Symbols, Characteristics, and PDE Type

The principal symbol is the part of a differential operator seen by very short-wavelength disturbances. Its zero set identifies characteristic covectors and characteristic hypersurfaces. Elliptic operators have no nonzero real characteristic covectors, hyperbolic operators organize propagation around real characteristic cones, and parabolic evolution uses an anisotropic balance between one time derivative and two spatial derivatives. These distinctions explain, at the structural level, why Laplace, wave, and heat equations require different data and produce different kinds of solutions.

Helpful background. Fourier Series, Fourier Transforms, and Plancherel Theory supplies the derivative-to-momentum correspondence used to interpret symbols.

From derivatives to covectors

Let $P$ be a scalar differential operator of order $m$ on an open set:

$$P = \sum_{|\alpha|\leq m} b_\alpha(x)\,\partial^\alpha, \qquad \partial^\alpha = \partial_1^{\alpha_1}\cdots\partial_n^{\alpha_n}.$$

With this site’s Fourier convention,

$$\widetilde u(p) = \int d^n x\,e^{+ip\cdot x}u(x), \qquad u(x) = \int\frac{d^n p}{(2\pi)^n}\, e^{-ip\cdot x}\widetilde u(p),$$

a derivative acts as $\partial_j\mapsto -ip_j$. It is therefore convenient to write $D_j=i\partial_j$, so that $D_j\mapsto p_j$. If

$$P = \sum_{|\alpha|\leq m}a_\alpha(x)D^\alpha,$$

then its full symbol and principal symbol in this convention are

$$p(x,\xi) = \sum_{|\alpha|\leq m}a_\alpha(x)\xi^\alpha, \qquad p_m(x,\xi) = \sum_{|\alpha|=m}a_\alpha(x)\xi^\alpha.$$

Equivalently, starting from the coefficients $b_\alpha$ of the $\partial^\alpha$ expression, substitute $\partial_j\mapsto-i\xi_j$ and keep the terms of degree $m$. Some mathematics references define $D_j=-i\partial_j$ because they use the opposite Fourier sign. That changes the displayed symbol by predictable powers of $-1$, but not the geometric question of whether its homogeneous leading part vanishes.

The principal symbol also follows directly from a high-frequency test. For a smooth real phase $\phi$ and amplitude $a$,

$$u_\lambda(x)=e^{-i\lambda\phi(x)}a(x), \qquad \lambda\longrightarrow+\infty,$$

one has

$$P u_\lambda = e^{-i\lambda\phi} \left[ \lambda^m p_m(x,d\phi)\,a +O(\lambda^{m-1}) \right].$$

Thus $p_m$ controls the leading response to rapidly oscillating waves. Because $d\phi$ is a covector, not a vector, symbols naturally live on the cotangent bundle. Under a change of coordinates, the covector and the principal symbol transform together, so the equation $p_m(x,\xi)=0$ is coordinate independent.

For a square $N$-component system, $p_m(x,\xi)$ is an $N\times N$ matrix. The relevant question is then whether this matrix is invertible. Systems that are not square require a rank condition rather than a determinant test.

Characteristics and admissible data surfaces

A nonzero real covector $\xi\in T_x^*M$ is characteristic for a scalar operator when

$$p_m(x,\xi)=0.$$

For a square system, it is characteristic when

$$\det p_m(x,\xi)=0.$$

Let a hypersurface $S$ be locally defined by $\phi(x)=0$ with $d\phi\neq0$ on $S$. It is characteristic at $x\in S$ precisely when its normal covector is characteristic:

$$p_m(x,d\phi_x)=0.$$

On a noncharacteristic surface, the highest normal derivative can locally be solved for in terms of tangential derivatives and lower-order data. This is the algebraic prerequisite for posing ordinary Cauchy data there. It is not, by itself, a proof of existence, uniqueness, stability, or global well-posedness.

Only the top-order coefficients enter $p_m$. Lower-order terms can change spectra, decay rates, masses, and Green functions, but they do not change the characteristic set of a fixed-order operator.

Elliptic, hyperbolic, and parabolic structures

Ellipticity

A scalar operator is elliptic at $x$ when

$$p_m(x,\xi)\neq0 \quad\text{for every real }\xi\neq0.$$

For systems, ellipticity means that $p_m(x,\xi)$ is invertible for every nonzero real $\xi$. Elliptic equations have no real characteristic hypersurfaces. Their basic problems are consequently boundary-value problems rather than finite-speed initial-value problems. Ellipticity underlies regularity: under suitable hypotheses, singularities in the solution must be tied to singularities in the forcing or boundary data.

For the Euclidean Laplacian,

$$P=-\Delta = -\sum_{j=1}^n\partial_j^2, \qquad p_2(\xi)=|\xi|^2,$$

so the only real zero is $\xi=0$ and $-\Delta$ is elliptic.

