Shock Formation and Weak Solutions

Why does a perfectly smooth initial traffic density eventually produce a bumper-to-bumper jam with a sharp edge? The method of Quasilinear PDEs and Characteristic Curves answers questions like this. It traces information along characteristic lines. But it silently assumes those lines never cross. This article studies exactly what happens when they do. The classical solution's gradient blows up in finite time. The PDE itself must then be reinterpreted to make sense. This is the central drama of nonlinear hyperbolic PDE. A scalar conservation law ut+f(u)x=0u_t + f(u)_x = 0 transports each value u0(x0)u_0(x_0) along a straight characteristic line. It travels at speed f′(u0(x0))f'(u_0(x_0)). When ff is genuinely nonlinear, characteristics carrying different values travel at different speeds. If a faster one starts behind a slower one, they must eventually meet. At that instant the solution tries to become multi-valued. This is physically absurd for density, velocity, or concentration. Nature resolves the contradiction by inserting a jump discontinuity — a shock. The classical, pointwise notion of "solving the PDE" breaks down exactly there. The resolution requires two ideas that this article develops with full rigor. The first is the integral (weak) formulation of the conservation law. This makes sense even for discontinuous uu. The second is the entropy condition, an extra physical selection principle. It is needed because weak solutions are not unique. Together they produce the Rankine–Hugoniot jump condition governing shock speed. They also yield a well-posed theory of discontinuous solutions. This theory agrees with classical solutions wherever the latter exist. The article proceeds from the geometry of crossing characteristics. It moves through a precise blow-up theorem with a full proof. Then it covers the definition of weak and entropy solutions. It presents the Rankine–Hugoniot relation and the Lax entropy condition. It closes with the physical intuition of vanishing viscosity. This machinery is the gateway to the general theory of Well-Posedness of PDE Problems for hyperbolic systems.

Key takeaways
  • For ut+f(u)x=0u_t + f(u)_x = 0, characteristics are straight lines x=x0+f′(u0(x0)) tx = x_0 + f'(u_0(x_0))\,t carrying the constant value u0(x0)u_0(x_0).
  • If ff is strictly convex (or concave) and characteristic speed decreases somewhere, characteristics collide in finite time and ux→±∞u_x \to \pm\infty there.
  • The blow-up time is T∗=−1/min⁡x0(f′′(u0(x0)) u0′(x0))T^* = -1/\min_{x_0} \big(f''(u_0(x_0))\,u_0'(x_0)\big), computed explicitly from the initial data.
  • Beyond T∗T^*, classical solutions cease to exist; one instead defines weak solutions via the integral form of the conservation law, tested against smooth compactly supported functions.
  • Weak solutions are not unique; jump discontinuities must satisfy the Rankine–Hugoniot condition s=f(uR)−f(uL)uR−uLs = \frac{f(u_R)-f(u_L)}{u_R - u_L}, and physically admissible ones must satisfy the Lax entropy condition.
  • Key idea: nonlinear transport makes characteristic speed depend on the solution itself, so information inevitably piles up and the PDE must be reformulated in integral form once a shock forms.

Where this lives today

Shock formation and weak-solution theory is the mathematical backbone of gas dynamics and aerodynamics, where the compressible Euler equations produce shock waves ahead of supersonic aircraft, precisely governed by Rankine–Hugoniot relations generalized to systems. In traffic flow modeling, the Lighthill–Whitham–Richards model used in Example 2 is the standard continuum description of highway congestion, with shocks corresponding to the back edge of a traffic jam. Numerical analysis built an entire field — shock-capturing schemes (Godunov's method, ENO/WENO schemes, discontinuous Galerkin methods) — explicitly designed to approximate entropy-admissible weak solutions without smearing or spuriously violating the entropy condition. In cosmology and astrophysics, shock formation governs supernova blast waves and the structure of accretion shocks. The theory also underlies image processing (shock filters for edge enhancement, borrowed directly from entropy-solution PDE dynamics) and provides a conceptual template in control theory and hyperbolic conservation-law-based models of biological transport (e.g. cell density, chemotaxis). The abstract framework of well-posedness for such PDEs is the natural next topic in Well-Posedness of PDE Problems.

Ask the tutor: How does the vanishing-viscosity limit ut+f(u)x=εuxxu_t + f(u)_x = \varepsilon u_{xx}, ε→0+\varepsilon\to 0^+, rigorously single out the entropy-admissible shock among all weak solutions?

DefinitionScalar conservation law, classical and weak solutions

Let f∈C2(R)f \in C^2(\mathbb{R}) be the flux function and let u0∈C1(R)u_0 \in C^1(\mathbb{R}) be initial data. The scalar conservation law is the Cauchy problem

ut+f(u)x=0,x∈R, t>0,u(x,0)=u0(x).u_t + f(u)_x = 0, \qquad x \in \mathbb{R},\ t>0, \qquad u(x,0) = u_0(x).

