x+w_{\eta\eta}\eta_x^2+w_\xi\xi_{xx}+w_\eta\eta_{xx},$$ $$u_{xy}=w_{\xi\xi}\xi_x\xi_y+w_{\xi\eta}(\xi_x\eta_y+\xi_y\eta_x)+w_{\eta\eta}\eta_x\eta_y+w_\xi\xi_{xy}+w_\eta\eta_{xy},$$ $$u_{yy}=w_{\xi\xi}\xi_y^2+2w_{\xi\eta}\xi_y\eta_y+w_{\eta\eta}\eta_y^2+w_\xi\xi_{yy}+w_\eta\eta_{yy}.$$ Multiply by $A$, $B$, $C$, add $a_1u_x+a_2u_y+a_0u$, and collect the coefficients of $w_{\xi\xi}$, $w_{\xi\eta}$, $w_{\eta\eta}$, $w_\xi$, $w_\eta$, $w$: this gives the five formulas, and $w$ keeps the coefficient $a_0$. The second-order coefficients are quadratic expressions in the rows $\mathbf p=(\xi_x,\xi_y)$ and $\mathbf q=(\eta_x,\eta_y)$ of $P$: $\bar A=\mathbf pM\mathbf p^{\mathsf T}$, $\tfrac12\bar B=\mathbf pM\mathbf q^{\mathsf T}$, $\bar C=\mathbf qM\mathbf q^{\mathsf T}$ (expand $\mathbf pM\mathbf q^{\mathsf T}=A\xi_x\eta_x+\tfrac B2(\xi_x\eta_y+\xi_y\eta_x)+C\xi_y\eta_y$). These are exactly the entries of $PMP^{\mathsf T}$, so $\bar M=PMP^{\mathsf T}$. Taking determinants, $\det\bar M=(\det P)^2\det M=J^2\det M$, and multiplying by $-4$ gives $\bar B^2-4\bar A\bar C=J^2(B^2-4AC)$. Since $J^2\gt0$, the two discriminants have the same sign. Finally, $\bar A,\bar B,\bar C$ are built from $A,B,C$ and the derivatives of $\xi,\eta$ only, so no dependence on $u$ is introduced; and $\bar M\ne0$ because $M\ne0$ and $P$ is invertible. ∎
Three consequences are used constantly. (i) Multiplying the equation by a nonvanishing function $h$ multiplies the discriminant by $h^2\gt0$, so the type is unchanged (even when $h\lt0$). (ii) Lower-order terms never affect the type: $u_{xx}-u_{yy}+100u_x+u^5=0$ is hyperbolic. (iii) If $A,B,C$ are continuous, the sets $\{B^2-4AC\gt0\}$ and $\{B^2-4AC\lt0\}$ are open (preimages of open intervals under a continuous function), so an equation that is hyperbolic or elliptic at a point is so in a whole neighbourhood; the parabolic set, by contrast, is closed and can be a curve, a point, a region or empty.
More variables. For $\sum_{i,j=1}^n a_{ij}u_{x_ix_j}+\dots=0$ with $a_{ij}=a_{ji}$, the type at a point is read from the eigenvalues of the symmetric matrix $(a_{ij})$: elliptic if they are all nonzero and of one sign, hyperbolic (normally hyperbolic) if they are all nonzero and exactly one has the opposite sign to the others, parabolic if exactly one eigenvalue is zero and the others have one sign, and ultrahyperbolic if at least two are of each sign. A linear change of variables replaces $(a_{ij})$ by $P(a_{ij})P^{\mathsf T}$, and by Sylvester's law of inertia the numbers of positive, negative and zero eigenvalues do not change. Remember to split the coefficient of a mixed term equally: $2u_{xy}$ contributes $1$ to the $(1,2)$ and to the $(2,1)$ entry.
Characteristic Curves
In Chapter 3 the characteristics of a first-order equation were the curves along which the equation reduces to an ODE. For second-order equations they have an equally concrete meaning: they are the curves on which knowing $u$ and its first derivatives does not determine the second derivatives (Theorem 4.3), and therefore the curves across which a solution may have a "kink" in its second derivatives — the wavefronts of a hyperbolic equation.
At a point $(x,y)$, a nonzero vector $(dx,dy)$ is a characteristic direction of the principal part $Au_{xx}+Bu_{xy}+Cu_{yy}$ if $$A\,dy^2-B\,dx\,dy+C\,dx^2=0 .$$ A smooth curve is a characteristic curve (a characteristic) if its tangent is a characteristic direction at every one of its points. A curve that is nowhere tangent to a characteristic direction is non-characteristic.
- A level curve $\varphi(x,y)=c$ of a $C^1$ function with $\nabla\varphi\ne0$ is a characteristic iff, along it, $$A\varphi_x^2+B\varphi_x\varphi_y+C\varphi_y^2=0 .$$ If $\varphi$ satisfies this first-order equation in a region, all its level curves there are characteristics.
- A graph $y=y(x)$ is a characteristic iff $$A\left(\frac{dy}{dx}\right)^2-B\,\frac{dy}{dx}+C=0$$ (the characteristic equation).
- At a point where the equation is hyperbolic there are exactly two characteristic directions, where it is parabolic exactly one, and where it is elliptic none. If $A\ne0$ the slopes are $\dfrac{dy}{dx}=\lambda_\pm=\dfrac{B\pm\sqrt{B^2-4AC}}{2A}$.
Proof. (1) Along a level curve the tangent is perpendicular to $\nabla\varphi$, so it is parallel to $(dx,dy)=(\varphi_y,-\varphi_x)$. Substituting, $A\,dy^2-B\,dx\,dy+C\,dx^2=A\varphi_x^2+B\varphi_y\varphi_x+C\varphi_y^2$, which proves (1). (2) The tangent of a graph is $(1,y')$, and $A\,y'^2-B\,y'+C=0$ is the condition of Definition 4.3. (3) Directions are determined up to scaling. If $A\ne0$, a characteristic direction cannot have $dx=0$ (that would force $A\,dy^2=0$), so we may take $dx=1$, and (2) is a quadratic in $y'$ with discriminant $B^2-4AC$: two real roots, a double root, or none. If $A=0$ then $A\,dy^2-B\,dx\,dy+C\,dx^2=dx\,(C\,dx-B\,dy)$; the directions are $dx=0$ and $C\,dx=B\,dy$, which are two distinct directions when $B\ne0$ (hyperbolic, $B^2-4AC=B^2\gt0$) and coincide when $B=0$, $C\ne0$ (parabolic). ∎
Beware of the minus sign in the characteristic equation: from $u_{xx}+2u_{xy}-3u_{yy}=0$ we get $y'^2-2y'-3=0$, i.e. $y'=3$ or $y'=-1$, not $y'=-3$ or $1$.
Let $u\in C^2$ solve the semilinear equation $Au_{xx}+Bu_{xy}+Cu_{yy}=\Phi(x,y,u,u_x,u_y)$ near a smooth curve $\Gamma:(x(\sigma),y(\sigma))$ with $(x',y')\ne(0,0)$, and suppose the Cauchy data $u$, $u_x$, $u_y$ are known along $\Gamma$. Then at a point of $\Gamma$ the second derivatives $u_{xx},u_{xy},u_{yy}$ are uniquely determined by the Cauchy data if and only if $\Gamma$ is non-characteristic there, i.e. $$A\,y'^2-B\,x'y'+C\,x'^2\ne0 .$$ Consequently, if $u$ is $C^1$ across a curve and $C^2$ up to it from each side, and solves the equation on both sides, then a jump of the second derivatives across the curve can occur only where the curve is characteristic.
