Partial Differential Equations
Method of Characteristics: Turning a PDE Into ODEs Along Curves
The method of characteristics performs a small miracle: it dissolves a first-order partial differential equation in n variables into a family of ordinary differential equations, one for each curve that threads through the domain. Along these privileged characteristic curves, the PDE stops being a statement about how a function couples its many partial derivatives and becomes a simple ODE describing how the solution is transported. Trace the curves back to where you know the data, integrate, and you have the solution — no Fourier transform, no Green's function, just curves and calculus.
Precisely: for a quasilinear first-order PDE a(x,y,u)uₓ + b(x,y,u)u_y = c(x,y,u), the solution surface z = u(x,y) in ℝ³ is woven from integral curves of the vector field (a, b, c). Those curves satisfy the characteristic system dx/dt = a, dy/dt = b, du/dt = c, and the graph of u is exactly the union of the characteristics launched from a curve carrying the initial data.
- FieldPartial differential equations (first-order)
- Applies toFirst-order PDEs: linear, semilinear, quasilinear, fully nonlinear
- Key hypothesisNon-characteristic (transversal) initial data: aΦ_y − bΦₓ ≠ 0 on the data curve
- Core statementA solution surface is a union of integral curves of the field (a, b, c)
- Proof techniquePicard–Lindelöf existence for the characteristic ODEs + inverse function theorem
- Attributed toCauchy, Hamilton, Charpit, Lagrange, Monge (18th–19th c.)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Precise statement: the characteristic system
Consider the quasilinear Cauchy problem a(x,y,u)uₓ + b(x,y,u)u_y = c(x,y,u), with u prescribed on an initial curve Γ ⊂ ℝ² parametrized by s ↦ (f(s), g(s)) carrying data u = h(s). Assume a, b, c ∈ C¹ in a neighborhood of the initial strip.
The characteristic ODEs are
dx/dt = a(x,y,z), dy/dt = b(x,y,z), dz/dt = c(x,y,z),
with initial conditions x(0,s) = f(s), y(0,s) = g(s), z(0,s) = h(s). If the non-characteristic (transversality) condition
J(s) = a(f,g,h)·g′(s) − b(f,g,h)·f′(s) ≠ 0
holds along Γ, then in a neighborhood of Γ the map (t,s) ↦ (x(t,s), y(t,s)) is a C¹ diffeomorphism, and u(x,y) := z(t(x,y), s(x,y)) is the unique C¹ solution of the Cauchy problem there. This is the local Cauchy–Kovalevskaya statement for first-order equations, but with a far softer hypothesis than analyticity.
The picture: weaving a surface from curves
Think of the solution as a surface z = u(x,y) sitting in ℝ³. The PDE a uₓ + b u_y = c can be rewritten as (a, b, c)·(uₓ, u_y, −1) = 0, which says the vector field V = (a, b, c) is tangent to the solution surface at every point (because (uₓ, u_y, −1) is the surface normal). So the solution surface must be everywhere tangent to V.
A surface tangent to a vector field is swept out by that field's integral curves. Hence: pick the space curve Γ̃ = (f(s), g(s), h(s)) sitting above the initial data, launch an integral curve of V from each of its points, and the ruled surface they trace is the graph of u. The PDE has become the geometry of flowing a curve along a field. The projections of these curves onto the (x,y)-plane are the characteristic base curves along which information propagates.
Key idea of the proof: ODE existence plus the inverse function theorem
The mechanism is a clean two-step reduction. Step 1 (transport is an ODE). By Picard–Lindelöf, the C¹ (hence locally Lipschitz) field (a,b,c) generates a unique local flow: for each fixed s the characteristic ODEs have a unique solution (x(t,s), y(t,s), z(t,s)) depending C¹ on (t,s). This produces a candidate parametric surface with no PDE in sight — only ODEs.
Step 2 (invert the parametrization). To read off u(x,y) we must solve (x,y) = (x(t,s), y(t,s)) for (t,s). The Jacobian of this map at t=0 is ∂(x,y)/∂(t,s) = a g′ − b f′ = J(s), the transversality determinant. When J ≠ 0 the inverse function theorem gives a C¹ inverse (t(x,y), s(x,y)) locally, and u := z∘(t,s) is well-defined and C¹. A short chain-rule computation confirms it solves the PDE and matches the data. Uniqueness follows because any solution's surface must contain the characteristics through Γ̃, which are unique.
Worked example: inviscid Burgers and shock formation
Take Burgers' equation u_t + u uₓ = 0 with u(x,0) = h(x) — quasilinear with (a,b,c) = (1, u, 0) in variables (t, x, u). The characteristic ODEs are dt/dτ = 1, dx/dτ = u, du/dτ = 0. So u is constant along each characteristic, and since dx/dt = u = const, the base characteristics are straight lines: x = h(s)·t + s, carrying value u = h(s).
Solving implicitly gives u(x,t) = h(x − u t). This is exact and elementary — a genuine reduction to ODEs. But watch the Jacobian: ∂x/∂s = 1 + h′(s) t. It vanishes at t* = −1/h′(s). Wherever h is decreasing (h′ < 0), characteristics collide at finite time t* = min_s(−1/h′(s)); the map (t,s) ↦ (t,x) stops being invertible and uₓ blows up. The classical solution breaks down and a shock forms, past which one continues with a weak (Rankine–Hugoniot) solution.
Why the hypotheses matter: transversality, smoothness, and what breaks
Transversality (J ≠ 0) is essential. If Γ is itself a characteristic base curve, then J ≡ 0: the data curve lies along the flow rather than across it. Then either the prescribed data is incompatible with the transport law (no solution), or it is compatible and infinitely many solutions exist (the data fails to pin down neighboring characteristics). Example: for uₓ = 0, prescribing u on the line x = 0 (transversal) works; prescribing u along a line y = const (a characteristic) generically fails.