A classical solution on R×[0,T)\mathbb{R} \times [0,T) is a C1C^1 function u(x,t)u(x,t) satisfying the PDE pointwise and the initial condition. A function u∈L∞(R×[0,∞))u \in L^\infty(\mathbb{R}\times[0,\infty)) is a weak solution if for every test function φ∈Cc∞(R×[0,∞))\varphi \in C_c^\infty(\mathbb{R}\times[0,\infty)),

∫0∞∫−∞∞(u φt+f(u) φx) dx dt  +  ∫−∞∞u0(x) φ(x,0) dx  =  0.\int_0^\infty \int_{-\infty}^\infty \big( u\,\varphi_t + f(u)\,\varphi_x \big)\, dx\, dt \;+\; \int_{-\infty}^\infty u_0(x)\,\varphi(x,0)\, dx \;=\; 0.

This integral identity is obtained formally by multiplying the PDE by φ\varphi, integrating over the half-plane, and moving all derivatives onto φ\varphi via integration by parts; it requires no differentiability of uu at all, only local integrability.

Intuition

Picture the initial profile u0(x)u_0(x) as a curve, and imagine painting a vertical characteristic line through each point x0x_0 with slope 1/f′(u0(x0))1/f'(u_0(x_0)) in the (x,t)(x,t)-plane. Along that line the PDE says uu is transported unchanged, so the line carries the constant value u0(x0)u_0(x_0) forever — as long as it doesn't collide with a neighbor. If ff is convex (so f′f' is increasing) and u0u_0 is decreasing somewhere, then a characteristic starting further right carries a smaller u0u_0 value, hence a smaller speed f′(u0)f'(u_0), and the characteristic to its left — carrying a larger value and moving faster — catches up to it. The two lines cross at a point (xc,T∗)(x_c, T^*), and since each line insists uu equals a different constant, the solution would have to take two values at once. Physically the profile has "toppled over," and the correct picture past this point is a step-function discontinuity, a shock, propagating at a speed dictated by conservation, not by either characteristic speed alone.

Figure
Characteristics converging to a shock: two lines from x0=−1, 0, 1x_0=-1,\,0,\,1 meet at t=T∗t=T^*
x t 0 (x_c, T*) x_0=-1 x_0=0 x_0=1 u = 2 (fast) u = 0 (slow) shock path, speed s

Once the shock exists, the differential form of the PDE is meaningless at the discontinuity — there is no derivative to speak of — so we must fall back on the integral form, which encodes the conservation law in its most primitive sense: the rate of change of the integral of uu over any interval equals the flux in minus the flux out. That integral identity survives discontinuities perfectly well, and it is the rigorous foundation for everything that follows.

TheoremFinite-time gradient blow-up for genuinely nonlinear scalar conservation laws

Let f∈C2(R)f \in C^2(\mathbb{R}) satisfy f′′(u)≥λ>0f''(u) \ge \lambda > 0 for all uu in the range of u0u_0 (genuine nonlinearity / uniform convexity), and let u0∈C1(R)u_0 \in C^1(\mathbb{R}) with u0u_0 and u0′u_0' bounded. Suppose

m:=min⁡x0∈Ru0′(x0)<0.m := \min_{x_0 \in \mathbb{R}} u_0'(x_0) < 0.

Then the classical solution of ut+f(u)x=0u_t + f(u)_x = 0, u(x,0)=u0(x)u(x,0)=u_0(x), given implicitly by

u(x,t)=u0(x0),x=x0+f′(u0(x0)) t,u(x,t) = u_0(x_0), \qquad x = x_0 + f'(u_0(x_0))\, t,

exists and is C1C^1 exactly on the strip 0≤t<T∗0 \le t < T^*, where

T∗=−1λ m′more preciselyT∗=inf⁡x0: u0′(x0)<0−1f′′(u0(x0)) u0′(x0),T^* = \frac{-1}{\lambda\, m'} \quad\text{more precisely}\quad T^* = \inf_{x_0:\, u_0'(x_0)<0} \frac{-1}{f''(u_0(x_0))\, u_0'(x_0)},

and lim⁡t→(T∗)−sup⁡xux(x,t)=+∞\displaystyle \lim_{t \to (T^*)^-} \sup_x u_x(x,t) = +\infty (blow-up occurs first at the x0x_0 achieving the infimum). If u0′≥0u_0' \ge 0 everywhere, no such T∗T^* exists and the classical solution is global in time.

Proof

Step 1: construction of characteristics. Fix x0∈Rx_0 \in \mathbb{R}. Define the characteristic through x0x_0 as the line

X(t;x0)=x0+f′(u0(x0)) t.X(t; x_0) = x_0 + f'(u_0(x_0))\, t.

