Laplace Transform Methods for PDEs

How does heat move down a long metal rod when you suddenly heat one end at time zero? This is exactly the kind of question where the Fourier Transform Methods for PDEs toolkit runs into a wall. That toolkit is built for problems on the whole line or with periodic data. Real experiments live on a half-line or finite interval. They start at a definite instant. They have boundary conditions imposed for all future time. The Laplace transform is the natural answer. It is designed precisely for functions defined on $t \ge 0$ . It converts initial-value problems in time into algebraic or ordinary differential problems in a complex frequency variable $s$ . The idea generalizes what you already know from ODEs. Transforming $\partial_t$ into multiplication by $s$ (minus initial data) turns a linear PDE into an ODE. That ODE is in the spatial variable, parametrized by $s$ . Solve that ODE, impose the spatial boundary conditions. Then invert the transform to recover $u(x,t)$ . This works cleanly for the heat equation, the wave equation on a half-line, and many diffusion-convection-reaction models. It is the standard route to explicit formulas involving the complementary error function and Duhamel-type convolution integrals. The big picture is a trade of one hard problem for two easier ones. First, an algebraic/ODE problem in $(x,s)$ . Second, a complex contour integral (the Bromwich inversion) to get back to $(x,t)$ . In practice the inversion step is often bypassed using tables, the convolution theorem, or asymptotic methods. But knowing that the rigorous inversion exists — and under what hypotheses — matters greatly. That is what makes the whole method trustworthy rather than a bag of tricks. This article gives the precise definition of the Laplace transform in time for a PDE. It states and proves the transform rules needed (differentiation, shifting, convolution). It proves the Bromwich inversion theorem under a clean hypothesis set. It carries a full worked solution of the heat equation on a semi-infinite rod. This worked solution proceeds through to the classical error-function answer.

Key takeaways

The Laplace transform in $t$ turns $\partial_t u$ into $sU - u(x,0)$ , converting an evolution PDE into an ODE (or algebraic equation) in $x$ , parametrized by $s\in\mathbb{C}$ .
It is the natural tool when the time domain is $t\ge 0$ with prescribed initial data and boundary conditions imposed for all $t>0$ — unlike Fourier Transform Methods for PDEs, which prefers the whole line or periodicity.
Solving the transformed ODE plus boundedness/boundary conditions gives $U(x,s)$ explicitly; recovering $u(x,t)$ is the Bromwich (Mellin) inversion integral, a contour integral in the complex $s$ -plane.
For the heat equation on a half-line with constant boundary temperature, this machinery reproduces the classical error-function solution exactly and rigorously.
The method requires functions of at most exponential growth in $t$ (transform existence) and analytic continuation of $U(x,s)$ to a right half-plane; dropping this breaks invertibility.
It sits between d'Alembert's Solution of the Wave Equation (explicit, transform-free) and the general theory of Well-Posedness of PDE Problems, which asks whether such solutions are unique and stable.
Key idea: transform time away, solve an ODE in space, then invert a complex contour integral to restore the time dependence.

Where this lives today

Laplace-transform methods for PDEs are the standard classical technique for heat conduction with time-dependent boundary flux or temperature in engineering heat-transfer texts, giving explicit error-function and convolution solutions used in metallurgy and building-thermal design. In electrical engineering and control theory, the same method (there usually applied to the telegrapher's/transmission-line equations, a damped wave equation) computes voltage and current transients on cables and is the origin of the "operational calculus" still taught alongside transfer functions and Laplace-domain circuit analysis. In financial mathematics, Laplace transforms in time solve the Black-Scholes PDE and related barrier-option problems, where the inversion integral becomes a contour integral evaluated by residues or numerical inverse-Laplace algorithms (e.g. the Talbot or Stehfest methods). In numerical analysis, this same viewpoint underlies "Laplace transform methods for evolution equations" used to solve stiff parabolic PDEs by evaluating the resolvent $(\mathcal A - sI)^{-1}$ at a handful of contour points rather than time-stepping — a technique tied to semigroup theory and functional-analytic treatments of Well-Posedness of PDE Problems. Finally, the transform's algebraic manipulation of differential operators (turning $\partial_t$ into $s$ ) is the direct ancestor of the symbolic "operator calculus" used throughout applied mathematics and signal processing, complementing the Fourier Transform Methods for PDEs used for spatial or periodic-in-time problems.