Proof. Write $P(\sigma)=u_x(x(\sigma),y(\sigma))$ and $Q(\sigma)=u_y(x(\sigma),y(\sigma))$; these, and $\Phi$ on $\Gamma$, are known. Differentiating $P$ and $Q$ along $\Gamma$ by the chain rule and adding the equation itself gives three linear equations for $(u_{xx},u_{xy},u_{yy})$: $$\begin{aligned}A\,u_{xx}+B\,u_{xy}+C\,u_{yy}&=\Phi,\\ x'\,u_{xx}+y'\,u_{xy}\phantom{{}+C\,u_{yy}}&=P',\\ x'\,u_{xy}+y'\,u_{yy}&=Q'.\end{aligned}$$ Its determinant is $$\det\begin{pmatrix}A&B&C\\ x'&y'&0\\ 0&x'&y'\end{pmatrix}=A\,y'^2-B\,x'y'+C\,x'^2 .$$ If this is nonzero the system has exactly one solution, so the second derivatives are determined. If it is zero the matrix is singular: either the system has no solution (the Cauchy data are incompatible with the equation) or its solutions form a line or plane, so two solutions with the same Cauchy data can have different second derivatives on $\Gamma$. For the last statement, the one-sided limits of $(u_{xx},u_{xy},u_{yy})$ on the two sides both solve this system with the same right-hand side (since $u$, $u_x$, $u_y$ are continuous across the curve); at a non-characteristic point the solution is unique, so the limits agree and there is no jump. ∎
This is the second-order version of the transversality condition of §3.2. When the data curve is non-characteristic and everything is analytic, differentiating again determines all higher derivatives on $\Gamma$, and the Cauchy–Kovalevskaya theorem produces a local analytic solution. For the wave equation $u_{tt}=u_{xx}$ the characteristics are the lines $x\pm t=\text{const}$; the initial line $t=0$ is non-characteristic, which is why "initial position and initial velocity" is the right data. For the heat equation $u_t=u_{xx}$ the characteristic equation in the $(x,t)$-plane is $dt^2=0$: the lines $t=\text{const}$ are characteristic, and indeed only $u(x,0)$ — not $u_t(x,0)$ — may be prescribed. An elliptic equation has no real characteristics at all, so its solutions cannot have kinks: this is the first hint of the smoothness of harmonic functions (§4.2).
Reduction to Canonical Form
Theorem 4.1 tells us how the principal part changes under a change of variables; Theorem 4.2 tells us how to make a coefficient vanish: $\bar A=A\xi_x^2+B\xi_x\xi_y+C\xi_y^2$ is zero exactly when the level curves of $\xi$ are characteristics. Choosing the new coordinate curves to be characteristics therefore simplifies the principal part as much as possible.
Consider the semilinear equation $Au_{xx}+Bu_{xy}+Cu_{yy}=\Phi(x,y,u,u_x,u_y)$ with $A,B,C$ of class $C^2$ near $(x_0,y_0)$ and not all zero there.
- Hyperbolic case. If $B^2-4AC\gt0$ at $(x_0,y_0)$, there are $C^2$ coordinates $\xi,\eta$ near $(x_0,y_0)$ with $J\ne0$ — the two families of characteristics, $\xi=\text{const}$ and $\eta=\text{const}$ — in which the equation becomes $$u_{\xi\eta}=\Phi_1(\xi,\eta,u,u_\xi,u_\eta)\qquad\text{(first canonical form)},$$ and, with $\alpha=\xi+\eta$, $\beta=\xi-\eta$, $\ u_{\alpha\alpha}-u_{\beta\beta}=\Phi_2$ (second canonical form).
- Parabolic case. If $B^2-4AC\equiv0$ in a neighbourhood of $(x_0,y_0)$, there are coordinates $\xi$ (the characteristic family) and $\eta$ (any function with $J\ne0$) in which $$u_{\eta\eta}=\Phi_1(\xi,\eta,u,u_\xi,u_\eta).$$
- Elliptic case. If $B^2-4AC\lt0$ at $(x_0,y_0)$ and $A,B,C$ are real-analytic, there are coordinates $\alpha,\beta$ in which $$u_{\alpha\alpha}+u_{\beta\beta}=\Phi_1(\alpha,\beta,u,u_\alpha,u_\beta).$$ They are the real and imaginary parts of a complex solution $\varphi=\alpha+i\beta$ of $A\varphi_x^2+B\varphi_x\varphi_y+C\varphi_y^2=0$. (The conclusion remains true for Hölder continuous coefficients — the theorem on isothermal coordinates of Korn and Lichtenstein — but the proof is much harder.)
Proof. Preliminary normalisation. We may assume $A\ne0$ near $(x_0,y_0)$. If $A(x_0,y_0)=0$ but $C(x_0,y_0)\ne0$, interchange $x$ and $y$. If $A=C=0$ at the point then $B\ne0$ there, and the linear change $x'=x+y$, $y'=x-y$ produces (Theorem 4.1 with $\xi=x+y$) the new coefficient $\bar A=A+B+C=B\ne0$ at the point. By continuity $A\ne0$ in a neighbourhood.
(1) Hyperbolic. By continuity $B^2-4AC\gt0$ near the point, so the slopes $\lambda_\pm=\dfrac{B\pm\sqrt{B^2-4AC}}{2A}$ are distinct and of class $C^2$. By the existence and uniqueness theorem and the differentiable dependence of solutions on initial values (§1.1), each ODE $y'=\lambda_\pm(x,y)$ has a local first integral: let $\xi(x,y)$ be the value at $x=x_0$ of the solution of $y'=\lambda_+$ through $(x,y)$, and define $\eta$ in the same way from $\lambda_-$. Then $\xi,\eta\in C^2$, $\xi(x_0,y)=y$ so $\xi_y=1$ at $(x_0,y_0)$ and hence $\xi_y\ne0$ nearby; likewise $\eta_y\ne0$. Since $\xi$ is constant along the solutions, $\xi_x+\lambda_+\xi_y=0$, i.e. $\xi_x/\xi_y=-\lambda_+$, and $A\lambda_+^2-B\lambda_++C=0$ gives $$\bar A=\xi_y^2\big(A\lambda_+^2-B\lambda_++C\big)=0,$$ and in the same way $\bar C=0$. The Jacobian is $J=\xi_x\eta_y-\xi_y\eta_x=(\lambda_--\lambda_+)\xi_y\eta_y\ne0$. Theorem 4.1 gives $\bar B^2=\bar B^2-4\bar A\bar C=J^2(B^2-4AC)\gt0$, so $\bar B\ne0$, and dividing the transformed equation by $\bar B$ gives the first canonical form. For the second, $\partial_\xi=\partial_\alpha+\partial_\beta$ and $\partial_\eta=\partial_\alpha-\partial_\beta$, so $u_{\xi\eta}=u_{\alpha\alpha}-u_{\beta\beta}$.
