Domains of Dependence and Influence

If a signal is dropped into a rope at one point, when does a distant observer feel it? When are they permanently safe from ever feeling it? This is not a rhetorical question. It is the mathematical content of finite propagation speed. This is one of the sharpest qualitative differences between the wave equation and its cousins like the heat and Laplace equations. Solutions of the wave equation carry information along characteristic lines at a fixed finite speed. So the value of a solution at a given point in spacetime is completely determined by data on a bounded interval. It is never determined by data from arbitrarily far away. This locality is captured precisely by two dual geometric objects: the domain of dependence of a point (the set of initial data that can possibly affect the solution there) and the domain of influence of a set (the set of future points that a given piece of initial data can possibly affect). These regions are literally triangles and cones carved out by characteristic lines. Understanding their shape is the key that unlocks causality, uniqueness, and well-posedness for hyperbolic equations. The big picture is this: the wave equation is hyperbolic, meaning it has real, distinct characteristic directions. These characteristics are exactly the boundaries of domains of dependence and influence. This stands in stark contrast to the heat equation and Laplace's equation. Their solutions at any point depend instantaneously on data everywhere. There is no notion of "finite speed" there at all. Recognizing which equations have this causal structure matters greatly. So does computing the exact triangles/cones involved. This is essential for physics (this is literally the light cone of special relativity in the wave-equation analogy). It is essential for numerical analysis (stability conditions like CFL are domain-of-dependence statements). It is also essential for the general theory of well-posed problems. This article builds the geometric picture from d'Alembert's formula. It states and proves the domain-of-dependence theorem rigorously (via an energy argument that generalizes beyond explicit formulas). It computes concrete examples with asymmetric data. It shows exactly what fails — and why — when finite propagation speed is dropped.

Key takeaways
  • The wave equation utt=c2uxxu_{tt} = c^2 u_{xx} propagates information along characteristics x±ct=constx \pm ct = \text{const}, at exactly speed cc.
  • The domain of dependence of (x0,t0)(x_0, t_0) is the interval [x0−ct0, x0+ct0][x_0 - ct_0,\ x_0 + ct_0] on the initial line t=0t=0.
  • The domain of influence of a point x0x_0 at time 00 is the "light cone" region {(x,t):∣x−x0∣≤ct}\{(x,t): |x - x_0| \le ct\}.
  • d'Alembert's formula makes this explicit; an independent energy method proves it without needing a closed-form solution.
  • The heat equation and Laplace's equation have no finite domain of dependence: disturbances propagate instantaneously.
  • Key idea: characteristics are causal boundaries — solution values only "see" initial data inside the backward characteristic triangle, never outside it.

Where this lives today

In numerical analysis, the Courant–Friedrichs–Lewy (CFL) condition for explicit finite-difference schemes is precisely a discrete domain-of-dependence requirement: the numerical stencil's domain of dependence must contain the true PDE's domain of dependence, or the scheme is unconditionally unstable. In special and general relativity, light cones are the exact physical incarnation of domains of influence for the wave equation governing electromagnetic and gravitational signals, encoding the impossibility of faster-than-light communication. In control theory and signal processing, finite propagation speed underlies causality constraints on filters and transfer functions. In seismology and acoustics, the leading and trailing edges of wavefronts (and, in three dimensions, the sharp Huygens' wake) are directly the boundary of the domain of influence of a localized source such as an earthquake rupture. In the general theory of PDE well-posedness — the subject this material feeds directly into — domains of dependence give the geometric mechanism behind uniqueness and continuous dependence for hyperbolic Cauchy problems, feeding into Well-Posedness of PDE Problems and contrasting sharply with the global-dependence behavior of elliptic and parabolic equations, which is why the classification of a PDE as hyperbolic, parabolic, or elliptic is inseparable from understanding d'Alembert's Solution of the Wave Equation and the broader theory of Characteristics of PDEs.

:::quiz[If initial data is supported only on [−2,1][-2,1] and c=5c=5, is the point (x,t)=(20,3)(x,t)=(20,3) inside the domain of influence?]

  • Yes, since 20≤5(3)+1=1620 \le 5(3) + 1 = 16 is false but influence extends further
  • No, since 1+5(3)=16<201 + 5(3) = 16 < 20 ✓
  • Yes, because all points eventually lie in the domain of influence
  • Cannot be determined without knowing the exact data values

:::faq[Can the domain of dependence be smaller than the full interval [x0−ct0,x0+ct0][x_0-ct_0, x_0+ct_0]?] Not in general for the constant-coefficient wave equation — every point of that interval genuinely can affect u(x0,t0)u(x_0,t_0) for suitable choices of data (take gg a narrow bump centered at any interior point of the interval and check the integral term is nonzero). So the stated domain of dependence is sharp: it is not merely an upper bound but the exact set of points whose data matter, in the sense that perturbing gg at any single interior point changes u(x0,t0)u(x_0,t_0). The two endpoints are special: only the pointwise value of ff there matters (not its derivatives), which is visible directly in d'Alembert's formula.

