Why two variables was never the real story
In the earlier guides of this rung you sorted second-order equations with a single number. Write the general second-order linear PDE in two variables as A u_xx + 2B u_xy + C u_yy + (lower-order terms) = 0, and the sign of the discriminant B^2 - A C tells you the type: negative means elliptic, zero means parabolic, positive means hyperbolic. It is clean, memorable, and — past two independent variables — completely stuck. The heat equation u_t = u_xx + u_yy lives in three variables; Laplace's equation in space lives in three or more. A single discriminant has no idea what to do with them.
The fix is not a bigger formula — it is a change of viewpoint. The discriminant was secretly asking a question about a quadratic form, and B^2 - A C was just that form's signature dressed up in two-variable clothing. Once we ask the question directly, the answer scales to any number of variables for free. The tool that asks it directly is called the symbol, and learning to read it is the capstone of this whole rung.
From the principal part to the symbol
Start with the only part that matters: the principal part, the bundle of highest-order derivatives. In n variables x_1, ..., x_n a second-order operator's principal part is a sum of terms a_ij u_(x_i x_j). Collect the coefficients into a symmetric n-by-n matrix A = (a_ij). For the 2D case this matrix is exactly [A, B; B, C], the same letters as the discriminant — which is your first hint that B^2 - A C was about this matrix all along.
Now perform the magic step that defines the symbol. Take each derivative and replace it by a frequency variable: u_(x_j) becomes a factor of xi_j, where xi = (xi_1, ..., xi_n) is a vector of "directions". A second derivative u_(x_i x_j) becomes xi_i xi_j. The principal part then turns into a purely algebraic expression in xi — the sum of a_ij xi_i xi_j — and that is the principal symbol. There are no derivatives left; the operator has become a polynomial in the direction vector xi.
operator u_x_i x_j principal symbol P(xi) --------------------------------------------------------- u_xx d^2/dx^2 --> xi_1^2 u_xx+u_yy Laplacian --> xi_1^2 + xi_2^2 u_xx-u_tt wave --> xi_1^2 - xi_2^2 u_xx, u_t heat --> xi_1^2 (the u_t term is lower order) rule: replace each d/dx_j by xi_j in the highest-order part only
Where does this substitution come from? It is not arbitrary. Feed the operator a plane wave u = e^(i xi . x) — a wave oscillating in the direction xi — and every derivative d/dx_j pulls down exactly a factor of i xi_j. The whole symbol of the operator is just the number the operator multiplies that plane wave by, and the principal symbol is its leading (top-degree) piece. So the symbol is literally the operator's response to pure frequencies, which is why it controls everything about wave-like behaviour.
Counting signs: the classification rule
The principal symbol P(xi) = sum of a_ij xi_i xi_j is a quadratic form, and the matrix A behind it is symmetric — so it can be diagonalized, and what survives a change of coordinates is the signs of its eigenvalues. That triple of counts — how many eigenvalues are positive, negative, and zero — is the form's signature, and it is exactly what classification by the symbol reads off. Everything reduces to counting plus, minus, and zero.
- Write the coefficient matrix A of the principal part and find the signs of its n eigenvalues (you rarely need the values — just whether each is +, -, or 0).
- If every eigenvalue has the same sign and none is zero (all + or all -), the equation is elliptic. The model is Laplace's equation, whose symbol xi_1^2 + ... + xi_n^2 vanishes only at xi = 0.
- If all are nonzero and exactly one has the opposite sign to the rest (n-1 of one sign, 1 of the other), the equation is hyperbolic. The model is the wave equation, symbol xi_1^2 - (xi_2^2 + ... + xi_n^2), which vanishes on a whole cone of directions.
- If exactly one eigenvalue is zero and the rest share a sign, the equation is parabolic — the symbol is degenerate in one direction. The model is the heat equation, where the missing direction is time and the second-order part is only the spatial Laplacian.
Notice how this exactly reproduces the old two-variable rule. For [A, B; B, C] the product of the two eigenvalues equals the determinant A C - B^2, which is the negative of the discriminant. Same signs (elliptic) means a positive determinant means B^2 - A C < 0; opposite signs (hyperbolic) means a negative determinant means B^2 - A C > 0; a zero eigenvalue (parabolic) means determinant zero means B^2 - A C = 0. The discriminant was the eigenvalue-sign test in disguise, and the symbol simply lets it breathe in higher dimensions.
What the signs actually mean
These sign-counts are not bookkeeping; they are the physics. An elliptic symbol never vanishes except at xi = 0, which means there is no real direction in which a pure oscillation costs nothing — no wave can propagate freely, and the solution settles into a steady, smoothed-out equilibrium. That is why Laplace's equation describes balance, why its solutions are infinitely smooth inside the region, and why prescribing Cauchy data for it is ill-posed: with no propagation, there is no notion of "evolving forward" to support such data.
A hyperbolic symbol, by contrast, vanishes on a whole cone of real directions — the characteristic cone. Those directions are the ones along which signals travel, and their existence is precisely why hyperbolic equations have a finite propagation speed, a domain of dependence, and waves that keep their sharp edges instead of smearing. The parabolic case sits between: one degenerate direction (time) and an elliptic spatial part, giving infinite propagation speed and instant smoothing — the heat equation's signature, and the reason running it backward in time is ill-posed.
Honest edges: when one label is not enough
The clean trichotomy hides three honest caveats. First, the type can change from point to point, because the coefficients a_ij can depend on x. The Tricomi equation u_yy = y u_xx is elliptic where y > 0 and hyperbolic where y < 0, switching across the line y = 0 — a genuine equation of mixed type that models transonic flow, where subsonic (elliptic) and supersonic (hyperbolic) regions really do meet.
Second, beyond three variables the eigenvalue signs admit patterns the three names cannot cover. If the form has several positive and several negative eigenvalues — say two and two — it is neither elliptic, parabolic, nor hyperbolic but ultrahyperbolic. These are mathematically real but physically exotic; the point is that "elliptic / parabolic / hyperbolic" is the menu for the common cases, not a law that every operator must obey.
Third, the symbol is fundamentally a linear idea — it reads the highest-order coefficients, which for a nonlinear equation depend on the unknown solution itself. One can still classify a nonlinear equation by freezing its coefficients at a given state (linearizing), but then the type can vary with the solution, not just with position. This is no academic footnote: it is why a single fluid equation can behave as different types in different regimes, and why the deepest questions — like global regularity for 3D Navier-Stokes — stay open. The symbol tells you the local character; it does not promise the global story is simple.