(2) Parabolic. Now $\lambda=B/(2A)$ is a double root, of class $C^2$. Let $\xi$ be a first integral of $y'=\lambda$ with $\xi_y\ne0$ as before, and take $\eta=x$; then $J=\xi_x\cdot0-\xi_y\cdot1=-\xi_y\ne0$. As in (1), $\bar A=0$. Theorem 4.1 gives $\bar B^2-4\bar A\bar C=J^2\cdot0=0$, so $\bar B=0$ as well. Finally $\bar C\ne0$, for otherwise $\bar M=PMP^{\mathsf T}$ would be the zero matrix although $M\ne0$ and $P$ is invertible. Divide by $\bar C$.
(3) Elliptic (sketch). The roots $\lambda,\bar\lambda$ of $A\lambda^2-B\lambda+C=0$ are now non-real. With analytic coefficients the complex first-order equation $\varphi_x+\lambda(x,y)\varphi_y=0$ has an analytic solution with $\varphi_y\ne0$ near the point (Cauchy–Kovalevskaya, or by solving the complex ODE $y'=\lambda$ with analytic dependence on the initial value). Since $\varphi_x=-\lambda\varphi_y$, we have $$Q(\nabla\varphi):=A\varphi_x^2+B\varphi_x\varphi_y+C\varphi_y^2=\varphi_y^2(A\lambda^2-B\lambda+C)=0 .$$ Write $\varphi=\alpha+i\beta$ and let $Q(\mathbf a,\mathbf b)$ denote the symmetric bilinear form with $Q(\mathbf a,\mathbf a)=Q(\mathbf a)$. Then $$0=Q(\nabla\alpha+i\nabla\beta)=\big[Q(\nabla\alpha)-Q(\nabla\beta)\big]+2i\,Q(\nabla\alpha,\nabla\beta),$$ and in the coordinates $(\alpha,\beta)$ this says exactly $\bar A=\bar C$ and $\bar B=0$. The Jacobian is nonzero: if $\nabla\alpha$ and $\nabla\beta$ were linearly dependent at a point, the ratio $\varphi_x/\varphi_y=-\lambda$ would be real there. So $\bar A=\bar C\ne0$ (again because $\bar M\ne0$), and dividing by $\bar A$ gives $u_{\alpha\alpha}+u_{\beta\beta}=\Phi_1$. ∎
- Classify in the region of interest with $B^2-4AC$; make sure $A,B,C$ do not all vanish there.
- Characteristic equation $A(y')^2-By'+C=0$ (mind the sign of $B$). Solve for $y'$ and integrate each ODE to a first integral $\varphi(x,y)=c$.
- New variables. Hyperbolic: $\xi=\varphi_+$, $\eta=\varphi_-$. Parabolic: $\xi=\varphi$ and $\eta$ any simple function with $J\ne0$ (usually $\eta=x$ or $\eta=y$). Elliptic: $\alpha=\operatorname{Re}\varphi$, $\beta=\operatorname{Im}\varphi$ for the complex first integral.
- Transform with Theorem 4.1. Compute only $\bar B$ (hyperbolic), $\bar C$ (parabolic) or $\bar A=\bar C$ (elliptic) — the other second-order coefficients vanish automatically, which is a good check — and the first-order coefficients $\bar a_1,\bar a_2$, which are the principal part applied to $\xi$ and $\eta$ plus $a_1\xi_x+a_2\xi_y$, etc.
- Rewrite all coefficients in terms of the new variables, divide, and, if a general solution is wanted, integrate the canonical form (e.g. $u_{\xi\eta}=0\Rightarrow u=f(\xi)+g(\eta)$; $u_{\eta\eta}=0\Rightarrow u=f(\xi)+\eta\,g(\xi)$).
For constant coefficients everything is explicit, and one can go one step further and remove the first-order terms as well.
Let $Au_{xx}+Bu_{xy}+Cu_{yy}+a_1u_x+a_2u_y+a_0u=g$ have real constant coefficients with $A\ne0$, and let $\lambda_\pm$ be the roots of $A\lambda^2-B\lambda+C=0$.
- The linear change of variables $\xi=y-\lambda_+x$, $\eta=y-\lambda_-x$ (hyperbolic); $\xi=y-\tfrac{B}{2A}x$, $\eta=x$ (parabolic); $\alpha=y-\mu x$, $\beta=\nu x$ where $\lambda_\pm=\mu\pm i\nu$, $\nu\ne0$ (elliptic) brings the principal part to a constant multiple of $u_{\xi\eta}$, $u_{\eta\eta}$ or $u_{\alpha\alpha}+u_{\beta\beta}$.
- Write the result as $u_{\xi\eta}+a\,u_\xi+b\,u_\eta+c\,u=h$ (hyperbolic). Then $u=e^{-b\xi-a\eta}\,v$ gives $$v_{\xi\eta}+(c-ab)\,v=e^{b\xi+a\eta}\,h .$$ For $u_{\eta\eta}+a\,u_\xi+b\,u_\eta+c\,u=h$ with $a\ne0$ (parabolic), $u=e^{\kappa\xi-b\eta/2}\,v$ with $\kappa=\dfrac{b^2/4-c}{a}$ gives $v_{\eta\eta}+a\,v_\xi=e^{-\kappa\xi+b\eta/2}\,h$, a heat equation. For $u_{\alpha\alpha}+u_{\beta\beta}+a\,u_\alpha+b\,u_\beta+c\,u=h$ (elliptic), $u=e^{-(a\alpha+b\beta)/2}\,v$ gives $$v_{\alpha\alpha}+v_{\beta\beta}+\Big(c-\frac{a^2+b^2}{4}\Big)v=e^{(a\alpha+b\beta)/2}\,h .$$
Proof. (1) The functions $y-\lambda x$ are first integrals of $y'=\lambda$, so by the proof of Theorem 4.4 the stated $\xi,\eta$ (respectively $\alpha+i\beta=y-\lambda_+x$) are characteristic coordinates; their Jacobians are $\lambda_--\lambda_+\ne0$, $-1$ and $-\nu\ne0$. In the elliptic case one checks directly that $\bar A=A\mu^2-B\mu+C$ and $\bar C=A\nu^2$ are equal (both equal $\tfrac{4AC-B^2}{4A}$, since $\mu=\tfrac{B}{2A}$ and $\nu^2=\tfrac{4AC-B^2}{4A^2}$) and that $\bar B=\nu(B-2A\mu)=0$. Second derivatives of linear functions vanish, so no first-order terms are created and the coefficients stay constant.