Ask the tutor: How does the domain-of-dependence argument change if the wave equation has a variable coefficient wave speed c(x)c(x), and what happens to the characteristic triangle then?

DefinitionDomain of dependence and domain of influence

Consider the one-dimensional wave equation on R×[0,∞)\mathbb{R} \times [0,\infty),

utt(x,t)=c2uxx(x,t),c>0,u_{tt}(x,t) = c^2 u_{xx}(x,t), \qquad c > 0,

with initial data

u(x,0)=f(x),ut(x,0)=g(x).u(x,0) = f(x), \qquad u_t(x,0) = g(x).

Domain of dependence. For a fixed point (x0,t0)(x_0, t_0) with t0>0t_0 > 0, the domain of dependence is the closed interval on the initial line

D(x0,t0):=[ x0−ct0, x0+ct0 ]⊂{t=0}.D(x_0,t_0) := \big[\, x_0 - c t_0,\ x_0 + c t_0 \,\big] \subset \{t=0\}.

It is the set of points xx such that the values f(x),g(x)f(x), g(x) can, in principle, influence u(x0,t0)u(x_0,t_0).

Domain of influence. For a fixed point x0x_0 on the initial line, the domain of influence is the region in the upper half-plane

I(x0):={ (x,t):t≥0, ∣x−x0∣≤ct }.I(x_0) := \big\{\, (x,t) : t \ge 0,\ |x - x_0| \le c t \,\big\}.

It is the set of spacetime points whose solution value can, in principle, be affected by data at x0x_0. More generally, for a set E⊂{t=0}E \subset \{t=0\} (e.g. the support of the data), its domain of influence is

I(E):=⋃x0∈EI(x0),I(E) := \bigcup_{x_0 \in E} I(x_0),

and the two notions are dual: (x,t)∈I(E)  ⟺  E∩D(x,t)≠∅(x,t) \in I(E) \iff E \cap D(x,t) \neq \emptyset. The boundary lines x−x0=±ctx - x_0 = \pm c t (equivalently x−x0=±c(t−t0)x - x_0 = \pm c(t-t_0) through a point (x0,t0)(x_0,t_0)) are called characteristics of the wave equation; they are the causal light rays of the problem.

Intuition

Picture the (x,t)(x,t)-plane with t≥0t \ge 0 as time running upward. Through any point (x0,t0)(x_0, t_0), draw the two characteristic lines of slope ±1/c\pm 1/c running backward to t=0t = 0. These two lines, together with the segment of the initial axis between their feet, bound a triangle — the characteristic triangle. d'Alembert's formula says explicitly that u(x0,t0)u(x_0,t_0) depends only on ff at the two triangle vertices and on gg integrated over the base of the triangle; nothing outside that base can possibly matter. This triangle is the domain of dependence, drawn out in spacetime rather than just recorded as an interval. Now flip the perspective: fix a single point x0x_0 (or a whole interval) on the initial line and ask which future points it can affect. Draw the two characteristics through x0x_0 going forward in time; the wedge between them, opening upward, is the domain of influence — the two-dimensional analogue of a light cone. If the initial data f,gf,g is supported only on some interval [a,b][a,b], the domain of influence is the union of wedges from every point of [a,b][a,b], which forms a trapezoidal "light cone" region: a solid triangle-with-flared-sides expanding at speed cc in both directions. Outside this expanding cone, the solution is identically zero for all time — the disturbance simply has not arrived yet, and once it has passed (in problems where there's a sharp trailing edge, as in odd space dimensions) it may leave silence behind again.