Ask the tutor: How does the choice of branch for $\sqrt{s}$ in $U(x,s)=G(s)e^{-x\sqrt{s/k}}$ affect the validity of the Bromwich inversion contour?

(x,t)

Intuition

Think of the Laplace transform as a weighted "totaling up" of a signal's future: multiplying $u(x,t)$ by the decaying exponential $e^{-st}$ and integrating over all $t \ge 0$ produces a single number (for each $x$ and $s$ ) that packages the whole time history. Differentiating in $t$ becomes, after integration by parts, multiplication by $s$ — but with a correction term that is exactly the initial data $u(x,0)$ . This is the miracle that makes initial-value problems tractable: the transform automatically carries the initial condition into the algebra. Geometrically, picture $u(x,t)$ as a stack of time-signals, one for each $x$ . The transform flattens each vertical time-slice into a complex number $U(x,s)$ , turning a genuinely two-variable PDE into a family of ODEs in $x$ , one for each value of $s$ in a right half-plane of the complex plane. Solving that ODE with the spatial boundary data pins down $U(x,s)$ completely as a function of $s$ (often involving $\sqrt{s}$ or $e^{-x\sqrt{s/k}}$ for diffusion problems). The final step, recovering $u(x,t)$ from $U(x,s)$ , is not a simple algebraic reversal — it requires integrating $U(x,s)e^{st}$ along a vertical line in the complex plane far enough to the right that $U$ is analytic, then closing or deforming the contour.

Figure

The Bromwich contour: a vertical line

\operatorname{Re}(s)=c

to the right of all singularities of

U(x,s)

Setting up the ODE in space

For concreteness, fix the heat equation $u_t = k u_{xx}$ on $x>0$ , $t>0$ , with $u(x,0)=0$ and $u(0,t)=g(t)$ . Transforming in $t$ and writing $U(x,s)=\mathcal L\{u\}(x,s)$ , $G(s) = \mathcal L\{g\}(s)$ , the equation becomes the ODE $sU = kU_{xx}$ , i.e.

U_{xx} - \frac{s}{k}U = 0, \qquad U(0,s) = G(s),

with general solution $U(x,s) = A(s)e^{-x\sqrt{s/k}} + B(s)e^{x\sqrt{s/k}}$ (principal branch, $\operatorname{Re}\sqrt{s/k}>0$ for $\operatorname{Re}(s)>0$ ). Boundedness as $x\to\infty$ forces $B(s)=0$ , and the boundary condition gives $A(s)=G(s)$ , so

U(x,s) = G(s)\, e^{-x\sqrt{s/k}}.

Inverting this transform is exactly what the theorem below makes rigorous.

u(x,\cdot)

x

Worked examples

Example 1 — Heat equation on a semi-infinite rod with time-varying boundary temperature. Solve $u_t = k u_{xx}$ for $x>0$ , $t>0$ , with $u(x,0)=0$ , $u(0,t) = T_0\, e^{-\alpha t}$ for constants $T_0>0$ , $\alpha>0$ , and $u(x,t)\to 0$ as $x\to\infty$ . Step 1 — transform. Since $u(x,0)=0$ , Part 1 of the theorem gives $\mathcal L\{u_t\} = sU(x,s)$ , and the PDE becomes $sU = kU_{xx}$ , i.e. $U_{xx} = (s/k)U$ . The boundary data transforms (a standard table entry, $\mathcal L\{e^{-\alpha t}\}(s) = 1/(s+\alpha)$ for $\operatorname{Re}(s)>-\alpha$ ) to

U(0,s) = \frac{T_0}{s+\alpha}.

Step 2 — solve the ODE. The bounded solution as $x\to\infty$ (for $\operatorname{Re}(s)>0$ , so $\operatorname{Re}\sqrt{s/k}>0$ ) is

U(x,s) = \frac{T_0}{s+\alpha}\,e^{-x\sqrt{s/k}}.