(2) For the hyperbolic case put $E=e^{-b\xi-a\eta}$, $u=Ev$. Then $u_\xi=E(v_\xi-bv)$, $u_\eta=E(v_\eta-av)$ and $u_{\xi\eta}=E(v_{\xi\eta}-av_\xi-bv_\eta+abv)$. Hence $$u_{\xi\eta}+au_\xi+bu_\eta+cu=E\big(v_{\xi\eta}-av_\xi-bv_\eta+abv+av_\xi-abv+bv_\eta-abv+cv\big)=E\big(v_{\xi\eta}+(c-ab)v\big),$$ and dividing $h$ by $E$ gives the claim. The other two cases are the same computation: with $u=e^{\kappa\xi+m\eta}v$ the parabolic operator becomes $$e^{\kappa\xi+m\eta}\big(v_{\eta\eta}+av_\xi+(b+2m)v_\eta+(a\kappa+bm+m^2+c)v\big),$$ and $m=-b/2$, $\kappa=(b^2/4-c)/a$ make the last two coefficients vanish; in the elliptic case the choice $e^{-(a\alpha+b\beta)/2}$ cancels both first-order terms and leaves $c-\tfrac{a^2+b^2}{4}$. ∎
So, up to lower-order terms that cannot be removed, every constant-coefficient equation is one of $v_{\xi\eta}+kv=h$ (the Klein–Gordon or telegraph type), $v_{\eta\eta}+av_\xi=h$ (heat), $v_{\eta\eta}=h$ (degenerate parabolic, when $a=0$) and $v_{\alpha\alpha}+v_{\beta\beta}+kv=h$ (Helmholtz type). The research project at the end of this section applies exactly this idea to the Black–Scholes equation of finance.
Linear Equations with Constant Coefficients
We now solve linear equations of any order whose coefficients are constants. It is traditional to call the unknown $z$ and to write $D=\dfrac{\partial}{\partial x}$, $D'=\dfrac{\partial}{\partial y}$, so that, for example, $z_{xx}-3z_{xy}+2z_{yy}=(D^2-3DD'+2D'^2)z$. Because the coefficients are constant, these operators commute and can be multiplied and factored like polynomials in two variables — the same algebra that made $1/f(D)$ work for ODEs in §2.2.
Let $F(X,Y)=\sum_{i+j\le n}c_{ij}X^iY^j$ be a polynomial with real constant coefficients and total degree $n\ge1$. The equation $$F(D,D')z=f(x,y),\qquad F(D,D')=\sum c_{ij}D^iD'^j,$$ is a linear equation of order $n$ with constant coefficients. The operator (and the equation) is homogeneous if every term has the same total order $n$ (as in $D^2-3DD'+2D'^2$) and non-homogeneous otherwise (as in $D^2-D'$ or $D-D'-1$). The general solution of $F(D,D')z=0$ is called the complementary function (CF); any one solution of $F(D,D')z=f$ is a particular integral (PI); and, by linearity, the general solution of $F(D,D')z=f$ is $z=\text{CF}+\text{PI}$.
$F(D,D')$ is reducible if it factors into first-order linear factors with constant (possibly complex) coefficients, $F(D,D')=\prod_k(\alpha_kD+\beta_kD'+\gamma_k)$, and irreducible otherwise. A homogeneous operator is always reducible over $\mathbb{C}$ (a homogeneous polynomial in two variables splits into linear factors: $F(X,Y)=c\prod(X-m_kY)$ up to factors $Y$), and over $\mathbb{R}$ when all the roots $m_k$ of $F(m,1)=0$ are real. The heat operator $D^2-D'$ is irreducible (Exercise IV.5).
Everything rests on one observation about a single first-order factor: in the coordinates that follow its characteristic lines, it becomes an ordinary derivative.
Let $m,c\in\mathbb{R}$ and write $z(x,y)=Z(s,w)$ with $s=x$, $w=y+mx$.
- $(D-mD'-c)\,z=Z_s-cZ$.
- Every $C^1$ solution of $(D-mD'-c)z=0$ on $\mathbb{R}^2$ (more generally, on a domain that meets each line $y+mx=\text{const}$ in an interval) is $$z=e^{cx}\,\varphi(y+mx),\qquad\varphi\in C^1 .$$
- For continuous $h$, a particular solution of $(D-mD'-c)z=h$ is $$z(x,y)=e^{cx}\int_{x_0}^{x}e^{-c\sigma}\,h\big(\sigma,\;y+mx-m\sigma\big)\,d\sigma .$$ In particular, for $c=0$: to compute $\dfrac{1}{D-mD'}\,h$, integrate $h(x,\,w-mx)$ with respect to $x$ treating $w$ as a constant, then replace $w$ by $y+mx$.
Proof. (1) $z(x,y)=Z(x,\,y+mx)$, so by the chain rule $z_x=Z_s+mZ_w$ and $z_y=Z_w$; hence $z_x-mz_y-cz=Z_s-cZ$. (2) The equation becomes $Z_s-cZ=0$, i.e. $\partial_s\big(e^{-cs}Z\big)=0$. On each line $w=\text{const}$, which meets the domain in an interval, $e^{-cs}Z$ is therefore constant; call its value $\varphi(w)$. So $Z=e^{cs}\varphi(w)$, i.e. $z=e^{cx}\varphi(y+mx)$, and $\varphi(w)=z(0,w)$ is $C^1$. Conversely every such function is a solution, by (1). (3) The function $Z(s,w)=e^{cs}\int_{x_0}^se^{-c\sigma}h(\sigma,w-m\sigma)\,d\sigma$ satisfies $Z_s-cZ=h(s,w-ms)$ by the fundamental theorem of calculus, and $h(s,w-ms)=h(x,y)$ because $y=w-mx$. ∎
This is the theorem on homogeneous linear first-order equations of §3.1 in operator language: $(D-mD')z=0$ says that $z$ is constant along the characteristic lines $y+mx=\text{const}$. A factor $D'-c$ (which occurs when $F$ has no $D^n$ term) is handled in the same way with the roles of $x$ and $y$ exchanged: its solutions are $e^{cy}\varphi(x)$. The next identities let us evaluate $F(D,D')$ on the functions that will appear.
Let $F$ be a polynomial, $a,b$ constants (real or complex), $V$ and $\varphi$ smooth functions. Then
- $F(D,D')\,e^{ax+by}=F(a,b)\,e^{ax+by}$;
- (exponential shift) $F(D,D')\big(e^{ax+by}V\big)$ $=e^{ax+by}\,F(D+a,\,D'+b)\,V$;
- if $F$ is homogeneous of degree $n$, then $F(D,D')\,\varphi(ax+by)=F(a,b)\,\varphi^{(n)}(ax+by)$;
- (Leibniz rule in $x$) $F(D,D')\big(xV\big)$ $=x\,F(D,D')V+F_X(D,D')\,V$, where $F_X=\partial F/\partial X$.
