Chapter 3 · First-Order Partial Differential Equations

3.1 Lagrange’s Method for Quasi-Linear Equations

First-order partial differential equations: how they arise by eliminating constants and functions, how they are classified, and Lagrange's method for the quasi-linear equation $Pp+Qq=R$ — characteristic curves, first integrals, the general integral $\varphi(u,v)=0$, integral surfaces through a curve and orthogonal surfaces.

A partial differential equation of the first order relates an unknown function $z(x,y)$ of two variables to its two first partial derivatives. Such equations are everywhere: the concentration of a pollutant carried by a river obeys $u_t+c\,u_x=-ku$; the travel time of a light or seismic wave obeys the eikonal equation $u_x^2+u_y^2=1/v^2$; the density of cars on a motorway, the action function of classical mechanics and the shape of a sand pile under steady pouring are all governed by first-order equations. Geometrically the equation prescribes, at every point of space, which tangent planes an integral surface $z=z(x,y)$ is allowed to have.

The great discovery of Lagrange (1770s) is that a quasi-linear first-order equation $P\,p+Q\,q=R$ is secretly a system of ordinary differential equations: its solution surfaces are woven out of the integral curves of the vector field $(P,Q,R)$, the characteristic curves $\tfrac{dx}{P}=\tfrac{dy}{Q}=\tfrac{dz}{R}$. Two independent first integrals $u=c_1$, $v=c_2$ of this system give every solution in the form $\varphi(u,v)=0$, where $\varphi$ is an arbitrary function — the role played by an arbitrary constant in an ODE is taken by an arbitrary function. This section develops that theory with full proofs. In CSIR NET it is examined directly: forming a PDE by eliminating constants or functions, finding the general solution of a Lagrange equation (often with clever multipliers), the integral surface through a given curve, surfaces orthogonal to a family, and "true or false" statements about uniqueness that hinge on the geometry of characteristics. Throughout, $z=z(x,y)$ and $p=z_x=\tfrac{\partial z}{\partial x}$, $q=z_y=\tfrac{\partial z}{\partial y}$.

First-Order PDEs and Their Classification

Let $D\subseteq\mathbb{R}^2$ be open and connected (a domain). We look for functions $z:D\to\mathbb{R}$ of class $C^1$; their graphs are surfaces in $\mathbb{R}^3$ with normal vector $(p,q,-1)$ at the point $(x,y,z(x,y))$.

📘 Definition 3.1 — First-order PDE, solution, integral surface

A first-order partial differential equation in two independent variables is a relation $$F(x,y,z,p,q)=0,$$ where $F$ is a given function on an open subset of $\mathbb{R}^5$ that depends on $p$ or $q$ (that is, $(F_p,F_q)\ne(0,0)$). A solution on a domain $D$ is a function $z\in C^1(D)$ such that $(x,y,z(x,y),z_x,z_y)$ lies in the domain of $F$ and $F\big(x,y,z(x,y),z_x(x,y),z_y(x,y)\big)=0$ for all $(x,y)\in D$. The graph of a solution is called an integral surface. A surface given implicitly by $\Phi(x,y,z)=0$ with $\Phi_z\ne0$ is an integral surface if the function $z(x,y)$ it defines (implicit function theorem) is a solution; then $p=-\Phi_x/\Phi_z$, $q=-\Phi_y/\Phi_z$.

📘 Definition 3.2 — Linear, semilinear, quasi-linear and nonlinear equations; order and degree

A first-order equation is called

  • linear if it has the form $a(x,y)\,p+b(x,y)\,q=c(x,y)\,z+d(x,y)$ — linear in $z,p,q$ together, with coefficients depending on $x,y$ only;
  • semilinear if it has the form $a(x,y)\,p+b(x,y)\,q=c(x,y,z)$ — the derivatives enter linearly with coefficients free of $z$, but $z$ may enter the right side in any way;
  • quasi-linear if it has the form $P(x,y,z)\,p+Q(x,y,z)\,q=R(x,y,z)$ — linear in $p,q$, with coefficients that may depend on $z$; this is Lagrange's equation;
  • nonlinear (fully nonlinear) otherwise, e.g. $pq=z$, $p^2+q^2=1$, $z=px+qy+pq$.

Thus linear $\subset$ semilinear $\subset$ quasi-linear $\subset$ all first-order equations. The order of a PDE is the order of the highest derivative that occurs; its degree is the power of the highest-order derivative once the equation has been made polynomial in the derivatives (when that is possible). So $p^2+q^2=1$ has order $1$ and degree $2$.

The classification matters because it decides the method. Linear and semilinear equations have characteristic projections in the $xy$-plane that do not depend on the solution; quasi-linear equations (Lagrange's method, this section) have characteristics that depend on $z$ but are still curves in $(x,y,z)$-space; fully nonlinear equations need Charpit's method and characteristic strips (§3.2). Superposition — "the sum of two solutions is a solution" — holds only for homogeneous linear equations ($d=0$).

✏️ Example 1 — Classify each equation
Classify as linear, semilinear (not linear), quasi-linear (not semilinear) or nonlinear, and give the order and degree: (a) $x\,p+y^2q=z+e^x$; (b) $y\,p-x\,q=z^2$; (c) $z\,p+q=0$; (d) $pq=z$; (e) $p^2+2pq+q^2=1$; (f) $\sqrt{1+p^2}=q$.
(a) Linear It is $a\,p+b\,q=c\,z+d$ with $a=x$, $b=y^2$, $c=1$, $d=e^x$. Order $1$, degree $1$.
(b) Semilinear The coefficients of $p,q$ depend on $x,y$ only, but the right side $z^2$ is not linear in $z$. Order $1$, degree $1$.
(c) Quasi-linear The coefficient of $p$ is $z$, so the equation is not semilinear; it is linear in $p,q$. This is the inviscid Burgers equation (with $y$ playing the role of time). Order $1$, degree $1$.
(d) Nonlinear The product $pq$ is not linear in $(p,q)$. Order $1$; it has total degree $2$ in $(p,q)$ but degree $1$ in each of $p$ and $q$ separately — for products of derivatives the word “degree” is ambiguous, and it is linearity, not degree, that decides the method.
(e) A disguised pair As written it is nonlinear of degree $2$. But it says $(p+q)^2=1$. If $z$ is a $C^1$ solution on a domain $D$, the continuous function $p+q$ takes only the values $\pm1$ on the connected set $D$, so it is constant: $z$ solves the linear equation $p+q=1$ or the linear equation $p+q=-1$.
(f) Nonlinear Rationalising gives $q^2=1+p^2$ together with $q\ge1$: order $1$, degree $2$.

Forming PDEs: Eliminating Arbitrary Constants and Functions

An ODE of order $n$ arises by eliminating $n$ constants from an $n$-parameter family of curves. For PDEs there are two mechanisms. A two-parameter family of surfaces $z=F(x,y,a,b)$ leads, after elimination of $a$ and $b$, to a first-order PDE (generally nonlinear). A family involving an arbitrary function $\varphi$ leads to a first-order PDE that is always quasi-linear. Reversing these processes is the business of the whole chapter.

📐 Theorem 3.1 — Eliminating two arbitrary constants
Let $F(x,y,a,b)$ be of class $C^2$ on an open set of $\mathbb{R}^4$ and suppose that at $(x_0,y_0,a_0,b_0)$ $$\det\begin{pmatrix}F_{xa}&F_{xb}\\F_{ya}&F_{yb}\end{pmatrix}\ne0 .$$ Put $p_0=F_x(x_0,y_0,a_0,b_0)$, $q_0=F_y(x_0,y_0,a_0,b_0)$. Then there are $C^1$ functions $A(x,y,p,q)$, $B(x,y,p,q)$ near $(x_0,y_0,p_0,q_0)$ such that every surface $z=F(x,y,a,b)$ with $(a,b)$ near $(a_0,b_0)$ satisfies, near $(x_0,y_0)$, the first-order PDE $$z=F\big(x,y,A(x,y,p,q),B(x,y,p,q)\big).$$

Proof. Consider the system $p=F_x(x,y,a,b)$, $q=F_y(x,y,a,b)$ as two equations for the two unknowns $a,b$. The Jacobian of the right-hand sides with respect to $(a,b)$ is the determinant in the hypothesis, so by the implicit function theorem there are a neighbourhood $N$ of $(x_0,y_0,p_0,q_0)$, a neighbourhood $M$ of $(a_0,b_0)$ and $C^1$ functions $A,B$ on $N$ such that, for $(x,y,p,q)\in N$ and $(a,b)\in M$, the equations hold exactly when $a=A(x,y,p,q)$, $b=B(x,y,p,q)$. Now fix $(a,b)\in M$ and let $z=F(x,y,a,b)$. For $(x,y)$ near $(x_0,y_0)$ its derivatives are $p=F_x(x,y,a,b)$, $q=F_y(x,y,a,b)$, so $(x,y,p,q)\in N$ (by continuity) and the uniqueness part of the implicit function theorem gives $A(x,y,p,q)=a$, $B(x,y,p,q)=b$. Hence $z=F(x,y,a,b)=F(x,y,A,B)$. ∎

In practice we do the elimination by algebra: differentiate $z=F$ with respect to $x$ and $y$ and eliminate $a,b$ from the three equations $z=F$, $p=F_x$, $q=F_y$. The theorem guarantees that this is possible when the constants are essential (the determinant is nonzero); in $z=(a+b)x+y$, for instance, $a$ and $b$ occur only through $a+b$, the determinant is $0$, and the family really has one constant. If the number of constants is larger than the number of independent variables, elimination needs second derivatives ($z=ax+by+c$ gives $z_{xx}=z_{xy}=z_{yy}=0$); if it is smaller, the eliminant is not unique ($z=ax+y$ satisfies both $q=1$ and $z=px+y$).