Figure
The backward characteristic triangle determining u(x0,t0)u(x_0,t_0)
x t 0 t = 0 (data f, g) (x₀, t₀) x₀ - ct₀ x₀ + ct₀ D(x₀,t₀) x + ct = x₀ + ct₀ x - ct = x₀ - ct₀
TheoremDomain of Dependence Theorem for the 1D Wave Equation

Let c>0c>0. Suppose u∈C2(R×[0,∞))u \in C^2(\mathbb{R}\times[0,\infty)) solves

utt=c2uxxon R×(0,∞),u_{tt} = c^2 u_{xx} \quad \text{on } \mathbb{R}\times(0,\infty),

with initial data

u(x,0)=f(x),ut(x,0)=g(x),x∈R,u(x,0) = f(x), \qquad u_t(x,0) = g(x), \qquad x \in \mathbb{R},

where f∈C2(R)f \in C^2(\mathbb{R}) and g∈C1(R)g \in C^1(\mathbb{R}). Fix (x0,t0)(x_0,t_0) with t0>0t_0>0 and let

D(x0,t0)=[x0−ct0, x0+ct0].D(x_0,t_0) = [x_0 - ct_0,\ x_0+ct_0].

Then u(x0,t0)u(x_0,t_0) is completely determined by f∣D(x0,t0)f\big|_{D(x_0,t_0)} and g∣D(x0,t0)g\big|_{D(x_0,t_0)}. Precisely: 1. (Explicit form.) u(x0,t0)u(x_0,t_0) is given by d'Alembert's formula

u(x0,t0)=f(x0−ct0)+f(x0+ct0)2+12c∫x0−ct0x0+ct0g(s) ds,u(x_0,t_0) = \frac{f(x_0-ct_0) + f(x_0+ct_0)}{2} + \frac{1}{2c}\int_{x_0-ct_0}^{x_0+ct_0} g(s)\,ds,

which manifestly depends only on data inside D(x0,t0)D(x_0,t_0). 2. (Uniqueness / stability consequence.) If u~\tilde u is another C2C^2 solution with initial data f~,g~\tilde f, \tilde g satisfying f~=f\tilde f = f and g~=g\tilde g = g on D(x0,t0)D(x_0,t_0), then u~(x0,t0)=u(x0,t0)\tilde u(x_0,t_0) = u(x_0,t_0), regardless of how f~,g~\tilde f,\tilde g differ from f,gf,g outside D(x0,t0)D(x_0,t_0).

Proof

We give two independent proofs of part (2) — one directly from the explicit formula, and one by an energy argument that does not use the formula at all. The energy method is the one that generalizes to variable coefficients, higher dimensions, and nonlinear hyperbolic systems where no closed form exists, so we present it in full.

Proof via d'Alembert's formula. Since uu is a C2C^2 solution of the wave equation with the given data, d'Alembert's formula (established from the standard factorization ∂t2−c2∂x2=(∂t−c∂x)(∂t+c∂x)\partial_t^2 - c^2\partial_x^2 = (\partial_t - c\partial_x)(\partial_t + c\partial_x) and integration along characteristics — see d'Alembert's Solution of the Wave Equation) gives

u(x0,t0)=f(x0−ct0)+f(x0+ct0)2+12c∫x0−ct0x0+ct0g(s) ds.u(x_0,t_0) = \frac{f(x_0-ct_0)+f(x_0+ct_0)}{2} + \frac{1}{2c}\int_{x_0-ct_0}^{x_0+ct_0} g(s)\,ds .

The right side involves ff evaluated only at the two endpoints x0±ct0x_0 \pm ct_0 of D(x0,t0)D(x_0,t_0), and gg integrated only over D(x0,t0)D(x_0,t_0). Hence if f~=f\tilde f = f and g~=g\tilde g = g pointwise on D(x0,t0)D(x_0,t_0), the same formula applied to u~\tilde u gives the identical numerical value, so u~(x0,t0)=u(x0,t0)\tilde u(x_0,t_0) = u(x_0,t_0). This proves part (2) directly, and part (1) is just the display itself. ■\blacksquare

Proof via the energy method (self-contained, formula-free). This proof is the one worth understanding deeply, because it is the template used for the multidimensional wave equation and for hyperbolic PDE with variable coefficients, where no explicit formula is available. Let w:=u~−uw := \tilde u - u. By linearity, ww solves the wave equation wtt=c2wxxw_{tt} = c^2 w_{xx} with initial data

w(x,0)=f~(x)−f(x)=:h(x),wt(x,0)=g~(x)−g(x)=:k(x),w(x,0) = \tilde f(x) - f(x) =: h(x), \qquad w_t(x,0) = \tilde g(x) - g(x) =: k(x),

and by hypothesis h≡0h \equiv 0 and k≡0k \equiv 0 on D(x0,t0)=[x0−ct0,x0+ct0]D(x_0,t_0) = [x_0-ct_0, x_0+ct_0]. We must show w(x0,t0)=0w(x_0,t_0) = 0. Step 1: the backward light-cone domain. For 0≤t≤t00 \le t \le t_0, define the shrinking interval

I(t):=[x0−c(t0−t), x0+c(t0−t)],I(t) := [x_0 - c(t_0 - t),\ x_0 + c(t_0-t)],

which shrinks from D(x0,t0)D(x_0,t_0) at t=0t=0 down to the single point {x0}\{x_0\} at t=t0t = t_0. Define the local energy

E(t):=12∫I(t)(wt(x,t)2+c2wx(x,t)2)dx.E(t) := \frac{1}{2}\int_{I(t)} \left( w_t(x,t)^2 + c^2 w_x(x,t)^2 \right) dx .