Proof. All four statements are linear in $F$, so it suffices to check them for a monomial $D^iD'^j$. (1) $D$ multiplies $e^{ax+by}$ by $a$ and $D'$ by $b$, so $D^iD'^je^{ax+by}=a^ib^je^{ax+by}$. (2) $D(e^{ax+by}V)=e^{ax+by}(D+a)V$ and $D'(e^{ax+by}V)=e^{ax+by}(D'+b)V$; iterating, $D^iD'^j(e^{ax+by}V)=e^{ax+by}(D+a)^i(D'+b)^jV$. (3) Each $D$ produces a factor $a$ and one more derivative of $\varphi$, each $D'$ a factor $b$: $D^iD'^j\varphi(ax+by)=a^ib^j\varphi^{(i+j)}(ax+by)$; when $i+j=n$ for every term, summing gives $F(a,b)\varphi^{(n)}$. (4) Since $D'x=0$, $D'^j(xV)=xD'^jV$; and by Leibniz' rule $D^i(xU)=xD^iU+iD^{i-1}U$ (all higher derivatives of $x$ vanish). Hence $D^iD'^j(xV)=xD^iD'^jV+iD^{i-1}D'^jV$, and $iX^{i-1}Y^j=\partial_X(X^iY^j)$. ∎
Suppose $F(D,D')=\prod_{k=1}^{K}(D-m_kD'-c_k)^{r_k}$ with distinct pairs $(m_k,c_k)$ and $r_1+\dots+r_K=n$. Then $$z=\sum_{k=1}^{K}e^{c_kx}\sum_{j=0}^{r_k-1}x^j\,\varphi_{kj}(y+m_kx)$$ is a solution of $F(D,D')z=0$ for all smooth functions $\varphi_{kj}$ — $n$ arbitrary functions in all — and every $C^\infty$ solution on $\mathbb{R}^2$ is of this form. In particular, for a homogeneous operator ($c_k=0$):
Proof. The listed functions are solutions. In the coordinates of Theorem 4.6, $e^{cx}x^j\varphi(y+mx)$ is $Z=e^{cs}s^j\varphi(w)$ and $(D-mD'-c)$ acts as $\partial_s-c$, which sends $e^{cs}s^j\varphi(w)$ to $j\,e^{cs}s^{j-1}\varphi(w)$. Applying it $r$ times gives $0$ whenever $j\lt r$. The factors of $F$ commute, so $F(D,D')$ kills every term.
Step 1: one factor, repeated. Let $S_r$ be the set of functions $e^{cx}\sum_{j\lt r}x^j\varphi_j(y+mx)$. We show by induction on $r$ that every smooth solution of $(D-mD'-c)^rz=0$ lies in $S_r$. For $r=1$ this is Theorem 4.6(2). If $(D-mD'-c)^rz=0$, then $\zeta=(D-mD'-c)z$ solves the equation with exponent $r-1$, so $\zeta\in S_{r-1}$: in the coordinates $(s,w)$, $Z_s-cZ=e^{cs}\sum_{j\lt r-1}s^j\psi_j(w)$. Multiplying by $e^{-cs}$ and integrating in $s$, $$Z=e^{cs}\Big[\varphi_0(w)+\sum_{j\lt r-1}\frac{s^{j+1}}{j+1}\psi_j(w)\Big]\in S_r .$$
Step 2: other factors map $S_r$ onto itself. Let $P=D-m'D'-c'$ with $(m',c')\ne(m,c)$. For $\psi$ smooth, $$P\big[e^{cx}x^j\psi(y+mx)\big]=e^{cx}\Big[x^j\big((m-m')\psi'+(c-c')\psi\big)(y+mx)+j\,x^{j-1}\psi(y+mx)\Big]$$ (compute $D$ and $D'$ directly). Given a target $e^{cx}x^j\chi(y+mx)$, solve the linear ODE $(m-m')\psi'+(c-c')\psi=\chi$ — possible because $m-m'$ and $c-c'$ are not both zero — and the error term involves the lower power $x^{j-1}$; descending in $j$ we reach any element of $S_r$. So $P(S_r)=S_r$.
Step 3: several factors. Induction on $K$. Write $F=(D-m_1D'-c_1)^{r_1}G$, where $G$ is the product of the remaining factors, and let $F(D,D')z=0$. Then $\zeta=G(D,D')z$ satisfies $(D-m_1D'-c_1)^{r_1}\zeta=0$, so $\zeta\in S_{r_1}$ by Step 1. By Step 2, applied factor by factor, there is $z_1\in S_{r_1}$ with $G(D,D')z_1=\zeta$. Then $G(D,D')(z-z_1)=0$, so by the induction hypothesis $z-z_1$ is a sum of elements of the sets belonging to the other factors, and $z$ has the stated form. ∎
When a root $m$ is repeated, $x^j\varphi_j(y+mx)$ may be replaced by $y^j\varphi_j(y+mx)$ if $m\ne0$, since $y=(y+mx)-mx$: the two families are the same. Complex roots, e.g. $D^2+D'^2=(D-iD')(D+iD')$, give $z=\varphi_1(y+ix)+\varphi_2(y-ix)$ with complex $\varphi$'s; the real solutions are the real parts $\operatorname{Re}\,\varphi(x+iy)$ of analytic functions — the harmonic functions of §4.2. For an irreducible operator no such formula exists, and we fall back on exponentials.
- For any $F$, $z=e^{ax+by}$ solves $F(D,D')z=0$ whenever $F(a,b)=0$; hence so does every finite sum $\sum_kA_ke^{a_kx+b_ky}$ with $F(a_k,b_k)=0$ (and every series of such terms that may be differentiated term by term). These form the CF of an irreducible equation.
- For a fixed $m$, $F(D,D')\varphi(y+mx)=0$ holds for every smooth $\varphi$ if and only if $D-mD'$ divides $F(D,D')$.
Proof. (1) is Theorem 4.7(1) and linearity. (2) Split $F=\sum_kF_k$ into homogeneous parts of degree $k$. By Theorem 4.7(3), $F(D,D')\varphi(y+mx)=\sum_kF_k(m,1)\,\varphi^{(k)}(y+mx)$. Choosing $\varphi(t)=e^{\mu t}$ shows that this vanishes for every $\varphi$ iff $\sum_kF_k(m,1)\mu^k=0$ for every $\mu$, i.e. iff $F_k(m,1)=0$ for every $k$. A homogeneous polynomial $F_k(X,Y)$ vanishes at $(m,1)$ iff $X-mY$ divides it. So the condition is that $X-mY$ divides every homogeneous part, which is equivalent to dividing $F$ (if $F=(X-mY)G$ then $F_k=(X-mY)G_{k-1}$). ∎
For the heat operator $D^2-D'$ the homogeneous parts are $D^2$ and $-D'$: $m^2=0$ forces $m=0$, but then $-1\ne0$. So no family $\varphi(y+mx)$ of solutions with an arbitrary function exists, while $e^{ax+a^2y}$ is a solution for every (complex) $a$; with $a=ik$ its real and imaginary parts give $e^{-k^2y}\cos kx$ and $e^{-k^2y}\sin kx$, the decaying modes that build every solution of the heat equation in §4.3.
Write $\dfrac1{F(D,D')}f$ for any particular integral of $F(D,D')z=f$.
- Exponentials. If $F(a,b)\ne0$: $\ \dfrac1{F(D,D')}e^{ax+by}=\dfrac{e^{ax+by}}{F(a,b)}$. If $F(a,b)=0$ but $F_X(a,b)\ne0$: $\ \dfrac1{F(D,D')}e^{ax+by}=\dfrac{x\,e^{ax+by}}{F_X(a,b)}$ (similarly $\dfrac{y\,e^{ax+by}}{F_Y(a,b)}$ if $F_Y(a,b)\ne0$).
- Functions of $ax+by$, homogeneous $F$ of degree $n$. If $F(a,b)\ne0$: $\ \dfrac1{F(D,D')}\varphi(ax+by)=\dfrac{1}{F(a,b)}\,\Phi_n(ax+by)$, where $\Phi_n$ is any $n$-th antiderivative of $\varphi$ ($\Phi_n^{(n)}=\varphi$). For a repeated factor: $\ \dfrac{1}{(D-mD')^r}\varphi(y+mx)=\dfrac{x^r}{r!}\,\varphi(y+mx)$.