📐 Theorem 3.2 — Eliminating an arbitrary function gives a quasi-linear equation
Let $u,v$ be $C^1$ functions on an open set $\Omega\subseteq\mathbb{R}^3$, let $\varphi$ be a $C^1$ function of two variables with $(\varphi_u,\varphi_v)\ne(0,0)$, and let $z\in C^1(D)$ have its graph in $\Omega$ and satisfy $\varphi\big(u(x,y,z),v(x,y,z)\big)=0$ on $D$. Then $z$ is a solution of the quasi-linear equation $$P'\,p+Q'\,q=R',$$ $$P'=\frac{\partial(u,v)}{\partial(y,z)},\quad Q'=\frac{\partial(u,v)}{\partial(z,x)},\quad R'=\frac{\partial(u,v)}{\partial(x,y)},$$ where $\frac{\partial(u,v)}{\partial(y,z)}=u_yv_z-u_zv_y$ and so on, evaluated at $(x,y,z(x,y))$. In vector form, $(P',Q',R')=\nabla u\times\nabla v$.

Proof. Differentiate the identity $\varphi(u,v)=0$ with respect to $x$ and to $y$, remembering that $z$ depends on $x,y$ (chain rule): $$\varphi_u\,(u_x+u_zp)+\varphi_v\,(v_x+v_zp)=0,$$ $$\varphi_u\,(u_y+u_zq)+\varphi_v\,(v_y+v_zq)=0.$$ This is a homogeneous linear system for the vector $(\varphi_u,\varphi_v)$, which is not zero; hence its determinant vanishes: $$(u_x+u_zp)(v_y+v_zq)-(u_y+u_zq)(v_x+v_zp)=0 .$$ Expanding, the term $u_zv_z\,pq$ cancels and we get $$(u_xv_y-u_yv_x)+p\,(u_zv_y-u_yv_z)+q\,(u_xv_z-u_zv_x)=0,$$ that is, $R'-p\,P'-q\,Q'=0$. ∎

Note what the theorem says geometrically: the vector $\nabla u\times\nabla v$ is tangent to every level curve $\{u=c_1,\ v=c_2\}$, and the equation $P'p+Q'q=R'$ says that this vector lies in the tangent plane of the surface. A special case worth remembering: $z=f(u)$ with $u=u(x,y)$ is $\varphi(u,z)=0$ with $\varphi=z-f(u)$.

✏️ Example 2 — Eliminating constants
Form the partial differential equations satisfied by the families (a) $z=(x-a)^2+(y-b)^2$ (paraboloids of revolution with vertex on the plane $z=0$); (b) $z=ax+by+a^2+b^2$ (planes); (c) $(x-a)^2+(y-b)^2+z^2=1$ (unit spheres with centre in the plane $z=0$).
(a) $p=2(x-a)$, $q=2(y-b)$, so $(x-a)^2+(y-b)^2=\tfrac{p^2+q^2}{4}$ and $$p^2+q^2=4z.$$ The determinant of Theorem 3.1 is $\det\begin{pmatrix}-2&0\\0&-2\end{pmatrix}=4\ne0$: both constants are essential.
(b) $p=a$, $q=b$; substituting, $$z=px+qy+p^2+q^2,$$ an equation of Clairaut form $z=px+qy+f(p,q)$, the PDE analogue of Clairaut's ODE (§1.2). We will solve it in §3.2.
(c) Differentiating implicitly, $2(x-a)+2zp=0$ and $2(y-b)+2zq=0$, so $x-a=-zp$, $y-b=-zq$. Substituting, $z^2p^2+z^2q^2+z^2=1$: $$z^2\left(1+p^2+q^2\right)=1.$$ Geometrically $1/\sqrt{1+p^2+q^2}$ is the cosine of the angle between the normal and the $z$-axis, so the equation says that the normal line at height $z$ meets the plane $z=0$ at distance exactly $1$ — the defining property of a unit sphere centred on that plane.
Check In (a), $z=(x-1)^2+(y+2)^2$ gives $p^2+q^2=4(x-1)^2+4(y+2)^2=4z$. ✓
✏️ Example 3 — Eliminating arbitrary functions
Eliminate the arbitrary functions: (a) $z=e^{y}f(x-y)$; (b) $z=f(x^2+y^2)$; (c) $\varphi\big(x+y+z,\ x^2+y^2+z^2\big)=0$; (d) $z=f(x+ay)+g(x-ay)$, $a\ne0$ a constant.
(a) $p=e^yf'(x-y)$ and $q=e^yf(x-y)-e^yf'(x-y)=z-p$. Hence $$p+q=z.$$
(b) $p=2x\,f'$, $q=2y\,f'$, so $yp-xq=0$. Surfaces of revolution about the $z$-axis are exactly the solutions of $y\,p-x\,q=0$ (Exercise II.9).
(c) Theorem 3.2 with $u=x+y+z$, $v=x^2+y^2+z^2$: $\nabla u=(1,1,1)$, $\nabla v=(2x,2y,2z)$, $$\nabla u\times\nabla v=\big(2z-2y,\ 2x-2z,\ 2y-2x\big).$$ Dividing by $-2$: $$(y-z)\,p+(z-x)\,q=x-y.$$
(d) Two arbitrary functions cannot be eliminated with first derivatives: $z_x=f'+g'$, $z_y=a(f'-g')$, and two more differentiations give $z_{xx}=f''+g''$, $z_{yy}=a^2(f''+g'')$, so $z_{yy}=a^2z_{xx}$ — the second-order wave equation of Chapter 4. Rule of thumb: one arbitrary function $\to$ first order, two $\to$ second order.

Lagrange's Method for $Pp+Qq=R$

Write the quasi-linear equation as a dot product: $$(P,Q,R)\cdot(p,q,-1)=0 .$$ Since $(p,q,-1)$ is normal to the surface $z=z(x,y)$, the equation says exactly this: at every point of an integral surface the vector $V=(P,Q,R)$ is tangent to the surface. So an integral surface should be a union of curves that are everywhere tangent to $V$ — the integral curves of the vector field $V$.

📘 Definition 3.3 — Characteristic curves, Lagrange's auxiliary equations, first integrals

Let $P,Q,R$ be $C^1$ on an open set $\Omega\subseteq\mathbb{R}^3$, not all zero at any point. The characteristic curves of $Pp+Qq=R$ are the solution curves of the autonomous system $$\frac{dx}{dt}=P(x,y,z),\qquad\frac{dy}{dt}=Q(x,y,z),\qquad\frac{dz}{dt}=R(x,y,z),$$ written without the parameter as Lagrange's auxiliary (subsidiary) equations $$\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}.$$ Their projections onto the $xy$-plane are the projected (base) characteristics. A first integral of the system is a $C^1$ function $u$ on $\Omega$ with $\nabla u\ne0$ that is constant along every characteristic curve; equivalently (differentiate along a solution) $$P\,u_x+Q\,u_y+R\,u_z=0\quad\text{in }\Omega .$$ Two first integrals $u,v$ are independent if $\nabla u\times\nabla v\ne0$ in $\Omega$.

By the Picard–Lindelöf theorem for systems (§1.3) exactly one characteristic passes through each point of $\Omega$. Locally they are the curves $u=c_1$, $v=c_2$: the intersections of the level surfaces of two independent first integrals.

📐 Theorem 3.3 — Integral surfaces are woven from characteristics
Let $P,Q,R\in C^1(\Omega)$ and let $z\in C^1(D)$ have graph $S\subseteq\Omega$.
  1. If $z$ is a solution of $Pp+Qq=R$ and $(x_0,y_0,z_0)\in S$, then the characteristic curve through $(x_0,y_0,z_0)$ lies in $S$ for all parameter values $t$ near $0$ (as long as its projection stays in $D$).
  2. Conversely, if through every point of $S$ there passes a characteristic curve lying in $S$, then $z$ is a solution.