Hyperbolicity

Let $p_m$ be a real homogeneous principal polynomial. One standard scalar definition says that it is hyperbolic with respect to a real covector $n$ if

$$p_m(x,n)\neq0$$

and, for every real $\xi$, all roots $\lambda$ of

$$\lambda\longmapsto p_m(x,\xi+\lambda n)$$

are real. The covector $n$ selects a time direction; characteristic roots then describe real propagation speeds relative to it. Distinct roots give strict hyperbolicity away from unavoidable degeneracies. For systems, real characteristic roots alone need not give stable evolution: strong hyperbolicity additionally requires sufficiently uniform control of the characteristic modes.

With the $(+---)$ metric convention, the wave operator is

$$\Box=\partial_t^2-\nabla^2.$$

Its principal symbol in the Fourier convention above is

$$p_2(\tau,\mathbf k) = -\tau^2+|\mathbf k|^2.$$

The characteristic set is the null cone $\tau^2=|\mathbf k|^2$. The normal $dt$ to a constant-time surface gives $p_2(dt)=-1$, so such a surface is noncharacteristic. By contrast, a surface $\phi(t,\mathbf x)=0$ is characteristic when

$$-(\partial_t\phi)^2+|\nabla\phi|^2=0.$$

Parabolic scaling

For a second-order scalar operator in two variables,

$$P_2 = A\,\partial_x^2 +2B\,\partial_x\partial_y +C\,\partial_y^2,$$

the traditional local classification uses

$$\Delta_{\mathrm{type}}=B^2-AC.$$

At a nonsingular point, $\Delta_{\mathrm{type}}<0$ is elliptic, $\Delta_{\mathrm{type}}>0$ is hyperbolic, and $\Delta_{\mathrm{type}}=0$ is parabolic. This discriminant is useful for reducing a scalar second-order equation to a local canonical form.

Evolutionary parabolic equations are more clearly described with weighted orders. For the heat operator

$$P_{\mathrm{heat}}=\partial_t-\kappa\Delta_{\mathbf x}, \qquad \kappa>0,$$

one time derivative balances two spatial derivatives. Assigning weight $2$ to the temporal frequency $\tau$ and weight $1$ to each spatial frequency $k_j$ gives the weighted principal symbol

$$p_{\mathrm{heat}}^{(2)}(\tau,\mathbf k) = -i\tau+\kappa|\mathbf k|^2.$$

This expresses the scaling $t\sim|\mathbf x|^2$. If ordinary isotropic second-order counting were used instead, the time derivative would be discarded and the result would miss the evolution structure. Parabolicity also carries a time orientation: spatial Fourier modes satisfy

$$\partial_t\widetilde u(t,\mathbf k) +\kappa|\mathbf k|^2\widetilde u(t,\mathbf k)=0, \qquad \widetilde u(t,\mathbf k) = e^{-\kappa|\mathbf k|^2t}\widetilde u(0,\mathbf k)$$

for forward time. Reversing that evolution amplifies high frequencies and is unstable.

Canonical operators at a glance

Operator	Relevant leading symbol	Structure	Immediate consequence
$-\Delta_{\mathbf x}+m^2$	$\lvert\mathbf k\rvert^2$	elliptic	no nonzero real characteristic covector
$\Box+m^2$	$-\tau^2+\lvert\mathbf k\rvert^2$	hyperbolic	null characteristic cone and admissible spacelike Cauchy surfaces
$\partial_t-\kappa\Delta_{\mathbf x}$	$-i\tau+\kappa\lvert\mathbf k\rvert^2$ with weights $(2,1)$	parabolic	forward smoothing with $t\sim\lvert\mathbf x\rvert^2$

The mass $m^2$ does not appear in either second-order principal symbol. It changes the full Fourier-space equation, but not the high-frequency characteristics.

Klein–Gordon and Dirac operators

For the Klein–Gordon equation,

$$(\Box+m^2)\phi=0,$$

the full plane-wave condition is

$$-p_0^2+|\mathbf p|^2+m^2=0, \qquad\text{or equivalently}\qquad p^2=m^2.$$

This massive dispersion relation is not the characteristic equation. The principal symbol omits $m^2$, so the characteristic cone remains $p^2=0$. Massive disturbances have subluminal group velocity, while the high-frequency propagation boundary is still null.

The free Dirac operator is first order:

$$P_D=i\gamma^\mu\partial_\mu-m.$$

Its principal symbol is

$$\sigma_1(P_D)(p) = \gamma^\mu p_\mu = p\!\!\!/.$$

The Clifford relation gives

$$(p\!\!\!/)^2=p^2\mathbb 1,$$

so the symbol fails to be invertible on the null cone. The same conclusion follows by composing the two mass signs:

$$(i\gamma^\mu\partial_\mu-m) (i\gamma^\nu\partial_\nu+m) = -(\Box+m^2)\mathbb 1.$$

After Euclidean continuation, $-\Delta_E+m^2$ is elliptic instead. This change of operator type is one reason Euclidean boundary-value methods and Lorentzian causal evolution must not be interchanged without specifying the continuation and its prescription.

The scalar-field application is developed on Klein–Gordon field and mode solutions.

What the classification does not decide

The principal symbol is deliberately incomplete.