- Sines and cosines. If $F(D,D')=G(D^2,DD',D'^2)$ and $G(-a^2,-ab,-b^2)\ne0$: $\ \dfrac1{F(D,D')}\sin(ax+by)=\dfrac{\sin(ax+by)}{G(-a^2,-ab,-b^2)}$, and likewise for $\cos$. In general use $\sin\theta=\operatorname{Im}e^{i\theta}$ and rule 1 with complex $a,b$.
- Exponential times anything. $\ \dfrac1{F(D,D')}\big(e^{ax+by}V\big)$ $=e^{ax+by}\,\dfrac{1}{F(D+a,D'+b)}\,V$.
- Polynomials. If $F(0,0)=c\ne0$, write $F=c\,(1-T)$, where $T$ has no constant term; then for a polynomial $f$ of degree $N$, $\ \dfrac1c\big(1+T+T^2+\dots+T^N\big)f$ is a PI. If $F(0,0)=0$, factor out a power of $D$ (or $D'$) and expand the rest in powers of $D'/D$, as in Example 7(a); only finitely many terms act, but check the result by substitution or obtain it by rule 6.
- The general method. Factor $F$ and apply Theorem 4.6(3) one factor at a time: $\dfrac1{D-mD'}h=\displaystyle\int h(x,\,w-mx)\,dx\Big|_{w=y+mx}$.
Proof. (1) The first formula is Theorem 4.7(1). For the second, Theorem 4.7(4) with $V=e^{ax+by}$ gives $$F(D,D')(xe^{ax+by})=xF(a,b)e^{ax+by}+F_X(a,b)e^{ax+by}=F_X(a,b)e^{ax+by}.$$ (2) The first formula is Theorem 4.7(3) applied to $\Phi_n$. For the second, $$(D-mD')\Big[\tfrac{x^r}{r!}\varphi(y+mx)\Big]=\tfrac{x^{r-1}}{(r-1)!}\varphi+\tfrac{x^r}{r!}\big(m\varphi'-m\varphi'\big)=\tfrac{x^{r-1}}{(r-1)!}\varphi(y+mx);$$ apply this $r$ times. (3) $D^2$, $DD'$, $D'^2$ multiply $\sin(ax+by)$ by $-a^2$, $-ab$, $-b^2$, so $G(D^2,DD',D'^2)\sin(ax+by)$ $=G(-a^2,-ab,-b^2)\sin(ax+by)$. (4) By Theorem 4.7(2), if $F(D+a,D'+b)W=V$ then $F(D,D')(e^{ax+by}W)=e^{ax+by}V$. (5) Every term of $T^{N+1}$ is a derivative of total order at least $N+1$, so $T^{N+1}f=0$. Hence $c(1-T)\cdot\tfrac1c(1+T+\dots+T^N)f$ $=(1-T^{N+1})f=f$. (6) is Theorem 4.6(3), used once per factor. ∎
- CF. For a homogeneous operator put $D\to m$, $D'\to1$ and solve the auxiliary equation $F(m,1)=0$; each root $m$ gives $\varphi(y+mx)$, a root of multiplicity $r$ gives $\sum_{j\lt r}x^j\varphi_j(y+mx)$. For a non-homogeneous operator try to factor into $(D-mD'-c)$'s (each gives $e^{cx}\varphi(y+mx)$); if it is irreducible, the CF is $\sum Ae^{ax+by}$ over the curve $F(a,b)=0$.
- PI, by the form of $f$. $e^{ax+by}$: rule 1 (check whether $F(a,b)=0$ first!). $\sin/\cos(ax+by)$: rule 3, or the imaginary part of rule 1. A function of $ax+by$ with homogeneous $F$: rule 2. Polynomials: rule 5. $e^{ax+by}\times$(anything): rule 4. Anything else: rule 6, one factor at a time.
- Check by applying $F(D,D')$ to your PI — or use the check-lab after Exercise card III.
- Answer $z=\text{CF}+\text{PI}$, with the correct number $n$ of arbitrary functions.
- The factor 2. With the principal part written $Au_{xx}+2Bu_{xy}+Cu_{yy}$ the test is $B^2-AC$; with $Au_{xx}+Bu_{xy}+Cu_{yy}$ it is $B^2-4AC$. Mixing them misclassifies $u_{xx}+2u_{xy}+2u_{yy}=0$: correctly $4-8\lt0$, elliptic; the mixed-up "$2^2-1\cdot2\gt0$" says hyperbolic.
- Type is a pointwise notion. $y\,u_{xx}+u_{yy}=0$ is elliptic, parabolic and hyperbolic in different parts of the plane; asking for "the" type of a variable-coefficient equation is meaningless without a region.
- The parabolic set can be anything closed. A line ($y\,u_{xx}+u_{yy}$), a circle (Example 1(b)), a single point ($u_{xx}+(x^2+y^2)u_{yy}$), the whole plane minus a point (Example 1(c)), the whole plane ($u_{xx}+4u_{xy}+4u_{yy}$), or empty ($u_{xx}+u_{yy}$).
- Points where $A=B=C=0$ have no type. $x^2u_{xx}-y^2u_{yy}=0$ is not of second order at the origin; do not call it parabolic there.
- Lower-order terms are irrelevant to type — $u_t=u_{xx}$ is parabolic although it "looks first order in $t$", and $u_{xx}-u_{yy}+100u_x=0$ is hyperbolic — and so is multiplication by a negative function: $-u_{xx}-u_{yy}=0$ is elliptic.
- Invariance needs $J\ne0$. The type is invariant under nonsingular changes of variables. Where the Jacobian vanishes (e.g. $\xi=x^3$ at $x=0$) the transformed coefficients can all vanish and nothing can be concluded.
- Quasi-linear equations: the type depends on the solution. $u_{xx}+u\,u_{yy}=0$ is elliptic along a solution where $u\gt0$ and hyperbolic where $u\lt0$.
- Sign in the characteristic equation. It is $A(y')^2-By'+C=0$. For $u_{xx}+2u_{xy}-3u_{yy}=0$ the slopes are $3$ and $-1$; writing $+By'$ gives the wrong pair $-3$, $1$.
- "Hyperbolic $\Rightarrow u=f(\xi)+g(\eta)$" is false. It needs the canonical form $u_{\xi\eta}=0$ with no lower-order terms (Example 5(d): $u_{xx}-u_{yy}+u=0$), and even then only on domains that meet the characteristic lines in intervals — on a slit plane $u_{xy}=0$ has solutions that are not $f(x)+g(y)$ (Exercise IV.4).
- Repeated factors need the extra $x$. The CF of $(D-mD')^2z=0$ is $\varphi_1(y+mx)+x\varphi_2(y+mx)$; writing $\varphi_1(y+mx)+\varphi_2(y+mx)$ is just one arbitrary function. An order-$n$ equation needs $n$ arbitrary functions.
- Irreducible operators have no "arbitrary-function" CF. For $D^2-D'$ the CF is a superposition of exponentials $e^{ax+a^2y}$; no formula $\sum\varphi_k(y+m_kx)$ exists (Theorem 4.9).
- The exponential rule fails when $F(a,b)=0$. $(D^2-D'^2)z=e^{x+y}$: $F(1,1)=0$, and the PI is $\tfrac{x}{2}e^{x+y}$ (rule 1 with $F_X=2X$), not "$e^{x+y}/0$".