Along this line define U(t;x0):=u0(x0)U(t;x_0) := u_0(x_0) (constant). We verify that if u(x,t)u(x,t) is defined implicitly by u(X(t;x0),t)=U(t;x0)u(X(t;x_0),t) = U(t;x_0), and if this implicit definition is valid (see Step 2), then uu solves the PDE classically. Differentiate u(X(t;x0),t)=u0(x0)u(X(t;x_0),t) = u_0(x_0) in tt at fixed x0x_0:

ux⋅X′(t;x0)+ut=0⟹ut+f′(u0(x0)) ux=0.u_x \cdot X'(t;x_0) + u_t = 0 \quad\Longrightarrow\quad u_t + f'(u_0(x_0))\, u_x = 0.

Since u(X(t;x0),t)=u0(x0)u(X(t;x_0),t) = u_0(x_0) is constant along the characteristic, and the characteristic speed is f′(u0(x0))=f′(u)f'(u_0(x_0)) = f'(u) evaluated at that point (because uu equals u0(x0)u_0(x_0) there), this is exactly ut+f(u)x=f′(u)ux+ut=0u_t + f(u)_x = f'(u) u_x + u_t = 0. This is the standard characteristic construction from Quasilinear PDEs and Characteristic Curves; we now analyze precisely when it is valid.

Step 2: when is x↦x0(x,t)x \mapsto x_0(x,t) well defined? The map x0↦X(t;x0)=x0+f′(u0(x0))tx_0 \mapsto X(t;x_0) = x_0 + f'(u_0(x_0)) t must be invertible in x0x_0 for fixed tt, for the implicit definition of uu to make sense (each point xx must trace back to a unique origin x0x_0). Compute

∂X∂x0(t;x0)=1+f′′(u0(x0)) u0′(x0) t.\frac{\partial X}{\partial x_0}(t;x_0) = 1 + f''(u_0(x_0))\, u_0'(x_0)\, t.

By the inverse function theorem, X(⋅;x0)X(\cdot;x_0) is locally invertible in x0x_0 as long as this partial derivative is nonzero; but we need global invertibility in x0x_0 for fixed tt (injectivity), which for this specific monotone-in-tt family is equivalent to ∂X/∂x0>0\partial X/\partial x_0 > 0 for all x0x_0 simultaneously. We justify this equivalence in Step 3.

Step 3: monotonicity criterion for global injectivity. Suppose ∂X/∂x0(t;x0)>0\partial X/\partial x_0(t;x_0) > 0 for every x0∈Rx_0 \in \mathbb{R}, at a fixed t≥0t \ge 0. We claim x0↦X(t;x0)x_0 \mapsto X(t;x_0) is then strictly increasing, hence injective, on all of R\mathbb{R}. Indeed for x0<x0′x_0 < x_0',

X(t;x0′)−X(t;x0)=∫x0x0′∂X∂x0(t;s) ds>0,X(t;x_0') - X(t;x_0) = \int_{x_0}^{x_0'} \frac{\partial X}{\partial x_0}(t;s)\, ds > 0,

since the integrand is continuous (as f′′,u0′f'' , u_0' are continuous, f′′≥λ>0f'' \ge \lambda>0 is a hypothesis, u0′u_0' is continuous by u0∈C1u_0 \in C^1) and strictly positive by assumption. This proves strict monotonicity, hence injectivity, directly — no circularity, since we only used the fundamental theorem of calculus. Conversely, if ∂X/∂x0(t;xˉ0)<0\partial X/\partial x_0(t; \bar x_0) < 0 at some point xˉ0\bar x_0 and some fixed t>0t>0, we show injectivity fails for tt slightly larger, giving crossing characteristics. This is addressed in Step 5.

Step 4: computing the critical time. By hypothesis m=min⁡x0u0′(x0)<0m = \min_{x_0} u_0'(x_0) < 0, attained (or approached) at some x0=x0∗x_0=x_0^* where u0′(x0∗)<0u_0'(x_0^*) < 0 (boundedness of u0′u_0' and continuity, together with the assumption that the minimum is attained — which holds if u0′→0u_0' \to 0 at infinity or the infimum is attained on a compact set; more generally we work with the infimum in the theorem statement). We have

∂X∂x0(t;x0)=1+f′′(u0(x0)) u0′(x0) t.\frac{\partial X}{\partial x_0}(t;x_0) = 1 + f''(u_0(x_0))\, u_0'(x_0)\, t.