Step 3 — invert using the shift and known pair. Recall the standard Laplace pair (derivable from the Bromwich integral applied to $e^{-a\sqrt s}$ , or from the known solution of $u_t=ku_{xx}$ with a Heaviside boundary condition):

\mathcal{L}^{-1}\left\{ \frac{1}{s} e^{-x\sqrt{s/k}} \right\}(t) = \operatorname{erfc}\!\left(\frac{x}{2\sqrt{kt}}\right), \qquad \operatorname{erfc}(z) = \frac{2}{\sqrt\pi}\int_z^\infty e^{-\eta^2}d\eta.

We need $e^{-x\sqrt{s/k}}/(s+\alpha)$ , not divided by $s$ ; write it via the derivative form using the known pair

\mathcal L\left\{ \frac{x}{2\sqrt{\pi k t^3}} e^{-x^2/4kt}\right\}(s) = e^{-x\sqrt{s/k}}

(the transform of the heat kernel derivative; this is the standard "Laplace transform of the boundary-flux kernel" table entry), and use the shifting rule together with convolution. By the convolution theorem (itself proved by Fubini on the double integral defining $\mathcal L\{p*q\}=\mathcal L\{p\}\mathcal L\{q\}$ ),

u(x,t) = T_0\int_0^t e^{-\alpha(t-\tau)}\; \frac{x}{2\sqrt{\pi k \tau^3}} e^{-x^2/(4k\tau)}\, d\tau.

This is the exact solution as a convolution integral. It can be checked at $\alpha = 0$ : then $u(x,t) = T_0\int_0^t \frac{x}{2\sqrt{\pi k\tau^3}}e^{-x^2/4k\tau}d\tau = T_0\operatorname{erfc}\!\left(\frac{x}{2\sqrt{kt}}\right)$ , matching the classical constant-boundary-temperature solution and confirming consistency of the table entries used. As $t\to\infty$ with $x$ fixed, $u(x,t)\to0$ since the boundary forcing decays like $e^{-\alpha t}$ and the diffused response inherits that decay — consistent with the exponential order hypothesis used throughout.

Example 2 — First-order transport equation with inflow boundary condition (subsidiary equation is only first order). Solve $u_t + c\,u_x = 0$ for $x>0,t>0$ (constant $c>0$ ), with $u(x,0) = 0$ and $u(0,t) = \sin(\omega t)$ . Step 1 — transform in $t$ . Using Part 1 with $u(x,0)=0$ : $\mathcal{L}\{u_t\} = sU(x,s)$ . The PDE becomes the first-order linear ODE in $x$

sU + cU_x = 0 \quad\Longrightarrow\quad U_x = -\frac{s}{c}U.

Step 2 — solve. $U(x,s) = U(0,s)\, e^{-sx/c}$ . The boundary data transforms via the standard pair $\mathcal L\{\sin\omega t\}(s) = \omega/(s^2+\omega^2)$ , valid for $\operatorname{Re}(s)>0$ , so

U(x,s) = \frac{\omega}{s^2+\omega^2}\, e^{-sx/c}.

Step 3 — invert using the time-shift (second shifting) theorem. The rule $\mathcal L^{-1}\{e^{-as}F(s)\}(t) = H(t-a) f(t-a)$ , where $H$ is the Heaviside step function and $f=\mathcal L^{-1}\{F\}$ (proved directly: $\int_0^\infty e^{-st}H(t-a)f(t-a)\,dt = \int_a^\infty e^{-st}f(t-a)\,dt = e^{-as}\int_0^\infty e^{-s\tau}f(\tau)\,d\tau$ by the substitution $\tau=t-a$ ), gives with $a = x/c$ and $f(t)=\sin(\omega t)$ :

u(x,t) = H\!\left(t - \tfrac{x}{c}\right)\sin\!\left(\omega\left(t - \tfrac{x}{c}\right)\right).