Since I(t0)={x0}I(t_0) = \{x_0\} has zero length, E(t0)=0E(t_0) = 0 automatically once we know EE makes sense; the content of the argument is to show EE is nonincreasing, so that E(0)=0E(0) = 0 forces w≡0w \equiv 0 on the whole triangle. Step 2: E(0)=0E(0) = 0. On I(0)=D(x0,t0)I(0) = D(x_0,t_0) we are given w(x,0)=h(x)≡0w(x,0) = h(x) \equiv 0, so differentiating along t=0t=0, wx(x,0)=h′(x)≡0w_x(x,0) = h'(x) \equiv 0 there too. Also wt(x,0)=k(x)≡0w_t(x,0) = k(x) \equiv 0 on I(0)I(0) by hypothesis. Hence the integrand of E(0)E(0) vanishes identically on I(0)I(0), giving E(0)=0E(0) = 0. Step 3: EE is nonincreasing. Differentiate under the integral sign (valid since w∈C2w \in C^2 and the endpoints of I(t)I(t) move smoothly with speed cc, using Leibniz's rule for a moving domain):

E′(t)=∫I(t)(wtwtt+c2wxwxt) dx  +  12[(wt2+c2wx2) r˙(t)]endpoints,E'(t) = \int_{I(t)} \big( w_t w_{tt} + c^2 w_x w_{xt} \big)\, dx \;+\; \frac{1}{2}\Big[ (w_t^2 + c^2 w_x^2)\, \dot{r}(t) \Big]_{\text{endpoints}},

where the boundary term comes from the two endpoints of I(t)I(t) moving inward at speed r˙(t)=c\dot r(t) = c (left endpoint moves right at rate +c+c, right endpoint moves left at rate −c-c; both contribute a factor with the same sign after accounting for orientation — we make this precise next). Write I(t)=[α(t),β(t)]I(t) = [\alpha(t), \beta(t)] with α(t)=x0−c(t0−t)\alpha(t) = x_0 - c(t_0-t), β(t)=x0+c(t0−t)\beta(t) = x_0 + c(t_0-t), so α′(t)=c\alpha'(t) = c and β′(t)=−c\beta'(t) = -c. By the Leibniz rule,

E′(t)=∫α(t)β(t)(wtwtt+c2wxwxt) dx+12(wt2+c2wx2)∣x=β(t)β′(t)−12(wt2+c2wx2)∣x=α(t)α′(t).E'(t) = \int_{\alpha(t)}^{\beta(t)} \big(w_t w_{tt} + c^2 w_x w_{xt}\big)\,dx + \frac{1}{2}(w_t^2+c^2w_x^2)\big|_{x=\beta(t)}\beta'(t) - \frac{1}{2}(w_t^2+c^2w_x^2)\big|_{x=\alpha(t)}\alpha'(t).

Since wtt=c2wxxw_{tt} = c^2 w_{xx}, the integrand of the bulk term is

wtwtt+c2wxwxt=c2(wtwxx+wxwxt)=c2∂x(wtwx),w_t w_{tt} + c^2 w_x w_{xt} = c^2\big(w_t w_{xx} + w_x w_{xt}\big) = c^2 \partial_x\big(w_t w_x\big),

which integrates exactly:

∫α(t)β(t)c2∂x(wtwx) dx=c2[wtwx]α(t)β(t)=c2(wtwx)∣β(t)−c2(wtwx)∣α(t).\int_{\alpha(t)}^{\beta(t)} c^2 \partial_x(w_t w_x)\, dx = c^2\Big[ w_t w_x \Big]_{\alpha(t)}^{\beta(t)} = c^2\big(w_t w_x\big)\big|_{\beta(t)} - c^2\big(w_t w_x\big)\big|_{\alpha(t)}.