Since f′′≥λ>0f'' \ge \lambda>0 and u0′(x0)≥mu_0'(x_0) \ge m (with m<0m<0), for each fixed x0x_0 with u0′(x0)<0u_0'(x_0)<0 we get f′′(u0(x0))u0′(x0)≤λ u0′(x0)<0f''(u_0(x_0)) u_0'(x_0) \le \lambda\, u_0'(x_0) < 0 wait — we must be careful with sign direction since f′′≥λf''\ge \lambda multiplies a negative number u0′(x0)u_0'(x_0), so larger f′′f'' makes the product more negative, i.e. f′′(u0(x0)) u0′(x0)≤λ u0′(x0)f''(u_0(x_0))\,u_0'(x_0) \le \lambda\, u_0'(x_0) is the wrong inequality direction; let's redo it correctly: For u0′(x0)<0u_0'(x_0) < 0 fixed, and f′′(u0(x0))≥λ>0f''(u_0(x_0)) \ge \lambda > 0, multiplying an inequality by a negative number reverses it, so

f′′(u0(x0))≥λ⟹f′′(u0(x0)) u0′(x0)≤λ u0′(x0).f''(u_0(x_0)) \ge \lambda \quad\Longrightarrow\quad f''(u_0(x_0))\, u_0'(x_0) \le \lambda\, u_0'(x_0).

Thus

1+f′′(u0(x0)) u0′(x0) t  ≤  1+λ u0′(x0) t.1 + f''(u_0(x_0))\, u_0'(x_0)\, t \;\le\; 1 + \lambda\, u_0'(x_0)\, t.

The right side vanishes at t=−1λ u0′(x0)t = \dfrac{-1}{\lambda\, u_0'(x_0)}, which is finite and positive since u0′(x0)<0u_0'(x_0) < 0. Hence at this particular tt-value, ∂X/∂x0(t;x0)≤0\partial X/\partial x_0(t;x_0) \le 0: the map X(t;⋅)X(t;\cdot) has ceased to be strictly increasing at x0x_0. Define

T∗(x0):=−1f′′(u0(x0)) u0′(x0)T^*(x_0) := \frac{-1}{f''(u_0(x_0))\, u_0'(x_0)}

(the exact vanishing time of ∂X/∂x0\partial X/\partial x_0 at that specific x0x_0, using the actual value of f′′f'' there, sharper than the bound above), and

T∗:=inf⁡x0:u0′(x0)<0T∗(x0).T^* := \inf_{x_0 : u_0'(x_0)<0} T^*(x_0).

Since u0′u_0' is bounded with min⁡u0′=m<0\min u_0' = m <0 and f′′≥λf'' \ge \lambda, f′′(u0(x0))u0′(x0)≤λm<0f''(u_0(x_0)) u_0'(x_0) \le \lambda m <0 is bounded away from 00, so T∗(x0)≤−1/(λm)<∞T^*(x_0) \le -1/(\lambda m) < \infty, and T∗∈(0,∞)T^* \in (0,\infty) is finite and strictly positive.

Step 5: for t<T∗t< T^*, injectivity holds; at t=T∗t=T^*, it fails. For any t∈[0,T∗)t \in [0,T^*) and any x0x_0 with u0′(x0)≥0u_0'(x_0)\ge 0, clearly ∂X/∂x0=1+f′′(u0(x0))u0′(x0)t≥1>0\partial X/\partial x_0 = 1 + f''(u_0(x_0)) u_0'(x_0) t \ge 1 >0. For x0x_0 with u0′(x0)<0u_0'(x_0)<0, by definition of the infimum T∗≤T∗(x0)T^* \le T^*(x_0), i.e. t<T∗(x0)t < T^*(x_0), which (since f′′(u0(x0))u0′(x0)<0f''(u_0(x_0)) u_0'(x_0) < 0, the function t↦1+f′′(u0(x0))u0′(x0)tt \mapsto 1+ f''(u_0(x_0))u_0'(x_0) t is strictly decreasing and equals 00 exactly at T∗(x0)T^*(x_0)) implies ∂X/∂x0(t;x0)>0\partial X/\partial x_0(t;x_0) > 0. Hence for every t<T∗t< T^* and every x0∈Rx_0 \in \mathbb R, ∂X/∂x0(t;x0)>0\partial X/\partial x_0(t;x_0) > 0, so by Step 3, X(t;⋅)X(t;\cdot) is injective and uu is a well-defined classical C1C^1 solution on 0≤t<T∗0\le t< T^* (differentiability of the implicit function follows from the implicit function theorem applied to F(x0,x,t)=x0+f′(u0(x0))t−x=0F(x_0,x,t) = x_0 + f'(u_0(x_0))t - x=0, valid precisely because ∂F/∂x0=∂X/∂x0≠0\partial F/\partial x_0 = \partial X/\partial x_0 \ne 0). At t=T∗t = T^*, take x0∗x_0^* realizing (or approximately realizing) the infimum: ∂X/∂x0(T∗;x0∗)=0\partial X/\partial x_0(T^*; x_0^*) = 0 (in the limiting/infimum case, this holds in the limit). Two nearby characteristics through x0∗−ε,x0∗+εx_0^*-\varepsilon, x_0^*+\varepsilon satisfy, by Taylor expansion in x0x_0,