This is exactly the d'Alembert-type traveling-wave answer: the boundary signal propagates undistorted at speed $c$ , arriving at position $x$ at time $x/c$ , echoing d'Alembert's Solution of the Wave Equation though here for a first-order (one-directional) transport equation rather than the second-order wave equation. It can be verified directly by substitution: for $t>x/c$ , $\partial_t u = \omega\cos(\omega(t-x/c))$ and $\partial_x u = -\tfrac{\omega}{c}\cdot(-1)\cos(\omega(t-x/c))\cdot(-1)$ ; carefully, $u_x = -\omega\cos(\omega(t-x/c))\cdot(-1/c) \cdot(-1)$ , and combining $u_t + cu_x = \omega\cos(\cdot) - \omega\cos(\cdot) = 0$ , confirming the PDE, and $u(0,t)=H(t)\sin(\omega t) = \sin(\omega t)$ for $t>0$ as required.

u(t) = e^{t^2}

Scope, edge cases & common traps

Situation	What to check	Common pitfall
Spatial domain is a half-line	Impose boundedness/decay as $x\to\infty$ to kill the growing exponential solution	Forgetting to discard $B(s)e^{x\sqrt{s/k}}$ , giving an unbounded, non-physical $U(x,s)$
Time-dependent boundary data $g(t)$	Confirm $g$ has a Laplace transform (exponential order) before transforming	Assuming $\mathcal L\{g\}$ exists without checking growth (see counterexample)
Branch of $\sqrt{s}$ or $\sqrt{s/k}$	Fix the principal branch with $\operatorname{Re}\sqrt{s/k} > 0$ for $\operatorname{Re}(s)>0$ , consistent across the whole right half-plane	Sign errors from an inconsistent branch choice, producing a spurious blow-up
Inverting via tables vs. Bromwich integral	Tables are shortcuts for exactly the same Bromwich contour computation; use convolution theorem to combine pieces	Multiplying two inverse transforms directly instead of convolving them
Finite interval $0<x<L$	Subsidiary ODE has two boundary conditions (at $x=0,L$ ); no boundedness condition needed, get $A(s),B(s)$ from linear system	Trying to invert an infinite series of poles (many closely spaced $s$ -singularities) without justifying term-by-term inversion

Common trapthe classic error: assuming

U(x,s)

is analytic on the whole right half-plane without checking, then inverting term by term

A very common mistake is to write down $U(x,s)$ from the subsidiary ODE, expand it as a series or partial fraction, and invert term by term — assuming this is legitimate because "each term looks like a table entry." The inversion theorem requires $U$ to be analytic on a full half-plane $\operatorname{Re}(s)>\gamma$ and satisfy the polynomial decay bound $|U(s)|\le M/|s|^p$ , $p>1$ , uniformly as $|s|\to\infty$ in that half-plane; branch cuts (from $\sqrt s$ ) or infinitely many poles (finite-interval problems) require extra justification — either deforming the Bromwich contour around a branch cut, or invoking a residue sum that must itself be shown to converge to the correct function. Skipping this check can silently produce a series that diverges, or that converges to the wrong function outside its radius of validity. The correct order of operations is: (1) verify $U$ 's analyticity domain and decay, (2) then invoke Part 3 or a residue-based variant, never the reverse.

The mathematician behind it

Origins · 1 minPierre-Simon Laplace and how this was discovered

Pierre-Simon Laplace

1749-1827 · French mathematician and astronomer, developer of the transform methods and celestial mechanics that bear his name

Laplace worked at the intersection of probability theory, mathematical physics, and celestial mechanics in Napoleonic-era France. He introduced the transform that bears his name in the context of probability generating functions and the theory of differential equations, decades before its systematic use for solving PDEs with initial and boundary data became standard engineering practice in the twentieth century, largely through the work of Oliver Heaviside and later rigorous treatments by mathematicians such as Bromwich and Doetsch.

History

- 1749: Pierre-Simon Laplace born in Beaumont-en-Auge, Normandy.

- 1782-1785: Laplace develops transform techniques in his work on probability and celestial mechanics, the origin of the "Laplace transform."

- 1893: Oliver Heaviside popularizes operational calculus methods for solving differential equations in electrical engineering, closely related to Laplace transform techniques.

- 1916-1917: Thomas John I'Anson Bromwich provides a rigorous complex-analytic foundation for the inversion integral, justifying Heaviside's operational methods.