Smoothness matters locally and globally. Even with C^∞ data, quasilinear equations lose classical solutions in finite time when characteristics cross (Burgers), so the theorem is inherently local. Only genuinely linear equations with globally transversal, non-intersecting characteristics give global classical solutions. This is where the theory hands off to weak solutions, entropy conditions, and viscosity solutions (Crandall–Lions, 1983) for Hamilton–Jacobi equations.
Applications and significance
The method is the backbone of first-order PDE theory. In conservation laws (gas dynamics, traffic flow, the Lighthill–Whitham–Richards model) characteristics predict exactly when and where shocks form and set up the Riemann problem. In geometric optics, the eikonal equation |∇u|² = n² is solved by characteristics that are precisely light rays, recovering Fermat's principle. In classical mechanics, the fully nonlinear case yields Hamilton–Jacobi theory: the characteristics of the Hamilton–Jacobi PDE are Hamilton's canonical equations, unifying wave and particle pictures.
It also underlies the classification of PDEs (characteristics distinguish hyperbolic, parabolic, elliptic types), continuous-time optimal control via the Hamilton–Jacobi–Bellman equation, and modern numerical schemes (upwinding, semi-Lagrangian, and characteristic-Galerkin methods) that integrate along numerically traced characteristics for stability. Wherever information propagates at finite speed along definite directions, characteristics reveal those directions.
| PDE class | General form | Characteristic ODEs | Do characteristics decouple from u? |
|---|---|---|---|
| Linear (homogeneous coeffs) | a(x,y)uₓ + b(x,y)u_y = 0 | dx/dt = a, dy/dt = b, du/dt = 0 | Yes — projected curves fixed in advance |
| Semilinear | a(x,y)uₓ + b(x,y)u_y = c(x,y,u) | dx/dt = a, dy/dt = b, du/dt = c(x,y,u) | Yes for x,y; u solves an ODE along them |
| Quasilinear | a(x,y,u)uₓ + b(x,y,u)u_y = c(x,y,u) | dx/dt = a, dy/dt = b, du/dt = c | No — the curves themselves depend on u |
| Fully nonlinear | F(x,y,u,uₓ,u_y) = 0 | Hamilton's 2n+1 system in (x, u, p) | No — must carry the gradient p as a variable |
Frequently asked questions
What is the non-characteristic (transversality) condition and why is it required?
It is the requirement J(s) = a·g′(s) − b·f′(s) ≠ 0 along the initial curve Γ = (f(s), g(s)) — geometrically, that Γ crosses the characteristic base curves transversally rather than running along one. This determinant is exactly the Jacobian of the parametrization map (t,s) ↦ (x,y) at t = 0. If it vanishes, the inverse function theorem fails, you cannot recover u(x,y) from the characteristic coordinates, and the Cauchy problem is either overdetermined (no solution) or underdetermined (infinitely many).
Why does the method work for first-order PDEs but not second-order ones?
For a first-order PDE the single equation says one vector field is tangent to the solution surface, and a surface tangent to a field is swept by its integral curves — a clean geometric reduction to ODEs. Second-order (and higher) equations impose conditions on curvature-like second derivatives that no single field of curves can encode; there characteristics still exist as surfaces where the equation degenerates and govern propagation of singularities, but they no longer determine the full solution. Hyperbolic second-order equations do admit related characteristic/Riemann-invariant methods, but the elementary curve-integration recipe is special to first order.
What happens to characteristics when a shock forms?
In a genuinely nonlinear (quasilinear) equation like Burgers', characteristics carrying different values can converge and cross at a finite time t* = min_s(−1/h′(s)) wherever the data is decreasing. At the crossing the map from characteristic coordinates to (x,y) becomes non-invertible (Jacobian = 0), uₓ blows up, and no single-valued classical solution exists past t*. One continues with a weak solution containing a shock discontinuity whose speed is fixed by the Rankine–Hugoniot jump condition and selected by an entropy condition.
How does the method extend to fully nonlinear equations F(x, u, ∇u) = 0?
You must augment the state with the gradient p = ∇u, because now the characteristic direction depends on p itself. This gives Charpit's / Hamilton's characteristic system: 2n+1 ODEs for (x, u, p), namely dx/dt = ∇_p F, dp/dt = −∇_x F − p ∂_u F, du/dt = p·∇_p F. These are exactly Hamilton's canonical equations when F is the Hamiltonian, which is why Hamilton–Jacobi theory in mechanics is the fully nonlinear instance of the method of characteristics.
Is the solution produced by the method unique?
Locally, yes, as a C¹ (classical) solution, provided the transversality condition holds. Any classical solution's graph must contain the unique integral curves of (a,b,c) passing through the initial strip, so the surface — hence u — is forced near Γ. Globally, uniqueness can fail once classical solutions cease to exist (shocks); among weak solutions uniqueness is restored only by imposing an admissibility/entropy condition (Kruzhkov's theorem for scalar conservation laws, or viscosity-solution uniqueness for Hamilton–Jacobi).
Does the initial data need to be analytic, as in Cauchy–Kovalevskaya?
No. That is a key advantage over the general Cauchy–Kovalevskaya theorem. For first-order PDEs the method needs only C¹ coefficients and C¹ non-characteristic data to produce a local C¹ solution, because the engine is Picard–Lindelöf for the ODEs plus the inverse function theorem — both of which require only Lipschitz/C¹ regularity, not real-analyticity. Analyticity buys you nothing extra here beyond a real-analytic solution if the data happens to be analytic.