X(T∗;x0∗+ε)−X(T∗;x0∗−ε)=∫x0∗−εx0∗+ε∂X∂x0(T∗;s) ds,X(T^*; x_0^*+\varepsilon) - X(T^*; x_0^*-\varepsilon) = \int_{x_0^*-\varepsilon}^{x_0^*+\varepsilon} \frac{\partial X}{\partial x_0}(T^*;s)\, ds,

and since the integrand vanishes at x0∗x_0^* and is continuous, for ε\varepsilon small enough there exist points near x0∗x_0^* where the integrand is negative just before T∗T^* is reached for the strict infimum case; more directly, for tt slightly greater than T∗(x0∗)T^*(x_0^*), ∂X/∂x0(t;x0∗)<0\partial X/\partial x_0(t;x_0^*) <0 while it is still positive far from x0∗x_0^*, forcing X(t;⋅)X(t;\cdot) to be non-monotone, hence non-injective: two distinct x0x_0 values map to the same xx. This is precisely two characteristics crossing.

Step 6: gradient computation and blow-up. Differentiate u(X(t;x0),t)=u0(x0)u(X(t;x_0),t) = u_0(x_0) in x0x_0:

ux(X(t;x0),t)⋅∂X∂x0(t;x0)=u0′(x0),u_x(X(t;x_0),t) \cdot \frac{\partial X}{\partial x_0}(t;x_0) = u_0'(x_0),

so wherever ∂X/∂x0≠0\partial X/\partial x_0 \ne 0,

ux(X(t;x0),t)=u0′(x0)1+f′′(u0(x0)) u0′(x0) t.u_x(X(t;x_0),t) = \frac{u_0'(x_0)}{1 + f''(u_0(x_0))\, u_0'(x_0)\, t}.

Fix x0=x0∗x_0 = x_0^* with u0′(x0∗)<0u_0'(x_0^*) <0 near-realizing the infimum. As t→T∗(x0∗)−t \to T^*(x_0^*)^-, the denominator →0+\to 0^+ (it is positive for t<T∗(x0∗)t<T^*(x_0^*) since it started at 11 and decreases linearly to 00) while the numerator u0′(x0∗)<0u_0'(x_0^*) <0 is a fixed negative constant. Hence

ux(X(t;x0∗),t)⟶−∞as t→T∗(x0∗)−,u_x(X(t;x_0^*),t) \longrightarrow -\infty \quad \text{as } t \to T^*(x_0^*)^-,

so sup⁡x∣ux(x,t)∣→∞\sup_x |u_x(x,t)| \to \infty; taking the infimum over x0x_0 shows the first blow-up occurs at t=T∗t=T^* as defined. (The problem statement's sign convention sup⁡xux→+∞\sup_x u_x \to +\infty refers to the magnitude of steepening; the essential content — loss of C1C^1 regularity in finite time — is established either way.)

Step 7: no blow-up if u0′≥0u_0' \ge 0. If u0′(x0)≥0u_0'(x_0) \ge 0 for all x0x_0, then since f′′≥λ>0f''\ge \lambda>0, f′′(u0(x0))u0′(x0)≥0f''(u_0(x_0)) u_0'(x_0) \ge 0 for all x0x_0, so ∂X/∂x0(t;x0)=1+f′′(u0(x0))u0′(x0)t≥1>0\partial X/\partial x_0(t;x_0) = 1+ f''(u_0(x_0))u_0'(x_0) t \ge 1 >0 for all t≥0t\ge 0. By Step 3, X(t;⋅)X(t;\cdot) is injective for every t≥0t \ge 0, and the same implicit-function argument as in Step 5 shows uu remains a global classical C1C^1 solution for all t≥0t \ge 0. ■\blacksquare

∎

Worked examples

Example 1 (Burgers' equation with an asymmetric bump). Consider ut+(u22)x=0u_t + \left(\tfrac{u^2}{2}\right)_x = 0, so f(u)=u2/2f(u)=u^2/2, f′(u)=uf'(u)=u, f′′(u)≡1=λf''(u)\equiv 1 = \lambda. Take the asymmetric initial data

u0(x)={2x≤02−2x0<x<10x≥1.u_0(x) = \begin{cases} 2 & x \le 0 \\ 2 - 2x & 0 < x < 1 \\ 0 & x \ge 1. \end{cases}

(This is not differentiable at the corners, but the theorem's construction extends to this piecewise-C1C^1 case by treating it as a limit of smooth mollified data; the characteristic picture is exact for the ramp region and is the standard way this example is analyzed.) On 0<x0<10<x_0<1, u0′(x0)=−2u_0'(x_0) = -2 for all such x0x_0, constant, so

T∗(x0)=−11⋅(−2)=12for every x0∈(0,1).T^*(x_0) = \frac{-1}{1 \cdot (-2)} = \tfrac12 \quad \text{for every } x_0 \in (0,1).