- 1937: Gustav Doetsch's systematic treatise establishes the Laplace transform as a standard rigorous tool for solving partial differential equations with initial-boundary data.

Why use the Laplace transform instead of the Fourier transform for these problems?

The Fourier transform in $t$ requires the function to be defined (and typically decaying or bounded) on the whole real line $-\infty<t<\infty$ , which is unnatural when a physical process genuinely starts at $t=0$ with a definite initial condition. The Laplace transform is built for functions on $t\ge0$ and automatically incorporates initial data through its differentiation rule $\mathcal L\{u_t\}=sU-u(0)$ . It also only requires exponential growth control rather than absolute integrability, so it applies to a much wider class of physically realistic (even growing) time signals. The two transforms coincide in spirit — both diagonalize $\partial_t$ — but Laplace is the correct tool whenever "time starts now" and initial conditions matter.

Do I always need to compute the Bromwich contour integral by hand?

No. In practice, most problems are solved by recognizing $U(x,s)$ as built from standard transform pairs (exponentials, powers of $s$ , $\sqrt s$ , rational functions) combined via shifting, scaling, and the convolution theorem, all of which are shortcuts derived once and for all from the Bromwich integral. The contour integral is mainly invoked directly when no elementary pair matches, often evaluated via residues if $U$ has only poles, or via a branch-cut integral if $U$ involves $\sqrt s$ (as in diffusion problems), producing integrals like the error function.

What if the transformed boundary condition itself does not have a Laplace transform?

Then the method cannot get started, exactly as illustrated in the counterexample with $u(t)=e^{t^2}$ . Before transforming, you must verify that all data (initial condition, boundary conditions, and forcing terms) are of exponential order. Most physically reasonable signals — bounded, polynomially growing, or exponentially decaying/growing signals — satisfy this, but pathological super-exponential growth (rare in applications, but possible in artificially constructed or blow-up scenarios) rules the method out entirely.

How does this method handle finite spatial intervals, like

0<x<L

The subsidiary ODE in $x$ is solved exactly as on the half-line, but now both boundary conditions at $x=0$ and $x=L$ are used (instead of one boundary condition plus a boundedness condition at infinity), giving a $2\times2$ linear system for the two constants in the general solution. The resulting $U(x,s)$ typically has infinitely many poles (at values of $s$ related to the eigenvalues of the associated Sturm-Liouville problem), and inverting via residues at all these poles reproduces the Fourier/eigenfunction series solution — connecting this method directly to separation-of-variables techniques.

Is the Laplace transform method restricted to linear PDEs?

Yes, in its standard form. The core algebraic trick — $\mathcal L\{u_t\}=sU-u(0)$ — relies on linearity of the transform and of the differential operator; the transform of a nonlinear term like $u^2$ or $uu_x$ does not reduce to a simple algebraic expression in $U(x,s)$ , but instead produces a convolution in $s$ , which usually makes the subsidiary equation as hard as (or harder than) the original PDE. For genuinely nonlinear evolution equations, other methods (energy estimates, fixed-point/semigroup theory, or specific nonlinear transform tricks like Cole-Hopf for Burgers' equation) are used instead.

How does this relate to well-posedness of the PDE?

Constructing an explicit solution via Laplace transform is a strong existence argument, and the injectivity result (Part 2 of the theorem) supplies uniqueness among solutions of the assumed growth class. Continuous dependence on data (the third pillar of well-posedness) is a separate question, usually addressed by direct energy or maximum-principle estimates rather than by transform algebra; the connections between these solvability arguments and the general framework are developed further under Well-Posedness of PDE Problems.

Can this method fail even when a classical solution exists?

Yes — the method can fail to produce the answer even when a perfectly good solution exists, if the growth or analyticity hypotheses of the inversion theorem are not verified, or if $U(x,s)$ cannot be found in closed form. In such cases the subsidiary ODE may still be solved, but the resulting inversion integral must be evaluated numerically (inverse Laplace transform algorithms such as Talbot's method), or asymptotically for large or small $t$ , rather than in exact closed form.