Substituting β′(t)=−c\beta'(t) = -c, α′(t)=c\alpha'(t) = c, we get

E′(t)=c2(wtwx)∣β−c2(wtwx)∣α−c2(wt2+c2wx2)∣β−c2(wt2+c2wx2)∣α.E'(t) = c^2(w_tw_x)|_{\beta} - c^2(w_tw_x)|_\alpha - \frac{c}{2}(w_t^2+c^2w_x^2)|_\beta - \frac{c}{2}(w_t^2+c^2w_x^2)|_\alpha .

Group the terms at x=β(t)x=\beta(t):

−c2(wt2−2c wtwx+c2wx2)∣β=−c2(wt−cwx)2∣β≤0,-\frac{c}{2}\Big(w_t^2 - 2c\,w_tw_x + c^2w_x^2\Big)\Big|_\beta = -\frac{c}{2}\big(w_t - c w_x\big)^2\Big|_\beta \le 0,

and the terms at x=α(t)x = \alpha(t):

−c2(wt2+2c wtwx+c2wx2)∣α=−c2(wt+cwx)2∣α≤0.-\frac{c}{2}\Big(w_t^2 + 2c\,w_tw_x + c^2w_x^2\Big)\Big|_\alpha = -\frac{c}{2}\big(w_t + c w_x\big)^2\Big|_\alpha \le 0.

Both are perfect squares with a minus sign, hence

E′(t)=−c2(wt−cwx)2∣x=β(t)−c2(wt+cwx)2∣x=α(t)≤0.E'(t) = -\frac{c}{2}(w_t - cw_x)^2\Big|_{x=\beta(t)} - \frac{c}{2}(w_t+cw_x)^2\Big|_{x=\alpha(t)} \le 0 .

So EE is nonincreasing on [0,t0][0,t_0]. Step 4: conclude. Since E(0)=0E(0) = 0 (Step 2), E≥0E \ge 0 everywhere (it is an integral of a sum of squares), and EE is nonincreasing (Step 3), we get E(t)=0E(t) = 0 for all t∈[0,t0]t \in [0,t_0]. Because the integrand wt2+c2wx2w_t^2 + c^2w_x^2 is continuous and nonnegative and integrates to zero over I(t)I(t) for every tt, it vanishes identically: wt(x,t)=wx(x,t)=0w_t(x,t) = w_x(x,t) = 0 for all (x,t)(x,t) in the closed triangle

T:={(x,t):0≤t≤t0, x∈I(t)}.\mathcal{T} := \{(x,t): 0 \le t \le t_0,\ x \in I(t)\}.

Hence ww is constant on the connected set T\mathcal{T}. Since w(x,0)=h(x)=0w(x,0) = h(x) = 0 for x∈I(0)x \in I(0), that constant is 00. In particular (x0,t0)∈T(x_0,t_0) \in \mathcal{T} (it is the apex, I(t0)={x0}I(t_0) = \{x_0\}), so w(x0,t0)=0w(x_0,t_0) = 0, i.e. u~(x0,t0)=u(x0,t0)\tilde u(x_0,t_0) = u(x_0,t_0). ■\blacksquare

∎

Worked examples

Example 1 — an asymmetric plucked-and-struck string. Take c=3c = 3 and initial data

f(x)={1−∣x−4∣∣x−4∣≤10otherwise,g(x)={20≤x≤10otherwise.f(x) = \begin{cases} 1 - |x-4| & |x-4|\le 1 \\ 0 & \text{otherwise}\end{cases}, \qquad g(x) = \begin{cases} 2 & 0\le x \le 1 \\ 0 & \text{otherwise.}\end{cases}

So ff is a small tent bump supported on [3,5][3,5] (centered away from the origin) and gg is a rectangular velocity pulse of height 22 supported on [0,1][0,1] (a different, non-overlapping location — this asymmetry matters, so we can watch the two pieces of data travel independently). We ask: does u(10,2)u(10, 2) depend on either piece of data? The domain of dependence is

D(10,2)=[10−3⋅2, 10+3⋅2]=[4,16].D(10,2) = [10 - 3\cdot 2,\ 10+3\cdot2] = [4, 16].

Check overlap with each support:

  • supp(f)=[3,5]\mathrm{supp}(f) = [3,5]. Intersection with [4,16][4,16] is [4,5][4,5], nonempty. So ff can affect u(10,2)u(10,2) — but only through its values at the two characteristic feet x0−ct0=4x_0-ct_0 = 4 and x0+ct0=16x_0+ct_0=16 by d'Alembert's formula, and f(16)=0f(16) = 0 while f(4)=1−∣4−4∣=1f(4) = 1-|4-4| = 1. So ff contributes via its value at x=4x=4 only.
  • supp(g)=[0,1]\mathrm{supp}(g) = [0,1]. Intersection with [4,16][4,16] is empty. So gg contributes nothing to u(10,2)u(10,2), even though the integral term in d'Alembert's formula would ordinarily involve gg — here that integral is ∫416g(s) ds=0\int_4^{16} g(s)\,ds = 0 since gg vanishes there entirely.