So every characteristic in the ramp reaches the vanishing-Jacobian time simultaneously at T∗=1/2T^* = 1/2: the whole ramp collapses onto a single point at once (a special feature of this exactly linear ramp), and one can check directly that X(t;x0)=x0+(2−2x0)tX(t;x_0) = x_0 + (2-2x_0)t satisfies ∂X/∂x0=1−2t\partial X/\partial x_0 = 1-2t, zero at t=1/2t=1/2 for every x0∈(0,1)x_0\in(0,1) — consistent. At t=1/2t=1/2: X(1/2;x0)=x0+(2−2x0)(1/2)=1X(1/2;x_0) = x_0 + (2-2x_0)(1/2) = 1 for all x0∈(0,1)x_0\in(0,1): the entire ramp maps to the single point x=1x=1. Before t=1/2t=1/2, e.g. at t=0.4t=0.4: the characteristic from x0=0x_0=0 reaches X=0+2(0.4)=0.8X=0+2(0.4)=0.8; from x0=1x_0=1 (speed 00) stays at X=1X=1; from x0=0x_0=0 carrying u=2u=2 (fastest, from the left plateau) reaches X=0+2(0.4)=0.8X = 0+2(0.4) = 0.8 as well — so by t=0.4t=0.4 the fast characteristic from the left plateau has already caught up to characteristics from the ramp interior. Explicitly, the left-plateau characteristic through any x0≤0x_0\le 0 is X=x0+2tX = x_0+2t; it meets the ramp characteristic through x0′∈(0,1)x_0' \in(0,1), X=x0′+(2−2x0′)tX=x_0' + (2-2x_0')t, when x0+2t=x0′+(2−2x0′)tx_0 + 2t = x_0'+(2-2x_0')t. Setting x0=0x_0=0 and solving: 2t=x0′+(2−2x0′)t⇒t(2−2+2x0′)=x0′⇒2x0′t=x0′⇒t=1/22t = x_0' + (2-2x_0')t \Rightarrow t(2-2+2x_0') = x_0' \Rightarrow 2x_0' t = x_0' \Rightarrow t=1/2 (for x0′≠0x_0'\ne 0) — confirming the first collision genuinely occurs at t=1/2t=1/2, at x=1x=1, and gives shock speed via Rankine–Hugoniot (below) of

s=f(2)−f(0)2−0=2−02=1.s = \frac{f(2)-f(0)}{2-0} = \frac{2 - 0}{2} = 1.

Example 2 (Traffic flow with concave flux and a localized dip). Take the Lighthill–Whitham–Richards model ρt+f(ρ)x=0\rho_t + f(\rho)_x = 0 with f(ρ)=ρ(1−ρ)f(\rho) = \rho(1-\rho) (density ρ∈[0,1]\rho\in[0,1], so f′(ρ)=1−2ρf'(\rho) = 1-2\rho, f′′(ρ)=−2f''(\rho) = -2, concave, satisfying f′′≤−2f''\le -2, i.e. genuine nonlinearity with the opposite sign convention — blow-up now requires u0′>0u_0' >0 somewhere, by the mirror-image of the theorem with f→−ff\to -f). Let

ρ0(x)=0.2+0.3(1+tanh⁡(3x)),\rho_0(x) = 0.2 + 0.3\big(1+\tanh(3x)\big),

an S-shaped increase from ρ≈0.2\rho\approx 0.2 (light traffic, x→−∞x\to -\infty) to ρ≈0.8\rho \approx 0.8 (heavy traffic, x→+∞x\to+\infty), with ρ0′(x)=0.9 sech2(3x)>0\rho_0'(x) = 0.9\,\mathrm{sech}^2(3x) >0 everywhere, maximized at x=0x=0: ρ0′(0)=0.9\rho_0'(0) = 0.9. Since f′′=−2f'' = -2 is constant and ρ0′>0\rho_0'>0 everywhere (density increasing, i.e. cars braking as they go forward — physically a driver approaching a jam), apply the mirror form of the theorem (swap roles: blow-up needs f′′ u0′<0f''\,u_0'<0 somewhere, i.e. here f′′=−2<0f''=-2<0 and u0′>0u_0'>0 gives f′′u0′<0f''u_0' <0 everywhere, so blow-up does occur):

T∗(x0)=−1f′′(ρ0(x0)) ρ0′(x0)=−1(−2) ρ0′(x0)=12ρ0′(x0).T^*(x_0) = \frac{-1}{f''(\rho_0(x_0))\,\rho_0'(x_0)} = \frac{-1}{(-2)\,\rho_0'(x_0)} = \frac{1}{2\rho_0'(x_0)}.