Proof. (1) The functions $\tilde P(x,y)=P(x,y,z(x,y))$ and $\tilde Q(x,y)=Q(x,y,z(x,y))$ are $C^1$ on $D$, hence locally Lipschitz. Let $(x(t),y(t))$ be the solution of the planar system $x'=\tilde P(x,y)$, $y'=\tilde Q(x,y)$ with $(x(0),y(0))=(x_0,y_0)$, and put $Z(t)=z(x(t),y(t))$. By the chain rule and the PDE, $$Z'(t)=p\,x'+q\,y'=p\,P+q\,Q=R\big(x(t),y(t),Z(t)\big).$$ So $(x(t),y(t),Z(t))$ solves the characteristic system with initial point $(x_0,y_0,z_0)$. Since $P,Q,R$ are $C^1$, that system has a unique solution with this initial point (Picard–Lindelöf), so this curve is the characteristic curve — and it lies on $S$ by construction.

(2) Let $(x_0,y_0,z_0)\in S$ and let $\gamma(t)$ be a characteristic in $S$ with $\gamma(0)=(x_0,y_0,z_0)$. Its tangent vector $\gamma'(0)=(P,Q,R)$ is tangent to $S$, and every tangent vector of the graph is orthogonal to its normal $(p,q,-1)$. Hence $Pp+Qq-R=0$ at $(x_0,y_0)$. ∎

📐 Theorem 3.4 — Lagrange's theorem: the general integral $\varphi(u,v)=0$
Let $P,Q,R\in C^1(\Omega)$ with $(P,Q,R)\ne0$ in $\Omega$, and let $u=c_1$, $v=c_2$ be two independent first integrals of $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$ in $\Omega$.
  1. If $\varphi$ is $C^1$ with $(\varphi_u,\varphi_v)\ne(0,0)$ and $z\in C^1(D)$ satisfies $\varphi(u,v)=0$ (with $u,v$ evaluated at $(x,y,z(x,y))$), then $z$ is a solution of $Pp+Qq=R$.
  2. Conversely, every solution $z\in C^1(D)$ with graph in $\Omega$ satisfies, near each point of a dense open subset of $D$, a relation $\varphi(u,v)=0$ of this kind (in fact $v=g(u)$ or $u=g(v)$ for a $C^1$ function $g$).
The relation $\varphi(u,v)=0$, with $\varphi$ arbitrary, is called the general integral (general solution) of Lagrange's equation.

Proof. Step 1: $\nabla u\times\nabla v$ is a multiple of $V=(P,Q,R)$. Since $u$ and $v$ are first integrals, $\nabla u\cdot V=0$ and $\nabla v\cdot V=0$. The vectors $\nabla u,\nabla v$ are linearly independent, so the vectors orthogonal to both form a line, spanned by $\nabla u\times\nabla v$; $V\ne0$ lies on this line. Hence $\nabla u\times\nabla v=\mu V$ with $\mu=(\nabla u\times\nabla v)\cdot V/|V|^2$ a continuous function that never vanishes.

Step 2: the key identity. For any $z\in C^1(D)$ with graph in $\Omega$ put $U(x,y)=u(x,y,z(x,y))$ and $W(x,y)=v(x,y,z(x,y))$. The computation in the proof of Theorem 3.2 shows $$\frac{\partial(U,W)}{\partial(x,y)}=(u_x+u_zp)(v_y+v_zq)-(u_y+u_zq)(v_x+v_zp)=R'-pP'-qQ',$$ where $(P',Q',R')=\nabla u\times\nabla v=\mu(P,Q,R)$. Therefore $$\frac{\partial(U,W)}{\partial(x,y)}=-\mu\,\big(Pp+Qq-R\big).$$ So $z$ is a solution if and only if the Jacobian of $U,W$ vanishes identically on $D$.

Step 3: part (1). If $\varphi(U,W)=0$ on $D$, differentiation gives $\varphi_u\nabla U+\varphi_v\nabla W=0$ with $(\varphi_u,\varphi_v)\ne0$, so $\nabla U$ and $\nabla W$ are linearly dependent and the Jacobian vanishes. By Step 2, $z$ is a solution.

Step 4: part (2). Let $z$ be a solution, so $\partial(U,W)/\partial(x,y)\equiv0$. First, $\nabla U$ and $\nabla W$ cannot both vanish on an open set $N$: otherwise $u$ and $v$ would be constant on the piece of $S$ over $N$, which would then lie in the curve $\{u=c_1,v=c_2\}$ — impossible for the graph of a function over an open set. So the set where $\nabla U\ne0$ or $\nabla W\ne0$ is open and dense. Near such a point, say with $U_x\ne0$, the map $(x,y)\mapsto(U(x,y),y)$ has Jacobian $U_x\ne0$ and is a local $C^1$ change of coordinates (inverse function theorem). Write $W$ in the new coordinates $(\xi,\eta)=(U,y)$; by the chain rule $W_y=W_\xi U_y+W_\eta$ and $W_x=W_\xi U_x$, so $$W_\eta=W_y-\frac{U_y}{U_x}W_x=\frac{U_xW_y-U_yW_x}{U_x}=0 .$$ Thus $W$ depends on $\xi=U$ alone: $W=g(U)$ with $g\in C^1$, i.e. $\varphi(u,v)=v-g(u)=0$. The cases $U_y\ne0$, $W_x\ne0$, $W_y\ne0$ are identical. ∎

The argument in Step 4 is the functional dependence theorem: two functions with identically vanishing Jacobian are, locally, functions of each other. In practice one writes the general integral in whichever explicit form is convenient, $v=f(u)$ or $u=f(v)$, remembering that each form may miss a few solutions (see the ⚠️ box).

Interactive 1 — Weaving an Integral Surface out of Characteristics

A curve $\Gamma$ (red) is given in space. From each of its points we follow the characteristic system $\dot x=P,\ \dot y=Q,\ \dot z=R$ forwards and backwards (blue). The curves sweep out the integral surface through $\Gamma$; green curves join points with equal parameter $t$. Press Play to grow the characteristics, and turn the view with the slider. The readouts follow the tip of the thick characteristic: its coordinates change, but the two first integrals $u$ and $v$ stay fixed.

parameter range—
tip (x, y, z)—
first integral u—
first integral v—
integral surface—
🧭 Problem-solving strategy — finding two independent first integrals
  1. Separate a pair. If two of the fractions $\frac{dx}{P},\frac{dy}{Q},\frac{dz}{R}$ involve only two variables (e.g. $\frac{dx}{x}=\frac{dy}{y}$), integrate that ODE directly.
  2. Use one integral in another pair. If $u=c_1$ is known, use it to eliminate a variable from another pair of fractions, integrate, and then substitute $c_1=u$ back so that the second integral is a function of $x,y,z$ only.
  3. Combine fractions. For any functions $\ell,m,n$: $$\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}=\frac{\ell\,dx+m\,dy+n\,dz}{\ell P+mQ+nR}.$$ Choose $\ell,m,n$ so that the new numerator is an exact differential $dw$ and the denominator is a simple multiple of $w$ or of another fraction.
  4. Multipliers. If $\ell P+mQ+nR=0$ and $\ell\,dx+m\,dy+n\,dz=dw$ is exact, then $w=c$ is a first integral (Theorem 3.5). Standard choices: $(1,1,1)$, $(x,y,z)$, $(1/x,1/y,1/z)$, $(y,x,0)$…
  5. Check independence and write the answer as $\varphi(u,v)=0$, or $v=f(u)$. Check one or two solutions by substituting back.
📐 Theorem 3.5 — The method of multipliers
Let $\ell,m,n$ be continuous on $\Omega$ with $\ell P+mQ+nR=0$ in $\Omega$, and suppose $\ell\,dx+m\,dy+n\,dz=dw$ for a $C^1$ function $w$ with $\nabla w\ne0$ (that is, $w_x=\ell$, $w_y=m$, $w_z=n$). Then $w=c$ is a first integral of $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$.

Proof. Along a characteristic $t\mapsto(x(t),y(t),z(t))$, $$\frac{d}{dt}w=w_xx'+w_yy'+w_zz'=\ell P+mQ+nR=0,$$ so $w$ is constant on it. ∎

✏️ Example 4 — Multipliers: $(y-z)p+(z-x)q=x-y$
Find the general integral of $(y-z)\,p+(z-x)\,q=x-y$, and describe the characteristic curves geometrically.
Multipliers $(1,1,1)$ $(y-z)+(z-x)+(x-y)=0$, and $dx+dy+dz=d(x+y+z)$. So $u=x+y+z$ is a first integral.
Multipliers $(x,y,z)$ $x(y-z)+y(z-x)+z(x-y)=0$, and $x\,dx+y\,dy+z\,dz=\tfrac12d(x^2+y^2+z^2)$. So $v=x^2+y^2+z^2$ is a first integral.
Independence $\nabla u\times\nabla v=(1,1,1)\times(2x,2y,2z)=2(z-y,\,x-z,\,y-x)=-2(P,Q,R)$, which is nonzero away from the line $x=y=z$.
Answer $$\varphi\big(x+y+z,\ x^2+y^2+z^2\big)=0,$$ $\varphi$ arbitrary — exactly the family of Example 3(c), as Theorem 3.4 predicts. The characteristics are the circles in which the planes $x+y+z=c_1$ meet the spheres $x^2+y^2+z^2=c_2$: circles centred on the line $x=y=z$ in planes perpendicular to it. Indeed $V=(y-z,z-x,x-y)=(x,y,z)\times(1,1,1)$, so the flow is a rotation about that line, and integral surfaces are the surfaces of revolution about the axis $x=y=z$ (Interactive 1, last case).