- A PI is not unique. For $(D-D')^2z=e^{x+y}$ both $\tfrac{x^2}2e^{x+y}$ and $-\tfrac{xy}{2}e^{x+y}$ are particular integrals; they differ by $\tfrac{x(x+y)}2e^{x+y}=x\,\varphi(y+x)$, a member of the CF. In a multiple-choice question check each option by substitution.
Exercises
- 1 Classify: (a) $u_{xx}+6u_{xy}+9u_{yy}=0$; (b) $2u_{xx}-3u_{xy}-2u_{yy}=0$; (c) $u_{xx}+u_{xy}+u_{yy}+u_x=0$; (d) $u_{xy}=u_x$.
- 2 Determine the type of $(1+x)u_{xx}+2xy\,u_{xy}-y^2u_{yy}=0$ at every point of the plane.
- 3 Find the regions where $(x^2-1)u_{xx}+2xy\,u_{xy}+(y^2-1)u_{yy}=0$ is hyperbolic, parabolic and elliptic.
- 4 For which real $\lambda$ is $u_{xx}+2\lambda u_{xy}+4u_{yy}+u=0$ hyperbolic, parabolic, elliptic?
- 5 Classify the telegraph equation $u_{tt}+2k\,u_t=c^2u_{xx}$ ($c\gt0$) and the diffusion equation $u_t=\big(k(x)u_x\big)_x$ with $k\gt0$ of class $C^1$, in the $(x,t)$-plane.
- 6 Transform $u_{xx}-u_{yy}=0$ by $\xi=x+y$, $\eta=x-y$ using the formulas of Theorem 4.1, and verify $\bar B^2-4\bar A\bar C=J^2(B^2-4AC)$.
- 7 Show that the substitution $u=e^{\alpha x+\beta y}v$ ($\alpha,\beta$ constants), and multiplication of the equation by a nowhere-zero function, do not change the type.
- 8 Prove that for continuous $A,B,C$ the hyperbolic and elliptic sets are open. Give an equation whose parabolic set has interior points and one whose parabolic set is a single point.
- 9 Classify the three-variable equations (a) $u_{xx}+2u_{yy}+3u_{zz}+2u_{xy}+2u_{yz}=0$; (b) $u_{xy}+u_{yz}+u_{zx}=0$.
- 1 Find the characteristics of $u_{xx}-2u_{xy}-3u_{yy}=0$, reduce it to canonical form and find its general solution.
- 2 Reduce $y^2u_{xx}-x^2u_{yy}=0$ to canonical form in the quadrant $x\gt0$, $y\gt0$.
- 3 Reduce $u_{xx}+2u_{xy}+u_{yy}=0$ to canonical form and find its general solution.
- 4 Show that $\alpha=x+y$, $\beta=2x$ reduces $u_{xx}-2u_{xy}+5u_{yy}=0$ to Laplace's equation.
- 5 Show that in $y\lt0$ the substitution $\xi=x+\tfrac23(-y)^{3/2}$, $\eta=x-\tfrac23(-y)^{3/2}$ reduces the Tricomi equation $y\,u_{xx}+u_{yy}=0$ to $u_{\xi\eta}=\dfrac{u_\xi-u_\eta}{6(\xi-\eta)}$.
- 6 Show that the characteristics of the wave equation $u_{tt}=c^2u_{xx}$ are the lines $x\pm ct=\text{const}$, and that every $C^2$ solution on $\mathbb{R}^2$ has the form $f(x+ct)+g(x-ct)$.
- 7 Remove the first-order terms from $u_{\xi\eta}+2u_\xi+3u_\eta=0$.
- 8 Show that $e^{2x}u_{xx}+2e^{x+y}u_{xy}+e^{2y}u_{yy}=0$ is parabolic everywhere, find its characteristics, and reduce it to canonical form.
- 1 Solve $(D^2-5DD'+6D'^2)z=0$.
- 2 Solve $(D^3-3D^2D'+2DD'^2)z=0$.
- 3 Solve $(D^2+4DD'+4D'^2)z=0$.
- 4 Solve $(D^2-D'^2)z=e^{x+2y}$.
- 5 Find a particular integral of $(D^2-D'^2)z=e^{x-y}$.
- 6 Find a particular integral of $(D^2-2DD'+D'^2)z=\sin(2x+3y)$.
- 7 Find a particular integral of $(D^2-3DD'+2D'^2)z=x+y$.
- 8 Use the general method (rule 6) to find a particular integral of $(D^2-DD'-2D'^2)z=(y-1)e^x$.
- 9 Solve $(D-D'-1)(D-D'-2)z=e^{2x-y}$, writing the operator out in full first.
- 10 Show that $e^{ax+a^2y}$ solves $(D^2-D')z=0$ for every complex $a$, and find a real solution of the form $X(x)Y(y)$ with $z(x,0)=\cos2x$.
- 1 Prove the Leibniz rule $F(D,D')(xV)$ $=x\,F(D,D')V+F_X(D,D')V$ and deduce the rule $\dfrac{1}{F(D,D')}e^{ax+by}=\dfrac{x\,e^{ax+by}}{F_X(a,b)}$ when $F(a,b)=0\ne F_X(a,b)$. What should one do when $F(a,b)=F_X(a,b)=0$?
- 2 Let $Au_{xx}+Bu_{xy}+Cu_{yy}$ have constant coefficients with $A\gt0$ and $B^2-4AC\lt0$. Find an explicit linear change of variables that turns it into $k(u_{\alpha\alpha}+u_{\beta\beta})$, and compute $k$.
- 3 Let $u(x,y)=\big(\max(x-y,0)\big)^2$. Show that $u\in C^1(\mathbb{R}^2)$, that $u_{xx}-u_{yy}=0$ off the line $x=y$, and that the second derivatives of $u$ jump across that line. Explain why this does not contradict Theorem 4.3.
- 4 Let $\Omega=\mathbb{R}^2\setminus\{(x,0):x\ge0\}$ (the plane slit along the positive $x$-axis). Construct $u\in C^\infty(\Omega)$ with $u_{xy}=0$ in $\Omega$ that is not of the form $f(x)+g(y)$.
- 5 Prove that $D^2-D'$ is irreducible, and find all solutions of $(D^2-D')z=0$ of the form $X(x)Y(y)$ with $XY\ne0$.
- 6 Prove rule 2 of Theorem 4.10 for homogeneous $F$ with $F(a,b)\ne0$, and use it to find a particular integral of $(D^2-D'^2)z=(x+2y)^{-2}$ in the region $x+2y\gt0$.
E1. CSIR Part B GATE MA The equation $u_{xx}+2y\,u_{xy}+x\,u_{yy}=0$ is hyperbolic exactly in the set (A) $\{(x,y):x\lt y^2\}$; (B) $\{(x,y):x\gt y^2\}$; (C) $\{(x,y):y\lt x^2\}$; (D) $\{(x,y):x\lt0\}$.
E2. CSIR Part B The equation $u_{xx}+2x\,u_{xy}+x^2u_{yy}+u_y=0$ is (A) parabolic, with characteristics $y-\tfrac12x^2=c$; (B) parabolic, with characteristics $y+\tfrac12x^2=c$; (C) hyperbolic, with characteristics $y\pm\tfrac12x^2=c$; (D) elliptic, with no real characteristics.