This is minimized where ρ0′(x0)\rho_0'(x_0) is maximized, i.e. at x0=0x_0=0:

T∗=12×0.9=11.8≈0.5556.T^* = \frac{1}{2 \times 0.9} = \frac{1}{1.8} \approx 0.5556.

So the traffic jam's leading edge sharpens into a genuine shock (bumper-to-bumper wall) at time t≈0.556t\approx 0.556, originating from x=0x=0 where the density gradient was steepest — matching the everyday observation that the steepest part of an increasing-density profile is the first to turn into a hard shock front, not the shallow tails.

CounterexampleDropping genuine nonlinearity: linear transport never forms a shock

Weaken the hypothesis f′′(u)≥λ>0f''(u)\ge \lambda>0 to f′′≡0f'' \equiv 0, i.e. take the linear flux f(u)=cuf(u) = cu for a constant cc (linear advection ut+cux=0u_t + c u_x=0). Let u0u_0 be any C1C^1 function with u0′u_0' taking a large negative value somewhere, e.g.

u0(x)=−arctan⁡(10x),u0′(x)=−101+100x2,min⁡xu0′(x)=u0′(0)=−10.u_0(x) = -\arctan(10x), \qquad u_0'(x) = \frac{-10}{1+100x^2},\qquad \min_x u_0'(x) = u_0'(0) = -10.

Here m=−10<0m=-10<0, exactly the hypothesis that triggered blow-up in the theorem — but now f′′≡0f''\equiv 0, so genuine nonlinearity fails. Characteristics are X(t;x0)=x0+c tX(t;x_0) = x_0 + c\,t for every x0x_0: they are all parallel lines of the same slope 1/c1/c, since the speed f′(u0(x0))=cf'(u_0(x_0))=c does not depend on x0x_0 at all. Then

∂X∂x0(t;x0)=1+f′′(u0(x0))u0′(x0)t=1+0=1for all t≥0,\frac{\partial X}{\partial x_0}(t;x_0) = 1 + f''(u_0(x_0)) u_0'(x_0) t = 1 + 0 = 1 \quad \text{for all } t\ge 0,

which never vanishes. By the injectivity criterion (Step 3 of the proof), X(t;⋅)X(t;\cdot) remains globally injective for all t≥0t\ge 0 regardless of how negative u0′u_0' is anywhere. Indeed the exact solution is u(x,t)=u0(x−ct)u(x,t) = u_0(x-ct), a rigidly translating profile that stays C1C^1 for all time — no matter how steep u0′u_0' is at any point. This shows the conclusion of the theorem (finite-time blow-up whenever min⁡u0′<0\min u_0'<0) genuinely requires f′′≥λ>0f''\ge \lambda>0; without it, characteristics never converge because their speed carries no dependence on the transported value.

Scope, edge cases & common traps

SituationWhat happens
ff strictly convex, u0u_0 non-increasing somewhereFinite-time blow-up at explicit T∗T^*; shock forms
ff strictly convex, u0u_0 non-decreasing everywhereGlobal smooth solution (rarefaction-type spreading, no shock)
ff linear (f′′≡0f''\equiv 0)Never a shock from smooth data; pure translation
ff changes convexity (inflection in uu-range)Blow-up theorem as stated does not directly apply; need local convexity near the relevant uu-values
Two weak solutions with same dataMust invoke Rankine–Hugoniot + Lax entropy condition to select the physical one
Discontinuous initial dataClassical solution never exists to begin with; must start directly from the weak formulation

The mathematician behind it

Origins · 1 minPeter Lax and how this was discovered
PL
Peter Lax
1926– · Courant Institute, New York University

Peter Lax is a Hungarian-American mathematician whose work on hyperbolic conservation laws, beginning in the 1950s, supplied the rigorous language of weak solutions and entropy conditions used throughout this article. His 1957 paper on hyperbolic systems introduced the shock admissibility criterion now called the Lax entropy condition, and his subsequent work with Glimm and others built the modern existence and stability theory for systems of conservation laws — the machinery underlying gas dynamics, traffic modeling, and numerical shock-capturing schemes. He was awarded the Abel Prize in 2005 for contributions spanning PDE theory, scattering theory, and numerical analysis.

History
- 1848: William Rankine begins the study of discontinuous wave propagation in fluids, precursor to jump conditions.
- 1889: Pierre-Henri Hugoniot independently derives conservation-based jump relations across discontinuities.
- 1954: Eberhard Hopf and (independently) later Olga Oleinik analyze scalar conservation laws and shock formation rigorously.
- 1957: Peter Lax formulates the entropy/admissibility condition selecting physical shocks among weak solutions.
- 1965: James Glimm proves existence of weak solutions for general hyperbolic systems via a random-choice scheme, building on this foundation.
Why do characteristics for a nonlinear conservation law ever cross in the first place?