For linear equations with $c=d=0$ the third auxiliary equation reads $dz=0$, and the theory collapses to a statement in the plane.

📐 Theorem 3.6 — Homogeneous linear equations $a\,p+b\,q=0$
Let $a,b\in C^1(D)$ with $(a,b)\ne(0,0)$ on the domain $D$. A function $z\in C^1(D)$ solves $a(x,y)\,z_x+b(x,y)\,z_y=0$ if and only if $z$ is constant along every solution curve of $\frac{dx}{a}=\frac{dy}{b}$ in $D$. If $u(x,y)=c$ is a first integral of $\frac{dx}{a}=\frac{dy}{b}$ with $\nabla u\ne0$, then every solution is locally of the form $z=f(u(x,y))$ with $f\in C^1$, and every such function is a solution.

Proof. Along a solution $(x(t),y(t))$ of $x'=a$, $y'=b$ we have $\tfrac{d}{dt}z(x(t),y(t))=a\,z_x+b\,z_y$; this vanishes along all such curves iff the PDE holds at all points (each point lies on such a curve). If $z=f(u)$ then $a z_x+bz_y=f'(u)\,(a u_x+bu_y)=0$. Conversely, $u$ and $z$ are both constant on the curves $u=c$, so their Jacobian $u_xz_y-u_yz_x$ vanishes (the vectors $\nabla u$, $\nabla z$ are both orthogonal to $(a,b)\ne0$), and the functional dependence argument of Theorem 3.4, Step 4, gives $z=f(u)$ near any point with $\nabla u\ne0$. ∎

So for $a\,p+b\,q=0$ the level curves of every solution are the projected characteristics, and a solution is completely determined by its values on a curve that crosses each projected characteristic exactly once. Interactive 2 makes this visible.

Interactive 2 — Data Travel Along Projected Characteristics

For $a\,z_x+b\,z_y=0$ the solution is constant along the curves $\frac{dx}{a}=\frac{dy}{b}$ (thin grey). Values are prescribed on the red data curve; the colour map shows the solution these data determine (red = positive, blue = negative, white = not reached by any characteristic from the data). Press Play: a tracer leaves the data curve along its characteristic. On the right, its coordinates $x(t)$, $y(t)$ change while $z$ stays flat. Move the slider to start from another data point.

tracer (x, y)—
first integral u(x, y)—
z at the tracer—
general solution—

Integral Surfaces Through a Curve, and Orthogonal Surfaces

The general integral contains an arbitrary function. Just as an initial condition picks one solution of an ODE, a curve picks one integral surface: we look for the integral surface that contains a given curve $\Gamma$. The precise existence and uniqueness theorem — the Cauchy problem — is proved in §3.2; here is the working method, and an example of how it can fail.

🧭 Problem-solving strategy — the integral surface through a given curve
  1. Find two independent first integrals $u(x,y,z)=c_1$, $v(x,y,z)=c_2$.
  2. Parametrise the curve: $\Gamma:\ x=x_0(s),\ y=y_0(s),\ z=z_0(s)$ (or use two equations for $\Gamma$).
  3. Along $\Gamma$ compute $c_1(s)=u(x_0,y_0,z_0)$ and $c_2(s)=v(x_0,y_0,z_0)$, and eliminate $s$ to get one relation $\varphi(c_1,c_2)=0$.
  4. The surface is $\varphi(u,v)=0$: the union of the characteristics through the points of $\Gamma$. Solve for $z$ if possible and check the PDE and that $\Gamma$ lies on the surface.
  5. If $c_1$ and $c_2$ are both constant along $\Gamma$, then $\Gamma$ is itself a characteristic and there are infinitely many integral surfaces through it; if the projection of $\Gamma$ runs along a projected characteristic while $\Gamma$ is not a characteristic, there is no solution $z(x,y)$ near $\Gamma$ (Example 7).
✏️ Example 5 — $x^2p+y^2q=(x+y)z$ through the line $y=1$, $z=x$
Find the general integral of $x^2p+y^2q=(x+y)\,z$ and the integral surface containing the straight line $y=1$, $z=x$.
First integral $\dfrac{dx}{x^2}=\dfrac{dy}{y^2}$ integrates to $-\tfrac1x=-\tfrac1y+c$, i.e. $u=\dfrac1y-\dfrac1x=\dfrac{x-y}{xy}=c_1$.
Second integral Combine the first two fractions with multipliers $(1,-1)$: $$\frac{dx-dy}{x^2-y^2}=\frac{dz}{(x+y)z}.$$ Since $x^2-y^2=(x-y)(x+y)$, the factor $x+y$ cancels: $\dfrac{d(x-y)}{x-y}=\dfrac{dz}{z}$, so $v=\dfrac{x-y}{z}=c_2$.
General integral $\varphi\!\left(\dfrac{x-y}{xy},\ \dfrac{x-y}{z}\right)=0$, $\varphi$ arbitrary; for instance $z=(x-y)\,f\!\left(\dfrac{x-y}{xy}\right)$.
Through the line On $\Gamma$ ($y=1$, $z=x$): $c_1=\dfrac{x-1}{x}$ and $c_2=\dfrac{x-1}{x}$. So $c_1=c_2$ along $\Gamma$, and the surface is $\dfrac{x-y}{xy}=\dfrac{x-y}{z}$. Away from the vertical plane $x=y$ (which is the union of the characteristics with $c_1=c_2=0$ and is not a graph), this is $$z=xy.$$
Check $p=y$, $q=x$: $x^2y+y^2x=xy(x+y)=(x+y)z$ ✓, and $y=1$ gives $z=x$ ✓. The projected line $y=1$ is never tangent to a projected characteristic: there $(P,Q)=(x^2,1)$ is never horizontal.

A classical geometric application: given a one-parameter family of surfaces $f(x,y,z)=c$, find the surfaces that cut every member of the family at right angles. Two surfaces meet orthogonally along a curve when their normals are orthogonal there.

📐 Theorem 3.7 — Surfaces orthogonal to a given family
Let $f\in C^1(\Omega)$ with $\nabla f\ne0$, so that the level sets $f=c$ form a one-parameter family of surfaces filling $\Omega$. A surface $z=z(x,y)$ with graph in $\Omega$ cuts every surface $f=c$ that it meets orthogonally if and only if $z$ solves the Lagrange equation $$f_x\,p+f_y\,q=f_z .$$ Its characteristics are the orthogonal trajectories of the family, the integral curves of $\nabla f$: $\dfrac{dx}{f_x}=\dfrac{dy}{f_y}=\dfrac{dz}{f_z}$.

Proof. At a common point the normal of $z=z(x,y)$ is $(p,q,-1)$ and the normal of the level surface through it is $\nabla f$. They are orthogonal iff $p\,f_x+q\,f_y-f_z=0$. ∎