E3. CSIR Part B GATE MA A particular integral of $(D^2-2DD'+D'^2)z=e^{x+y}$ is (A) $x\,e^{x+y}$; (B) $\tfrac12x^2e^{x+y}$; (C) $\tfrac12e^{x+y}$; (D) $xy\,e^{x+y}$.
E4. CSIR Part C Consider $x\,u_{xx}+2x\,u_{xy}+y\,u_{yy}=0$. Which of the following are true? (One or more options may be correct.) (A) It is hyperbolic at $(1,0)$. (B) It is elliptic at $(1,2)$. (C) It is parabolic at every point of the line $y=-x$ other than the origin. (D) Near $(-1,1)$ there is a nonsingular $C^2$ change of variables transforming it into $u_{\xi\eta}+(\text{first-order terms})=0$.
E5. CSIR Part C Which of the following statements are true? (A) Every $C^2$ solution of $u_{xx}-4u_{yy}=0$ on $\mathbb{R}^2$ is of the form $f(y+2x)+g(y-2x)$. (B) $z=\varphi_1(y-x)+x\,\varphi_2(y-x)$, with $\varphi_1,\varphi_2$ arbitrary, is the general solution of $(D+D')^2z=0$. (C) $z=e^{x+2y}$ satisfies $(D^2-D')z=0$. (D) There are constants $m_1,m_2$ such that $\varphi_1(y+m_1x)+\varphi_2(y+m_2x)$ solves $(D^2-D')z=0$ for all smooth $\varphi_1,\varphi_2$.
E6. CSIR Part C For the equation $x\,u_{xx}+u_{yy}=0$, which of the following are true? (A) It is hyperbolic in the half-plane $x\lt0$. (B) In $x\lt0$ its characteristics are the curves $y\pm2\sqrt{-x}=\text{const}$. (C) In $x\gt0$ the substitution $\xi=y$, $\eta=2\sqrt x$ transforms it into $u_{\xi\xi}+u_{\eta\eta}-\tfrac1\eta u_\eta=0$. (D) It is parabolic on the line $y=0$.
Suggested plan — 3 lectures × 60 min.
- Type. Start from the conic analogy and the three model equations. Definitions 4.1–4.2, then Theorem 4.1 with the matrix proof $\bar M=PMP^{\mathsf T}$ (ten minutes; the chain-rule formulas can be handed out). Example 1 on the board; project Interactive 1 on the Tricomi and "circle" equations and let the class predict the colours before switching. Finish with Example 5(a)–(c).
- Characteristics and canonical forms. Theorem 4.2 and the sign warning; Theorem 4.3 via the $3\times3$ determinant — this is the conceptual heart, so draw the curve and the three unknowns $r,s,t$. Example 2, then Theorem 4.4 (do the hyperbolic and parabolic proofs fully, the elliptic one as a sketch), the strategy box and Examples 3 and 4(a). Theorem 4.5 in five minutes with Exercise II.7.
- Constant coefficients. Theorem 4.6 (the substitution $s=x$, $w=y+mx$ is everything), Theorems 4.7–4.8 with the repeated-factor case, Theorem 4.9 for the heat operator, Theorem 4.10 and Examples 6–7. Project Interactive 2: wave, then the repeated factor, then Laplace — the blow-up of the Cauchy problem always provokes questions. End with E3 and E5 done live.
Discussion questions. (1) Why should a smooth change of coordinates — which only relabels points — be unable to change the type, while a change of the unknown such as $u=e^{\alpha x}v$ can change the lower-order terms? (2) The heat equation has characteristics $t=\text{const}$. What does Theorem 4.3 then say about prescribing $u_t(x,0)$ for it? (3) Interactive 2 shows the Cauchy problem for Laplace's equation exploding. Which data would be "right" for an elliptic equation, and why?
Misconceptions to watch for. Using $B^2-AC$ with the full coefficient of $u_{xy}$; giving a single type to a variable-coefficient equation; calling points where $A=B=C=0$ parabolic; sign errors in $A(y')^2-By'+C=0$; believing every hyperbolic equation has the solution $f(\xi)+g(\eta)$; forgetting the factor $x$ for repeated roots; dividing by $F(a,b)=0$; thinking a PI is unique.
Suggested homework. I(1, 2, 4, 8, 9), II(1, 2, 3, 6, 8), III(2, 3, 5, 6, 8, 10), IV(3, 4, 5); Exam corner E1, E4, E6.
Research question. The Black–Scholes equation $V_t+\tfrac12\sigma^2S^2V_{SS}+rSV_S-rV=0$ prices options on every stock exchange in the world. What is its type, and can the ideas of this section turn it into the plain heat equation $u_\tau=u_{xx}$ — so that pricing an option becomes a heat-conduction problem?
Why it matters. Fischer Black, Myron Scholes and Robert Merton derived the equation in 1973 (Scholes and Merton received the 1997 Nobel Memorial Prize in Economics). Its reduction to the heat equation is what makes closed-form prices possible, and it explains why numerical schemes for option pricing are the same as those for heat flow: explicit schemes need a stability restriction, implicit ones (Crank–Nicolson) do not.
What to do.
- Classify the equation in the half-plane $S\gt0$ and find its characteristics. Which data does Theorem 4.3 suggest — a value at the expiry time $t=T$, or a value and a time derivative?
- Substitute $S=e^x$ and $\tau=\tfrac12\sigma^2(T-t)$ and show that the equation becomes $V_\tau=V_{xx}+(k-1)V_x-kV$ with $k=2r/\sigma^2$.
- Use the exponential substitution of Theorem 4.5(2), $V=e^{\alpha x+\beta\tau}u$, and find $\alpha,\beta$ that leave $u_\tau=u_{xx}$.
- For a call option ($V=\max(S-K,0)$ at $t=T$) with $S=K=100$, $r=0.05$, $\sigma=0.2$, $T=1$, compute the price from the closed-form Black–Scholes formula, and then by solving $u_\tau=u_{xx}$ numerically with an explicit finite-difference scheme on a truncated $x$-interval. How small must the time step be compared with $\Delta x^2$ for the scheme to be stable?
Tools: Python with NumPy/SciPy (or a spreadsheet for step 4); a computer algebra system is helpful for steps 2–3.
Hand in: A two-page note with the classification, the two substitutions, your values of $\alpha$ and $\beta$, and a table comparing the numerical and closed-form prices for three grid sizes.
Going further. Read about the Heston model, in which the volatility is itself random: its pricing equation has two space variables $S$ and $v$ and a mixed term $\rho\sigma vS\,V_{Sv}$. Classify it with the matrix test of this section — when is it degenerate?
🧪 Self-check quiz
Q1. The equation $u_{xx}+4u_{xy}+4u_{yy}+u_x=0$ is
Q2. Under a $C^2$ change of variables with Jacobian $J\ne0$, the discriminant $B^2-4AC$
Q3. At a point where an equation is elliptic, the number of real characteristic directions is
Q4. The general solution of $(D-2D')^2z=0$ is
Q5. A particular integral of $(D^2-D'^2)z=e^{2x+y}$ is