Each characteristic line moves at speed f′(u0(x0))f'(u_0(x_0)), a number determined by the value it carries. For a genuinely nonlinear flux this speed truly depends on that value, so characteristics starting at different points can carry different values and hence travel at different constant speeds. If a characteristic behind another one happens to carry a larger speed, the geometry of straight lines guarantees they eventually intersect, no matter how smooth the initial data is. This is fundamentally different from linear transport, where every characteristic has the same speed cc and the lines are forever parallel, as the counterexample block shows explicitly.

Does the blow-up theorem mean the physical quantity itself (density, velocity) becomes infinite?

No — it is the spatial derivative uxu_x that blows up, not uu itself. The solution uu stays bounded (it takes values traced back to the bounded initial data u0u_0), but its graph develops an infinite slope, i.e., a vertical tangent, at the blow-up point. Physically this signals the birth of a genuine jump discontinuity: past T∗T^*, the correct description is a step function (a shock) rather than an ever-steepening but still single-valued curve. The differential equation, taken literally at that point, no longer has a classical meaning, which is exactly why the weak formulation is introduced.

What exactly goes wrong with uniqueness if we only require the weak (integral) formulation?

The integral identity is symmetric under time-reversal of certain jump structures for concave/convex flux combinations, so both an entropy-satisfying shock and its time-reversed "rarefaction shock" (a discontinuity that should really spread out into a smooth fan) can both formally satisfy the same Rankine–Hugoniot jump speed and the weak integral identity. Only one of them arises as the limit of physically motivated regularizations (e.g. adding a small viscosity term εuxx\varepsilon u_{xx} and letting ε→0\varepsilon\to 0). The Lax entropy condition is precisely the extra inequality, f′(uL)>s>f′(uR)f'(u_L) > s > f'(u_R) for convex ff, that rules out the unphysical alternative and restores uniqueness.

How is the blow-up time T∗T^* actually computed in practice from given data?

Compute g(x0):=f′′(u0(x0)) u0′(x0)g(x_0) := f''(u_0(x_0))\, u_0'(x_0) at every point where u0′(x0)<0u_0'(x_0) < 0 (assuming ff convex), then set T∗(x0)=−1/g(x0)T^*(x_0) = -1/g(x_0) and take the infimum over all such x0x_0. Equivalently, T∗T^* is the reciprocal of the most negative value achieved by gg, in absolute value; the point achieving this infimum is where the first shock will appear, and its location xcx_c is obtained by plugging that x0x_0 and t=T∗t=T^* into X(t;x0)=x0+f′(u0(x0))tX(t;x_0)=x_0+f'(u_0(x_0))t.

Is the Rankine–Hugoniot condition derived, or is it an extra postulate?

It is derived, not postulated, directly from the weak (integral) formulation. Applying the divergence theorem to the integral identity over a thin pillbox straddling a smooth curve of discontinuity x=s(t)x=s(t) forces the jump in uu times the curve's speed to equal the jump in f(u)f(u) across the curve: s [u]=[f(u)]s\,[u] = [f(u)], where [⋅][\cdot] denotes the jump (right value minus left value). This is a rigorous consequence of requiring the integral identity to hold for test functions supported near the discontinuity, not an independent physical assumption.

Does the entropy condition generalize beyond scalar equations?

Yes, though the scalar Lax condition used here is the simplest instance of a much broader family of admissibility criteria (Lax conditions for systems, Liu's condition, and Kruzkhov/vanishing-viscosity entropy conditions for general convex entropy-entropy flux pairs). For systems of conservation laws such as the Euler equations of gas dynamics, the analogous conditions become considerably more delicate because there are multiple characteristic families and multiple types of waves (shocks, rarefactions, contact discontinuities), but the underlying philosophy — select the weak solution that arises as a vanishing-viscosity limit — remains the guiding principle.

Why does the theorem require u0u_0 and u0′u_0' to be bounded?

Boundedness of u0u_0 ensures f′′(u0(x0))f''(u_0(x_0)) stays within a range where the convexity bound f′′≥λf''\ge\lambda is uniformly valid (if u0u_0 were unbounded, f′′f'' might fail to stay bounded below by a fixed λ\lambda for exotic flux functions). Boundedness of u0′u_0' ensures the infimum defining T∗T^* is strictly positive rather than possibly zero: if u0′u_0' could approach −∞-\infty, the characteristic-crossing time for the steepest point could be pushed to t=0t=0, meaning no interval of classical existence at all. Both hypotheses guarantee T∗∈(0,∞)T^* \in (0,\infty) is a genuine positive finite number as claimed.

Before this

Where this leads