✏️ Example 6 — Orthogonal surfaces of a family of ellipsoids
Find all surfaces orthogonal to the ellipsoids $x^2+y^2+2z^2=c$ $(c\gt0)$, and the one that contains the circle $x^2+y^2=1$, $z=1$.
The PDE $\nabla f=(2x,2y,4z)$, so Theorem 3.7 gives $2xp+2yq=4z$, i.e. $$x\,p+y\,q=2z .$$
First integrals $\dfrac{dx}{x}=\dfrac{dy}{y}$ gives $u=\dfrac yx$; $\dfrac{dx}{x}=\dfrac{dz}{2z}$ gives $\log z=2\log x+\text{const}$, $v=\dfrac{z}{x^2}$. General integral: $z=x^2f(y/x)$ — the functions homogeneous of degree $2$ (Euler's theorem in reverse).
Through the circle On $\Gamma$: $z=1$, so $c_2=\dfrac1{x^2}$, and $1+c_1^2=1+\dfrac{y^2}{x^2}=\dfrac{x^2+y^2}{x^2}=\dfrac1{x^2}$. Hence $c_2=1+c_1^2$ on $\Gamma$, and the surface is $\dfrac{z}{x^2}=1+\dfrac{y^2}{x^2}$: $$z=x^2+y^2 .$$
Check $xp+yq=2x^2+2y^2=2z$ ✓; on the circle $z=1$ ✓. Directly: the paraboloid's normal $(2x,2y,-1)$ and the ellipsoid's normal $(2x,2y,4z)$ have dot product $4x^2+4y^2-4z=0$ on the paraboloid ✓. The orthogonal trajectories $y=c_1x$, $z=c_2x^2$ are parabolas in vertical planes through the $z$-axis.
✏️ Example 7 — Prove or disprove: "there is exactly one integral surface through every curve"
For the equation $p+q=1$, decide whether there is a solution $z(x,y)$, defined near the projection of the curve, whose graph contains (a) $\Gamma_1:\ x=s,\ y=-s,\ z=0$; (b) $\Gamma_2:\ x=y=z=s$; (c) $\Gamma_3:\ x=y=s,\ z=0$. When it exists, is it unique?
General integral $\dfrac{dx}1=\dfrac{dy}1=\dfrac{dz}1$ gives $u=x-y$, $v=z-x$, so $z=x+f(x-y)$, $f\in C^1$ arbitrary (Theorem 3.4; here $u=x-y$ is a function of $(x,y)$ with $\nabla u\ne0$ everywhere, so this form is global on any domain on which the lines $x-y=c$ meet in intervals). The projected characteristics are the lines $x-y=c$; the characteristics are the lines of direction $(1,1,1)$.
(a) Exactly one On $\Gamma_1$: $0=s+f(2s)$, so $f(w)=-w/2$ and $z=x-\tfrac{x-y}{2}=\dfrac{x+y}{2}$. The projected curve $y=-x$ crosses every line $x-y=c$ exactly once, so $f$ is determined everywhere: the solution exists and is unique.
(b) Infinitely many $\Gamma_2$ has direction $(1,1,1)$: it is a characteristic. The condition $s=s+f(0)$ only says $f(0)=0$. Every $z=x+f(x-y)$ with $f(0)=0$ contains $\Gamma_2$ — for example $z=x$, $z=y$, $z=x+\sin(x-y)$.
(c) None $\Gamma_3$ projects onto the projected characteristic $x=y$ but is not a characteristic ($z$ does not increase with $x$ along it). A solution would need $0=s+f(0)$ for all $s$ — impossible. Directly: along the line $x=y$ every solution satisfies $\tfrac{d}{ds}z(s,s)=p+q=1$, while the data have slope $0$.
Moral Existence and uniqueness hold when the projected data curve is transversal to the projected characteristics; when it runs along one, there are either infinitely many solutions (data compatible: $\Gamma$ is a characteristic) or none. This is the Cauchy problem of §3.2.
⚠️ Counterexamples and pitfalls
  1. Quasi-linear is not linear. $z=\dfrac{x}{1+y}$ solves $z\,p+q=0$, but $2z$ does not: $2z\cdot\tfrac{2}{1+y}-\tfrac{2x}{(1+y)^2}=\tfrac{2x}{(1+y)^2}\ne0$. Superposition holds only for homogeneous linear equations.
  2. Two integrals must be independent. For $(y-z)p+(z-x)q=x-y$, both $x+y+z$ and $(x+y+z)^3$ are first integrals, but $\varphi\big(x+y+z,(x+y+z)^3\big)=0$ only gives the planes $x+y+z=c$. You need $\nabla u\times\nabla v\ne0$.
  3. $xyz$ is not a first integral of every "symmetric" equation. For $(y-z)p+(z-x)q=x-y$, $V\cdot\nabla(xyz)=(x-y)(x-z)(y-z)\ne0$. Always verify $Pu_x+Qu_y+Ru_z=0$.
  4. Fewer constants than variables: the eliminant is not unique. $z=ax+y$ satisfies both $q=1$ and $z=px+y$.
  5. More constants than variables: higher order. No nontrivial first-order equation is satisfied by every plane $z=ax+by+c$: at a fixed point $(x,y)$ the triple $(z,p,q)=(ax+by+c,a,b)$ takes every value in $\mathbb{R}^3$. Elimination needs second derivatives: $z_{xx}=z_{xy}=z_{yy}=0$.
  6. Inessential constants. In $z=(a+b)x+y$ the constants combine; the family has one parameter, and $p=a+b$, $q=1$.
  7. "Nonlinear-looking" may be linear. $p^2+2pq+q^2=1$ is the pair of linear equations $p+q=\pm1$ (Example 1(e)).
  8. Not every curve carries exactly one integral surface. For $p+q=1$: a transversal curve carries one, a characteristic carries infinitely many, a non-characteristic curve over a projected characteristic carries none (Example 7).
  9. Closed characteristics impose symmetry. A $C^1$ solution of $y\,p-x\,q=0$ on all of $\mathbb{R}^2$ with $z(x,0)=g(x)$ exists iff $g$ is even, because each circle $x^2+y^2=r^2$ meets the $x$-axis at $\pm r$. So $z(x,0)=x$ has no solution on $\mathbb{R}^2$, although $z=\sqrt{x^2+y^2}$ solves it near any point $(a,0)$, $a\gt0$.
  10. The explicit form $v=f(u)$ can miss solutions. For $x\,p+y\,q=z$ with $u=z/x$, $v=y/x$, the planes $z=cx$ are solutions ($u=c$) but are not of the form $y/x=f(z/x)$. Use $\varphi(u,v)=0$, or solve for the right variable ($z=x\,f(y/x)$ contains all of them).
  11. A characteristic curve is not an integral surface. It is a curve; integral surfaces are one-parameter families of characteristics.