We can compute u(10,2)u(10,2) exactly:

u(10,2)=f(4)+f(16)2+12⋅3∫416g(s) ds=1+02+16(0)=12.u(10,2) = \frac{f(4) + f(16)}{2} + \frac{1}{2\cdot 3}\int_4^{16} g(s)\,ds = \frac{1+0}{2} + \frac{1}{6}(0) = \frac{1}{2}.

Notice the qualitative punch line: even though gg is "closer" to the origin than ff is far, it is gg that fails to reach (10,2)(10,2) in time (the velocity pulse needs to travel from [0,1][0,1], and even its rightmost point x=1x=1 would need to reach x=10x=10 or x=−...x=-... within time 22 at speed 33, covering at most distance 66 — from x=1x=1 that reaches only to x=7x=7, short of 1010). Meanwhile ff's right edge at x=5x=5 combined with leftward propagation reaches 5−3(2)=−15 - 3(2) = -1 to 5+3(2)=115+3(2) = 11, and its left foot at x=4x=4 sits exactly inside [4,16][4,16], so part of ff does arrive. This is exactly the content of the domain-of-dependence theorem: only overlap with D(x0,t0)D(x_0,t_0) matters, and the overlap can be partial and asymmetric between the two pieces of data f,gf,g.

Example 2 — domain of influence of an off-center, asymmetric pulse. Now let c=2c=2 and suppose f≡0f\equiv 0 but gg is supported on the asymmetric interval [−1,4][-1, 4] (length 5, not centered at the origin), with gg otherwise arbitrary (nonzero somewhere in that interval, say g(x)=x(4−x)(x+1)g(x) = x(4-x)(x+1) there, zero outside — the exact shape does not matter for the influence region, only the support does). The domain of influence of the support E=[−1,4]E = [-1,4] is

I(E)=⋃x0∈[−1,4]{(x,t):t≥0, ∣x−x0∣≤2t}.I(E) = \bigcup_{x_0 \in [-1,4]} \{(x,t): t\ge 0,\ |x-x_0|\le 2t\}.

Geometrically this is bounded by:

  • the leftmost characteristic emanating from the left end of EE, going leftward: x=−1−2tx = -1 - 2t;
  • the rightmost characteristic emanating from the right end of EE, going rightward: x=4+2tx = 4 + 2t.

So explicitly,

I(E)={(x,t):t≥0, −1−2t≤x≤4+2t}.I(E) = \{(x,t) : t \ge 0,\ -1-2t \le x \le 4+2t\}.

Check a specific point: is (x,t)=(−5,1.5)(x,t) = (-5, 1.5) inside the cone? We need −1−2(1.5)=−4≤−5-1 - 2(1.5) = -4 \le -5? No: −5<−4-5 < -4, so (−5,1.5)∉I(E)(-5,1.5) \notin I(E), meaning u(−5,1.5)=0u(-5,1.5) = 0 regardless of the details of gg, because the disturbance from [−1,4][-1,4] has only had time to reach as far left as x=−4x=-4 by t=1.5t=1.5. Now check (x,t)=(−3,1.5)(x,t) = (-3,1.5): is −4≤−3≤4+3=7-4 \le -3 \le 4+3=7? Yes, so this point lies inside the influence cone and u(−3,1.5)u(-3,1.5) may be (and generically will be) nonzero. We can verify this last claim using duality with the domain of dependence: (x,t)=(−3,1.5)∈I(E)(x,t)=(-3,1.5) \in I(E) iff D(−3,1.5)∩E≠∅D(-3,1.5) \cap E \ne \emptyset. Compute D(−3,1.5)=[−3−3,−3+3]=[−6,0]D(-3,1.5) = [-3-3, -3+3] = [-6,0]. Indeed [−6,0]∩[−1,4]=[−1,0]≠∅[-6,0]\cap[-1,4] = [-1,0] \ne \emptyset, confirming consistency between the two dual definitions.