• It does not select retarded, advanced, Feynman, or other Green functions. Those choices require boundary conditions, support conditions, or an explicit $i0$ prescription.
• It does not replace an energy estimate or a well-posedness theorem. Hyperbolicity is the starting structure for such results, not their conclusion.
• It does not encode masses or potentials, because these are lower-order terms.
• It can vary with position. A variable-coefficient equation may change type across the domain, and conclusions valid in one region need not cross that interface unchanged.
• For gauge systems or constrained equations, the unreduced principal symbol may be singular for structural reasons. Gauge fixing, constraints, or a reduced system must be specified before invertibility is interpreted.

The symbol definitions and ellipticity criterion used here follow Dyatlov 2022, Definition 12.17, §13.3, and Definition 14.1, PDF. The model PDE classifications can be compared with Evans 2010, Chapters 2, 6, and 7, the characteristic-root discussion with Melrose 2016, Hyperbolic Differential Operators, PDF, and the Klein–Gordon and Dirac examples with Schwartz 2014, §§2.3 and 10.3.

For existence and stability in declared function spaces, continue to Weak Solutions, Sobolev Spaces, and Well-Posedness. For inverse kernels, continue to Fundamental Solutions and Green Operators; that page and the symbol page are the two hard inputs to the hyperbolic and elliptic branches.

Common pitfalls

Using the full dispersion relation as the characteristic equation. Characteristics come from the homogeneous top-order symbol. For $\Box+m^2$, the mass shell is $p^2=m^2$, but the characteristic cone is $p^2=0$.

Classifying the heat operator by ordinary spacetime order alone. The terms $\partial_t$ and $\Delta_{\mathbf x}$ balance under parabolic scaling, even though their ordinary derivative orders differ. Dropping $\partial_t$ erases the direction of evolution.

Treating a characteristic covector as a propagation trajectory. A covector is normal to a phase surface. Rays or bicharacteristics arise only after the symbol is used to generate a Hamiltonian flow on phase space.

Ignoring Fourier-sign translations. Under the convention used here, $\partial_\mu\mapsto-ip_\mu$ and $i\partial_\mu\mapsto p_\mu$. A source using $D=-i\partial$ must be translated before formulas are compared term by term.

Exercises

1. For $P=A\partial_x^2+2B\partial_x\partial_y+C\partial_y^2$, show that the characteristic-curve slopes are real precisely when $B^2-AC\geq0$, assuming $A\neq0$.

   Solution

   A curve $\phi(x,y)=0$ is characteristic when

   $$A(\partial_x\phi)^2 +2B(\partial_x\phi)(\partial_y\phi) +C(\partial_y\phi)^2=0,$$

   where an irrelevant overall sign from the Fourier substitution has been removed. Along a level curve $y=y(x)$, $\partial_x\phi+\partial_y\phi\,dy/dx=0$. Setting $s=dy/dx$ therefore gives

   $$As^2-2Bs+C=0, \qquad s=\frac{B\pm\sqrt{B^2-AC}}{A}.$$

   The slopes are real exactly when $B^2-AC\geq0$. Equality gives the repeated characteristic direction of the parabolic case.

2. Let $\phi(t,\mathbf x)=t-f(\mathbf x)$. Determine when the level surfaces of $\phi$ are characteristic for the wave operator.

   Solution

   The normal covector is $d\phi=(1,-\nabla f)$. Substitution into the wave principal symbol gives

   $$p_2(d\phi)=-1+|\nabla f|^2.$$

   Hence the surface is characteristic exactly when $|\nabla f|=1$. Constant-time surfaces have $f=\text{constant}$ and are noncharacteristic.

3. Solve the spatial Fourier modes of $(\partial_t-\kappa\Delta)u=0$ and explain why backward evolution is unstable.

   Solution

   Since $\Delta\mapsto-|\mathbf k|^2$, each mode obeys

   $$\partial_t\widetilde u +\kappa|\mathbf k|^2\widetilde u=0.$$

   Thus

   $$\widetilde u(t,\mathbf k) = e^{-\kappa|\mathbf k|^2(t-t_0)} \widetilde u(t_0,\mathbf k).$$

   For $t>t_0$, high-frequency modes are strongly damped. Recovering earlier data multiplies them by $e^{+\kappa|\mathbf k|^2(t-t_0)}$, so arbitrarily small high-frequency errors can become arbitrarily large.

References

• Semyon Dyatlov (2022), Lecture Notes for 18.155: Distributions, Elliptic Regularity, and Applications to PDEs, PDF, Definition 12.17 for constant-coefficient full and principal symbols, § 13.3 for the invariant principal symbol, and Definition 14.1 for ellipticity. The notes use a different $D$ convention, translated above.
• Lawrence C. Evans (2010), Partial Differential Equations, second edition, Chapters 2, 6, and 7, for the classical elliptic, hyperbolic, and evolution equations.
• Richard B. Melrose (2016), Hyperbolic Differential Operators, PDF, for hyperbolic polynomials and characteristic roots.
• Matthew D. Schwartz (2014), Quantum Field Theory and the Standard Model, § 2.3 for the free scalar and Klein–Gordon equation, and § 10.3, especially equations (10.66), (10.92), and (10.94), for the Clifford relation, Dirac operator, and its square.