Exercises

I
Classification and formation of PDEs.
  • 1 Classify (linear / semilinear / quasi-linear / nonlinear) and give order and degree: (a) $x^2p+y\,q=z^3$; (b) $z\,p-y\,q=x$; (c) $p^2+q^2=z$; (d) $(\sin y)\,p+e^xq=xyz$; (e) $(p-q)^2=0$.
  • 2 Eliminate $a,b$ from $z=ax+by+ab$.
  • 3 Eliminate $a,b$ from $z=(x^2+a)(y^2+b)$.
  • 4 Eliminate $a,b$ from $(x-a)^2+(y-b)^2=z^2\cot^2\alpha$, where $0\lt\alpha\lt\pi/2$ is fixed (right circular cones with vertex on $z=0$ and semi-vertical angle $\tfrac{\pi}{2}-\alpha$).
  • 5 Eliminate $f$ from $z=f(x^2-y^2)$.
  • 6 Eliminate $f$ from $z=xy+f(x^2+y^2)$.
  • 7 Eliminate $\varphi$ from $\varphi\big(x^2+y^2+z^2,\ z^2-2xy\big)=0$.
  • 8 Eliminate $f$ and $g$ from $z=f(x+2y)+g(x-2y)$. Why can no first-order equation do this?
1 (a) semilinear (right side $z^3$), order 1, degree 1; (b) quasi-linear, not semilinear (coefficient $z$ of $p$), degree 1; (c) nonlinear, degree 2; (d) linear ($c=xy$, $d=0$), degree 1; (e) nonlinear as written (degree 2), but equivalent to the linear equation $p=q$.
2 $p=a$, $q=b$: $z=px+qy+pq$ (Clairaut form).
3 $p=2x(y^2+b)$, $q=2y(x^2+a)$, so $pq=4xy(x^2+a)(y^2+b)$: $$pq=4xyz.$$
4 Differentiating, $x-a=zp\cot^2\alpha$ and $y-b=zq\cot^2\alpha$. Substituting, $z^2\cot^4\alpha\,(p^2+q^2)=z^2\cot^2\alpha$, so $$p^2+q^2=\tan^2\alpha$$ — the surface makes the constant angle $\alpha$ with the horizontal plane, as a cone of this kind does.
5 $p=2xf'$, $q=-2yf'$: $y\,p+x\,q=0$.
6 $p=y+2xf'$, $q=x+2yf'$: $yp-xq=y^2-x^2$.
7 $u=x^2+y^2+z^2$, $v=z^2-2xy$: $\nabla u\times\nabla v=(2x,2y,2z)\times(-2y,-2x,2z)=\big(4z(x+y),\,-4z(x+y),\,4(y^2-x^2)\big)$. Dividing by $4(x+y)$: $$z\,(p-q)=y-x.$$
8 $z_{xx}=f''+g''$, $z_{yy}=4(f''+g'')$: $z_{yy}=4z_{xx}$. At a fixed point $(x,y)$ the four numbers $f(x+2y)$, $g(x-2y)$, $f'(x+2y)$, $g'(x-2y)$ can be chosen freely, so $(z,p,q)=\big(f+g,\ f'+g',\ 2(f'-g')\big)$ takes every value in $\mathbb{R}^3$; hence no nontrivial relation $F(x,y,z,p,q)=0$ holds for every member of the family.
II
General integrals by Lagrange's method.
  • 1 $2p+3q=1$.
  • 2 $x\,p+y\,q=nz$ ($n$ a real constant). Interpret the answer.
  • 3 $y\,p+x\,q=z$.
  • 4 $xz\,p+yz\,q=xy$.
  • 5 $(y+z)\,p-(z+x)\,q=y-x$ (find suitable multipliers).
  • 6 $p+q=\dfrac{z}{x+y}$ in the region $x+y\gt0$.
  • 7 $x\,p+y\,q=z+x^2+y^2$.
  • 8 $\dfrac{p}{x}+\dfrac{q}{y}=\dfrac{1}{z}$ in the region $x,y,z\gt0$.
  • 9 Show that the solutions of $y\,p-x\,q=0$ on an annulus $0\lt a\lt x^2+y^2\lt b$ are exactly the functions $z=f(x^2+y^2)$, i.e. the surfaces of revolution about the $z$-axis.
1 $\tfrac{dx}2=\tfrac{dy}3$: $3x-2y=c_1$; $\tfrac{dx}2=\tfrac{dz}1$: $x-2z=c_2$. General integral $\varphi(3x-2y,\,x-2z)=0$, i.e. $z=\tfrac x2+f(3x-2y)$.
2 $\tfrac{dx}x=\tfrac{dy}y$: $y/x=c_1$; $\tfrac{dx}x=\tfrac{dz}{nz}$: $z/x^n=c_2$. So $z=x^nf(y/x)$ (for $x\gt0$, say): the solutions are exactly the functions homogeneous of degree $n$ — this is the converse of Euler's theorem on homogeneous functions.
3 $\tfrac{dx}y=\tfrac{dy}x$: $x^2-y^2=c_1$. Adding numerators and denominators, $\tfrac{dx+dy}{x+y}=\tfrac{dz}z$: $\tfrac{x+y}{z}=c_2$. Answer $\varphi\big(x^2-y^2,\tfrac{x+y}{z}\big)=0$.
4 $\tfrac{dx}{xz}=\tfrac{dy}{yz}$: $y/x=c_1$. Then $\tfrac{dx}{xz}=\tfrac{dz}{xy}$ with $y=c_1x$ gives $c_1x\,dx=z\,dz$, so $c_1x^2-z^2=c_2$; substituting $c_1=y/x$: $xy-z^2=c_2$. Answer $\varphi\big(\tfrac yx,\ xy-z^2\big)=0$. (Check: $V\cdot\nabla(xy-z^2)=xz\cdot y+yz\cdot x-xy\cdot2z=0$.)
5 Multipliers $(x,y,z)$: $x(y+z)-y(z+x)+z(y-x)=0$, giving $x^2+y^2+z^2=c_1$. Multipliers $(1,1,-1)$: $(y+z)-(z+x)-(y-x)=0$, giving $x+y-z=c_2$. Answer $\varphi\big(x^2+y^2+z^2,\ x+y-z\big)=0$ — surfaces of revolution about the axis through the origin in direction $(1,1,-1)$.
6 $dx=dy$: $x-y=c_1$. Then $\tfrac{d(x+y)}{2}=dx=\tfrac{(x+y)\,dz}{z}$, so $\tfrac{dz}{z}=\tfrac{d(x+y)}{2(x+y)}$ and $\tfrac{z^2}{x+y}=c_2$. Answer $\varphi\big(x-y,\ \tfrac{z^2}{x+y}\big)=0$, e.g. $z^2=(x+y)\,f(x-y)$.
7 $y=c_1x$. Along it $\tfrac{dz}{dx}=\tfrac{z+x^2(1+c_1^2)}{x}$, a linear ODE: $\tfrac{d}{dx}\big(\tfrac zx\big)=1+c_1^2$, so $\tfrac zx=x(1+c_1^2)+c_2$, i.e. $\tfrac{z-x^2-y^2}{x}=c_2$. Answer $$z=x^2+y^2+x\,f(y/x).$$ Check: $x^2+y^2$ contributes $2(x^2+y^2)$ to $xp+yq$, and $xf(y/x)$, homogeneous of degree 1, contributes itself.
8 Multiply by $xyz$: $yz\,p+xz\,q=xy$. $\tfrac{dx}{yz}=\tfrac{dy}{xz}$: $x^2-y^2=c_1$; $\tfrac{dx}{yz}=\tfrac{dz}{xy}$: $x^2-z^2=c_2$. Answer $\varphi(x^2-y^2,\ x^2-z^2)=0$.
9 By Theorem 3.6 a solution is constant along the projected characteristics $\tfrac{dx}{y}=\tfrac{dy}{-x}$, the circles $x^2+y^2=r^2$, each of which lies entirely in the annulus. So $z(x,y)=f(r^2)$ with $f(w)=z(\sqrt w,0)$, which is $C^1$ for $a\lt w\lt b$. Conversely $z=f(x^2+y^2)$ gives $yp-xq=2xyf'-2xyf'=0$.
III
Integral surfaces through curves; orthogonal surfaces.
  • 1 Find the integral surface of $x\,p+y\,q=z$ that contains the circle $x^2+y^2=1$, $z=2$.
  • 2 Find the integral surface of $y\,p-x\,q=0$ containing the curve $x=s$, $y=0$, $z=s^2$ $(s\gt0)$.
  • 3 Find the integral surface of $p+q=2z$ containing the line $x=s$, $y=-s$, $z=1$.
  • 4 Find the integral surface of $x\,p+y\,q=2z$ containing the parabola $y=1$, $z=x^2+1$.
  • 5 Find the surfaces orthogonal to the family $xyz=c$, and the one containing the circle $x=0$, $y^2+z^2=1$.
  • 6 Show that the surfaces orthogonal to the spheres $x^2+y^2+z^2=cx$ satisfy $(x^2-y^2-z^2)\,p+2xy\,q=2xz$, and find them.
  • 7 For $x\,p+y\,q=z$, decide whether there is a solution near $(1,1)$ whose graph contains (a) the curve $y=x$, $z=1$; (b) the curve $y=x$, $z=2x$. If so, how many?
  • 8 Find the integral surface of $(y-z)\,p+(z-x)\,q=x-y$ containing the circle $x^2+y^2=1$, $z=1$.
1 $u=y/x$, $v=z/x$. On the circle $v=2/x$ and $1+u^2=1/x^2$, so $v^2=4(1+u^2)$: $\tfrac{z^2}{x^2}=4\tfrac{x^2+y^2}{x^2}$, $$z=2\sqrt{x^2+y^2}$$ (the branch with $z=2\gt0$ on $\Gamma$) — a cone.
2 $u=x^2+y^2$, $v=z$; on $\Gamma$, $u=s^2=v$. So $z=x^2+y^2$.
3 $u=x-y$, and $\tfrac{dz}{dx}=2z$ gives $v=ze^{-2x}$. On $\Gamma$: $u=2s$, $v=e^{-2s}=e^{-u}$. So $ze^{-2x}=e^{-(x-y)}$: $$z=e^{x+y}.$$ Check: $p+q=2e^{x+y}=2z$, and $z=1$ when $y=-x$.
4 $u=y/x$, $v=z/x^2$. On $\Gamma$: $x=1/u$, $v=\tfrac{x^2+1}{x^2}=1+u^2$. So $z=x^2+y^2$.