:::svg[Domain of influence of an off-center pulse on [−1,4][-1,4], speed c=2c=2] x t E = [-1,4] x = -1-2t x = 4+2t (-3, 1.5) (-5,1.5) outside -1 4

CounterexampleWhy finite propagation speed needs hyperbolicity — the heat equation

The domain-of-dependence theorem relies crucially on the equation being the wave equation — specifically on it being hyperbolic with real characteristic speeds ±c\pm c. Drop this and replace the wave equation with the heat equation,

ut=uxx,u(x,0)=f(x),u_t = u_{xx}, \qquad u(x,0) = f(x),

and finite propagation speed fails completely, even for compactly supported, smooth, nonnegative data. Take f(x)=e−x2f(x) = e^{-x^2} (a smooth Gaussian, or for compact support take ff any smooth bump supported on [−1,1][-1,1], nonnegative, not identically zero). The solution is given by convolution with the heat kernel,

u(x,t)=∫−∞∞14πt e−(x−y)2/4t f(y) dy,t>0.u(x,t) = \int_{-\infty}^{\infty} \frac{1}{\sqrt{4\pi t}}\, e^{-(x-y)^2/4t}\, f(y)\, dy, \qquad t>0.

If f≥0f \ge 0 and ff is not identically zero on some set of positive measure, then for every xx and every t>0t>0, however large ∣x∣|x| is and however small tt is,

u(x,t)=∫14πte−(x−y)2/4tf(y) dy>0,u(x,t) = \int \frac{1}{\sqrt{4\pi t}} e^{-(x-y)^2/4t} f(y)\, dy > 0,

because the heat kernel 14πte−(x−y)2/4t\frac{1}{\sqrt{4\pi t}}e^{-(x-y)^2/4t} is strictly positive for every yy, so the integrand is strictly positive on the support of ff, giving a strictly positive integral. Concretely: with ff supported on [−1,1][-1,1] and x=106x = 10^6, t=10−9t = 10^{-9}, the value u(106,10−9)u(10^6, 10^{-9}) is an unimaginably small but strictly positive number — the "signal" from [−1,1][-1,1] reaches x=106x=10^6 instantaneously (in fact at every positive time, however small). There is no interval D(x0,t0)D(x_0,t_0) outside of which data can be ignored: the value of uu at any point depends on the entire real line's worth of initial data, for any t>0t>0. This is the precise failure of the hypothesis "the PDE is hyperbolic with finite characteristic speeds" — the heat equation is parabolic, its only "characteristic direction" is t=constt=\text{const} (a degenerate double characteristic), and no backward light cone exists. The same instantaneous-influence phenomenon occurs for Laplace's equation, where by contrast there is no time variable at all and boundary data everywhere on ∂Ω\partial\Omega affects the harmonic function at every interior point simultaneously (via the Poisson kernel, which again is everywhere strictly positive).

Scope, edge cases & common traps

SituationWhat holdsWhat can go wrong
1D wave equation, f∈C2f\in C^2, g∈C1g\in C^1Domain of dependence theorem holds exactly as stated; d'Alembert's formula appliesIf ff only C0C^0 or gg only Lloc1L^1_{loc}, formula still defines a weak solution but C2C^2 regularity claims fail
Wave equation with a source term utt−c2uxx=F(x,t)u_{tt}-c^2u_{xx}=F(x,t)Domain of dependence becomes the full triangle (source contributes via a double integral over the triangle, not just its base)Forgetting the source term entirely changes the dependence set from an interval to a 2D region
Variable wave speed c(x)c(x)Energy method generalizes (with care) but characteristics are curved, not straight linesAssuming straight-line light cones when speed is non-constant
Heat equation / Laplace's equationNo finite domain of dependence at allMistakenly importing "finite propagation speed" intuition from the wave equation
Higher space dimensions (n=3n=3)Domain of dependence is a ball, and sharp Huygens' principle holds: influence is exactly on the boundary sphereIn n=2n=2, no sharp Huygens' principle — influence persists inside the whole disk, not just its boundary circle (a genuinely different phenomenon, still finite-speed but with a "wake")

The mathematician behind it

Origins · 1 minJean le Rond d'Alembert and how this was discovered
JL
Jean le Rond d'Alembert
1717–1783 · Paris, Enlightenment mathematical physics

D'Alembert derived the explicit traveling-wave solution of the vibrating string equation in 1747, in the course of a debate with Euler and Daniel Bernoulli over what kinds of functions could represent a plucked string. His formula — expressing the displacement as a sum of a left-moving and a right-moving wave — is the computational engine behind the entire theory of domains of dependence and influence for the one-dimensional wave equation. The underlying geometric idea, that information travels along characteristic lines at finite speed, became one of the conceptual seeds of the modern theory of hyperbolic PDE and, much later, of the light-cone structure central to relativity.