5 $\nabla f=(yz,xz,xy)$: $yz\,p+xz\,q=xy$, with first integrals $x^2-y^2$ and $x^2-z^2$ (Exercise II.8): orthogonal surfaces $\varphi(x^2-y^2,\,x^2-z^2)=0$. On $\Gamma$: $c_1=-y^2$, $c_2=-z^2$, so $c_1+c_2=-1$: $$y^2+z^2-2x^2=1,$$ a hyperboloid of one sheet. Check: $(-4x,2y,2z)\cdot(yz,xz,xy)=-4xyz+2xyz+2xyz=0$ ✓.
6 $f=(x^2+y^2+z^2)/x$, $\nabla f=\tfrac1{x^2}\big(x^2-y^2-z^2,\,2xy,\,2xz\big)$; Theorem 3.7 gives the stated equation. $\tfrac{dy}{2xy}=\tfrac{dz}{2xz}$: $y/z=c_1$. With multipliers $(x,y,z)$: $\tfrac{x\,dx+y\,dy+z\,dz}{x(x^2+y^2+z^2)}=\tfrac{dz}{2xz}$, so $\tfrac{d(x^2+y^2+z^2)}{x^2+y^2+z^2}=\tfrac{dz}{z}$ and $\tfrac{x^2+y^2+z^2}{z}=c_2$. Orthogonal surfaces: $\varphi\big(\tfrac yz,\ \tfrac{x^2+y^2+z^2}{z}\big)=0$ (they include the spheres $x^2+y^2+z^2=c\,z$).
7 The projected characteristics are rays from the origin, and $y=x$ is one of them. (a) $\Gamma=(s,s,1)$ has direction $(1,1,0)$, not parallel to $V=(s,s,1)$: not a characteristic. Every solution has $z/x$ constant on the ray (Theorem 3.3), but $z/x=1/s$ varies: no solution. (b) $\Gamma=(s,s,2s)$ is a characteristic; every $z=x\,f(y/x)$ with $f(1)=2$ contains it: infinitely many, e.g. $2x$, $x+y$, $(x^2+y^2)/x$.
8 $u=x+y+z$, $v=x^2+y^2+z^2$ (Example 4). On $\Gamma$: $v=1+1=2$ for every point, while $u$ varies. The relation is $v=2$: the sphere $$x^2+y^2+z^2=2,$$ i.e. $z=\sqrt{2-x^2-y^2}$ near $\Gamma$. Check: $p=-x/z$, $q=-y/z$, $(y-z)p+(z-x)q=\tfrac{-xy+xz-yz+xy}{z}=x-y$ ✓.
IV
Proofs and challenges.
  • 1 Prove: if $u$ and $v$ are first integrals of $\tfrac{dx}{P}=\tfrac{dy}{Q}=\tfrac{dz}{R}$ and $\Phi$ is $C^1$ with $\nabla\big(\Phi(u,v)\big)\ne0$, then $\Phi(u,v)$ is a first integral. Must three first integrals always be functionally dependent?
  • 2 Prove that the surfaces generated by straight lines through a fixed point $(a,b,c)$ (cones with vertex $(a,b,c)$) are exactly the solutions of $(x-a)\,p+(y-b)\,q=z-c$.
  • 3 Prove that the cylinders whose generators are parallel to a fixed vector $(\ell,m,n)$, $\ell\ne0$, are the solutions of $\ell p+mq=n$, and find the general integral.
  • 4 Show that $(y-z)\,p+(z-x)\,q=x-y$ has the integral surface $x^2+y^2+z^2=1$ containing the circle $\Gamma:\ x^2+y^2=1$, $z=0$, yet there is no solution $z(x,y)$ defined on a neighbourhood of any point of the unit circle in the $xy$-plane whose graph contains the nearby part of $\Gamma$. Explain the failure using the projected characteristic directions along $\Gamma$.
  • 5 Let $g\in C^1(\mathbb{R})$. Prove that $y\,p-x\,q=0$ has a $C^1$ solution on all of $\mathbb{R}^2$ with $z(x,0)=g(x)$ for all $x$ if and only if $g$ is even.
  • 6 Prove that $x\,p+y\,q=z$ has no solution on all of $\mathbb{R}^2$ other than the linear functions $z=ax+by$. (Hint: a solution is homogeneous of degree $1$ and differentiable at $0$.)
  • 7 (Applied.) A pollutant of concentration $c(x,t)$ is carried by a canal with speed $v=0.5\ \text{m/s}$ and decays at rate $k=0.1\ \text{h}^{-1}$: $c_t+v\,c_x=-kc$. At $t=0$ the concentration is $c_0(x)=8e^{-(x/200)^2}\ \text{mg/L}$ ($x$ in metres). Find $c(x,t)$ and the peak concentration and its position after $2$ hours.
1 $V\cdot\nabla\Phi(u,v)=\Phi_u(V\cdot\nabla u)+\Phi_v(V\cdot\nabla v)=0$. Yes: at each point all gradients of first integrals are orthogonal to $V\ne0$, so they lie in a plane; any three are linearly dependent, their Jacobian $\partial(u,v,w)/\partial(x,y,z)$ vanishes, and (locally, when $u,v$ are independent) $w=\Phi(u,v)$.
2 The characteristics of $(x-a)p+(y-b)q=z-c$ satisfy $\tfrac{dx}{x-a}=\tfrac{dy}{y-b}=\tfrac{dz}{z-c}$: $x-a=Ae^t$, $y-b=Be^t$, $z-c=Ce^t$ — the half-lines from $(a,b,c)$. By Theorem 3.3 a surface is a solution iff it is a union of such half-lines, i.e. a cone with vertex $(a,b,c)$. The general integral is $\varphi\big(\tfrac{y-b}{x-a},\tfrac{z-c}{x-a}\big)=0$.
3 Characteristics are the lines of direction $(\ell,m,n)$, so solutions are unions of such lines (Theorem 3.3), i.e. cylinders. First integrals: $mx-\ell y$, $nx-\ell z$ (check: $\ell m-m\ell=0$, $\ell n-n\ell=0$). General integral $\varphi(mx-\ell y,\ nx-\ell z)=0$, i.e. $z=\tfrac{n}{\ell}x+f(mx-\ell y)$.
4 The sphere is an integral surface (Exercise III.8 with $v=1$) and contains $\Gamma$. Along $\Gamma=(\cos s,\sin s,0)$ the projected characteristic direction is $(P,Q)=(\sin s,-\cos s)$, which is tangent to the unit circle at every point, while $V=(\sin s,-\cos s,\cos s-\sin s)$ is not tangent to $\Gamma$ (its $z$-component is nonzero except where $\cos s=\sin s$). So the projected data curve runs along projected characteristics without $\Gamma$ being characteristic — the situation of Example 7(c). Concretely: suppose $z(x,y)$ is a solution near a point of the circle with $z(\cos s,\sin s)=0$ for $s$ in an interval. Differentiating in $s$: $-\sin s\,p+\cos s\,q=0$ there. But at these points $z=0$, and the PDE reads $\sin s\,p-\cos s\,q=\cos s-\sin s$. Together they force $\cos s=\sin s$ on a whole interval — impossible. (On the sphere itself, the tangent plane is vertical along $\Gamma$: the sphere is not a graph over a neighbourhood of the circle.)
5 If $z$ is a solution on $\mathbb{R}^2$, it is constant on each circle (Theorem 3.6), and the circle of radius $|x|$ contains $(x,0)$ and $(-x,0)$: $g(x)=g(-x)$. Conversely, if $g$ is even put $z(x,y)=g\big(\sqrt{x^2+y^2}\big)$. Away from $0$ it is $C^1$ and solves the equation. At $0$: $g'(0)=0$ ($g'$ is odd), so $z(x,y)-g(0)=g(r)-g(0)=o(r)$ and $z$ is differentiable at $0$ with zero gradient; and $\nabla z=g'(r)(x,y)/r\to0$ as $r\to0$. So $z\in C^1(\mathbb{R}^2)$.
6 On each open ray from $0$ the function $z/r$ is constant (the projected characteristics are rays; $z/x$, or $z/y$, is a first integral). So $z(\lambda x,\lambda y)=\lambda z(x,y)$ for $\lambda\gt0$. Differentiability at $0$ (and $z(0,0)=0$ by continuity, letting $\lambda\to0$) gives $z(\lambda x,\lambda y)=\lambda\big(ax+by\big)+o(\lambda)$ with $(a,b)=\nabla z(0,0)$; dividing by $\lambda$ and letting $\lambda\to0^+$: $z(x,y)=ax+by$. ($z=\sqrt{x^2+y^2}$ solves the equation away from $0$ only.)
7 Lagrange: $\tfrac{dt}{1}=\tfrac{dx}{v}=\tfrac{dc}{-kc}$ gives $x-vt=c_1$, $ce^{kt}=c_2$, so $c=e^{-kt}c_0(x-vt)$. In consistent units ($t$ in hours, $v=0.5\cdot3600=1800\ \text{m/h}$): after $2$ h the peak is at $x=3600\ \text{m}$ with value $8e^{-0.2}\approx6.55\ \text{mg/L}$. The shape (width $200\ \text{m}$) is unchanged: pure advection does not spread.
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

E1. CSIR Part B GATE MA The partial differential equation obtained by eliminating the arbitrary function $f$ from $z=e^{x}f(x-y)$ is (A) $p-q=z$; (B) $p+q=z$; (C) $p+q=0$; (D) $pq=z$.

Idea Differentiate and eliminate $f'$.
Computation $p=e^xf+e^xf'=z+e^xf'$ and $q=-e^xf'$. Adding, $p+q=z$.
Answer (B). (Check with $f\equiv1$: $z=e^x$ gives $p=e^x$, $q=0$.)

E2. CSIR Part B Let $z(x,y)$ be the solution of $x\,p+y\,q=2z$ whose graph contains the curve $x=1$, $z=2y+3$. Then $z(2,3)$ equals (A) $18$; (B) $21$; (C) $24$; (D) $30$.

Idea First integrals $u=y/x$, $v=z/x^2$; relate them on the curve.
Relation On $x=1$: $u=y$, $v=2y+3=2u+3$. Hence $\tfrac{z}{x^2}=\tfrac{2y}{x}+3$, $z=2xy+3x^2$. Check: $xp+yq=x(2y+6x)+2xy=4xy+6x^2=2z$ ✓.
Answer $z(2,3)=12+12=24$: (C).

E3. CSIR Part B For the equation $(y-z)\,p+(z-x)\,q=x-y$, which of the following is not a first integral of the auxiliary equations? (A) $x+y+z$; (B) $x^2+y^2+z^2$; (C) $xy+yz+zx$; (D) $xyz$.

Idea $u$ is a first integral iff $Pu_x+Qu_y+Ru_z=0$; functions of first integrals are first integrals.
Options (A), (B): multipliers $(1,1,1)$ and $(x,y,z)$ (Example 4). (C) $xy+yz+zx=\tfrac12\big[(x+y+z)^2-(x^2+y^2+z^2)\big]$ is a function of (A) and (B). (D) $(y-z)yz+(z-x)xz+(x-y)xy=(x-y)(x-z)(y-z)$, not identically $0$.
Answer (D).

E4. CSIR Part C Consider $x\,p+y\,q=z$. Which of the following statements are true? (One or more options may be correct.) (A) $z=\sqrt{x^2+y^2}$ is a solution on $\mathbb{R}^2\setminus\{(0,0)\}$. (B) Every $C^1$ solution on the half-plane $x\gt0$ has the form $z=x\,f(y/x)$ for some $C^1$ function $f$. (C) There is a solution on $\mathbb{R}^2\setminus\{(0,0)\}$ equal to $1$ on the circle $x^2+y^2=1$. (D) There is a solution on the half-plane $x\gt0$ with $z(x,x)=1$ for all $x\gt0$.

(A) True $p=x/r$, $q=y/r$, so $xp+yq=r=z$.
(B) True In $x\gt0$ the projected characteristics are the open rays, each meeting the line $x=1$ once, and $z/x$ is constant on each (Theorem 3.3). So $z(x,y)/x=z(1,y/x)$; put $f(t)=z(1,t)$, which is $C^1$.
(C) True $z=\sqrt{x^2+y^2}$ from (A).
(D) False $y=x$ is a ray, so $z(x,x)/x$ must be constant; $1/x$ is not. (The curve $(s,s,1)$ projects onto a projected characteristic but is not a characteristic.)
Answer (A), (B), (C).

E5. CSIR Part C Which of the following correctly pair a family of surfaces ($a,b$ arbitrary constants) with a first-order PDE satisfied by every member of the family? (A) $z=ax+by+ab$ and $z=px+qy+pq$. (B) $z=(x-a)^2+(y-b)^2$ and $p^2+q^2=4z$. (C) $z=ax+b$ and $z=xp$. (D) $z=a(x+y)+b$ and $p=q$.

(A) True $p=a$, $q=b$; substitute.
(B) True $p=2(x-a)$, $q=2(y-b)$ (Example 2(a)).
(C) False $p=a$, so $xp=ax=z-b\ne z$ unless $b=0$. (For $z=ax+b$ the first-order equation satisfied is $q=0$; the constant $b$ cannot be eliminated with first derivatives because the family does not depend on $y$.)
(D) True $p=q=a$.
Answer (A), (B), (D).

E6. CSIR Part C GATE MA For the equation $y\,z_x-x\,z_y=0$ on $\mathbb{R}^2$, which of the following are true? (A) $z=x^2+y^2+1$ is a solution. (B) $z=xy$ is a solution. (C) Every $C^1$ solution on $\mathbb{R}^2$ is constant on each circle $x^2+y^2=r^2$. (D) There is a $C^1$ solution on $\mathbb{R}^2$ with $z(x,0)=x$ for all $x\in\mathbb{R}$.

(A) True $y\cdot2x-x\cdot2y=0$.
(B) False $y\cdot y-x\cdot x=y^2-x^2\not\equiv0$.
(C) True Theorem 3.6: the projected characteristics are the circles, and each lies in $\mathbb{R}^2$.
(D) False By (C), $z(x,0)=z(-x,0)$, so $g(x)=x$ would have to be even (Exercise IV.5).
Answer (A), (C).
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Classification and formation. Open with the river-pollutant equation $c_t+vc_x=-kc$ and the eikonal equation as motivation. Definition 3.2 with Example 1 (ask the class to classify before revealing; item (e) always surprises). Theorems 3.1–3.2 with Examples 2–3; emphasise the slogan "constants → nonlinear, functions → quasi-linear" and the Jacobian formula $(P',Q',R')=\nabla u\times\nabla v$. End with the question: "can we reverse Theorem 3.2?"
  2. Lagrange's method. The geometric reading $(P,Q,R)\cdot(p,q,-1)=0$; Theorem 3.3 with the ODE-uniqueness proof; project Interactive 1 (case 4 first: the rotating circles make "union of characteristics" unforgettable). Theorem 3.4 via the key identity $\partial(U,W)/\partial(x,y)=-\mu(Pp+Qq-R)$. The strategy box and Example 4 on the board; Theorem 3.6 with Interactive 2 (hyperbolas case: the left half-plane stays white). Finish with Examples 5 and 7 and E4.

Discussion questions. (1) Why does eliminating an arbitrary function always produce an equation that is linear in $p$ and $q$? (2) In Example 7, the data curve $\Gamma_3$ looks perfectly harmless. What exactly goes wrong, and how would you detect it from $P,Q$ and the parametrisation of $\Gamma$ alone? (3) For $y\,p-x\,q=0$, why does a local solution near $(1,0)$ with $z(x,0)=x$ exist although no global one does?

Misconceptions to watch for. Treating a single first integral as the general solution; using two dependent integrals; forgetting to substitute $c_1$ back after using it to find $c_2$; believing that superposition holds for quasi-linear equations; assuming that every curve determines a unique integral surface; writing $v=f(u)$ and missing the solutions $u=$ const; sign errors in the cyclic Jacobians $\partial(u,v)/\partial(z,x)$.

Suggested homework. I(1, 3, 4, 7), II(3, 4, 5, 7), III(1, 3, 5, 7), IV(2, 5, 7); Exam corner E2, E3, E6.

🔬 Research–Teaching NexusResearch-tutored · optional, about 45 minutes

Research question. After an accidental chemical spill into a river, when does the contamination reach a town downstream, and how concentrated is it on arrival — and how much do the answers change when the river's speed varies along its course?

Why it matters. Water utilities along large rivers (the Rhine after the 1986 spill near Basel, the Danube, the Ganga, the Mississippi) close their intakes on the basis of exactly this forecast. The simplest model, $c_t+v(x)\,c_x=-kc$ (advection with first-order decay, dispersion neglected), is a linear first-order PDE that Lagrange's method solves in closed form along characteristics.

What to do.

  1. Solve $c_t+v\,c_x=-kc$, $c(x,0)=c_0(x)$, for constant $v$ by Lagrange's method, and verify your answer by substitution.
  2. For variable speed $v(x)\gt0$, show that the characteristics are $t-\tau(x)=$ const with travel time $\tau(x)=\int_0^x\frac{d\xi}{v(\xi)}$, and that a pulse released at $x=0$ arrives at $x=L$ at time $\tau(L)$ with its peak multiplied by $e^{-k\tau(L)}$.
  3. Take $L=120\ \text{km}$, $k=0.03\ \text{h}^{-1}$, and compare (a) $v\equiv0.8\ \text{m/s}$ with (b) $v(x)=0.8\,(1+x/L)\ \text{m/s}$ (a river that speeds up downstream). Compute arrival times and the surviving fraction of the peak.
  4. Show that in case (b) the pulse is stretched in space by the factor $v(L)/v(0)$ but not in time: plot $c(L,t)$ for a Gaussian release.

Tools: pen and paper for the solution; a spreadsheet or Python for the numbers and the plot.

Hand in: a one-page note with the solution formula, the two arrival times and surviving fractions, and one plot of $c(L,t)$.

Going further. Add dispersion, $c_t+vc_x=Dc_{xx}-kc$ (a second-order equation, Chapter 4), and look up how its fundamental solution modifies the arrival-time forecast; read about how real spill models are calibrated with tracer dye experiments.

Instructor note: (a) $\tau=120000/0.8\ \text{s}=41.7\ \text{h}$, surviving fraction $e^{-1.25}\approx0.29$. (b) $\tau=\frac{L}{v_0}\log2\approx28.9\ \text{h}$, fraction $e^{-0.867}\approx0.42$. The time-profile at $x=L$ in case (b) is the initial time-profile at $x=0$ delayed by $\tau(L)$ and damped by $e^{-k\tau(L)}$; in space it is twice as wide because $v(L)=2v(0)$.

🧪 Self-check quiz

0/5 correct

Q1. The equation $z\,p+q=z^2$ is

Q2. Lagrange's auxiliary equations for $P\,p+Q\,q=R$ are

Q3. Eliminating $f$ from $z=f(x^2+y^2)$ gives

Q4. A $C^1$ solution of $y\,p-x\,q=0$ on all of $\mathbb{R}^2$ with $z(x,0)=g(x)$ for every $x$ exists if and only if

Q5. The general solution of $x\,p+y\,q=2z$ (for $x\gt0$) is