History
- 1747 — d'Alembert derives the traveling-wave solution of the vibrating string equation, the origin of the explicit formula.
- 1749–1753 — Euler and Daniel Bernoulli dispute the class of admissible initial shapes, sharpening what "solution" should mean.
- 1860s — Riemann develops the method of characteristics systematically for hyperbolic systems, generalizing the triangle/cone picture.
- 1920s — Jacques Hadamard formalizes well-posedness, tying domains of dependence directly to continuous dependence on data.
- 1950s — Courant, Friedrichs, and Lewy's stability condition for numerical schemes is recognized as a discrete domain-of-dependence requirement.
Is the domain of dependence always exactly an interval of length 2ct02ct_0?

Yes, for the one-dimensional constant-coefficient wave equation with speed cc, D(x0,t0)=[x0−ct0,x0+ct0]D(x_0,t_0) = [x_0-ct_0, x_0+ct_0] always has length exactly 2ct02ct_0, growing linearly with t0t_0. This reflects the fact that both characteristic families travel at exactly speed cc in either direction, so the backward light cone from (x0,t0)(x_0,t_0) widens at a constant rate as you trace it down to t=0t=0. If the wave speed is not constant (e.g. c=c(x)c=c(x)), the domain of dependence is still an interval, but its endpoints are determined by solving the (now curved) characteristic ODEs dx/dt=±c(x)dx/dt = \pm c(x) backward from (x0,t0)(x_0,t_0), rather than by the simple linear formula.

Does the domain-of-dependence theorem say u(x0,t0)u(x_0,t_0) equals zero outside the interval?

No — this is the most common misreading. The theorem is about which data can influence the value, not about the value itself being zero anywhere. The correct dual statement about vanishing is the domain-of-influence direction: if the initial data f,gf,g vanish outside some set EE, then u(x,t)=0u(x,t) = 0 for all (x,t)(x,t) outside I(E)I(E), the forward light cone generated by EE. These are two separate, dual statements, and conflating "DD" (an interval on the initial line, about inputs) with "II" (a region in spacetime, about outputs) is exactly the trap flagged in the warning above.

How does the domain of dependence change in three spatial dimensions?

In 3D, for the wave equation utt=c2Δuu_{tt} = c^2\Delta u, the domain of dependence of a point (x0,t0)(x_0,t_0) is the closed ball Bˉ(x0,ct0)\bar B(x_0, ct_0) at t=0t=0. But something sharper happens: Kirchhoff's formula shows u(x0,t0)u(x_0,t_0) actually depends only on data on the boundary sphere ∂B(x0,ct0)\partial B(x_0,ct_0) (for the position data ff, differentiated) together with the velocity data gg averaged over that same sphere — not the full solid ball. This sharp version is called the (strong) Huygens' principle, and it fails in 2D, where the dependence genuinely fills the whole disk (a "wake" persists after the wavefront passes), making odd dimensions qualitatively different from even ones.

Why does the energy method matter if d'Alembert's formula already proves everything in 1D?

The explicit formula is special to the 1D constant-coefficient case; it does not exist for variable wave speed c(x,t)c(x,t), for the multidimensional wave equation on general domains, or for nonlinear hyperbolic systems (e.g. gas dynamics). The energy method proves the same domain-of-dependence conclusion using only the PDE structure and integration by parts, with no formula required, which is exactly why it is the technique that survives in the general theory of hyperbolic PDE and underlies the broader study of Well-Posedness of PDE Problems.

Is finite propagation speed equivalent to hyperbolicity?

Within the classical second-order linear classification (hyperbolic/parabolic/elliptic), finite propagation speed is a hallmark of hyperbolicity, but the equivalence should be stated carefully: hyperbolic equations with real, distinct characteristic roots have finite propagation speed bounded by the largest characteristic speed, while parabolic and elliptic equations have infinite propagation speed (as the heat-equation counterexample shows). The precise general statement, covering variable-coefficient and systems cases, is part of the broader theory of characteristics and energy estimates, and it is one of the organizing principles used when classifying any given PDE before attempting to solve it.

How does this relate to the CFL condition in numerical methods?

An explicit finite-difference scheme for the wave equation updates the grid value at (x0,t0+Δt)(x_0, t_0+\Delta t) using only grid values in some fixed numerical stencil at time t0t_0. The Courant–Friedrichs–Lewy condition requires that this numerical domain of dependence contain the true PDE domain of dependence, i.e. Δt≤Δx/c\Delta t \le \Delta x / c for standard explicit schemes; violating it means the scheme is trying to predict a value using less information than the true solution actually requires, which provably causes instability. This is a direct, quantitative descendant of the theorem proved here.

Before this

Where this leads