Laplace Transform Inversion

Every Laplace transform $F(s)=\int_0^\infty f(t)e^{-st}\,\mathrm{d}t$ is a bookkeeping device: it trades a function of time for an analytic function of a complex frequency $s$ , defined on a half-plane $\operatorname{Re}s>\sigma_0$ . Inversion is the return trip — given $F(s)$ , recover the $f(t)$ it came from. The answer is not a table lookup but a single object: the Bromwich integral, a contour integral running up a vertical line $\operatorname{Re}s=c$ placed to the right of every singularity of $F$ . It is the exact inverse of $\mathcal L$ , and the engine underneath solving differential equations by transform methods and building Green's functions from integral transforms. The deepest reading is that inversion is a Fourier transform wearing a disguise. Multiplying $f$ by $e^{-ct}$ tames any function growing no faster than $e^{\sigma_0 t}$ , so $g(t)=e^{-ct}f(t)\,\theta(t)$ becomes absolutely integrable and its Fourier transform is exactly $F(c+i\omega)$ . Undo the Fourier inversion and the damping, and the Bromwich formula falls out — the vertical Bromwich line is the real Fourier frequency axis, slid rightward into the region where the integral converges. In practice one almost never grinds through the line integral: one closes it with a giant semicircle and reads off residues. For $t>0$ the factor $e^{st}$ decays on the left, so we close left and collect a residue per pole; for $t<0$ it decays on the right, where $F$ is analytic, so the integral is zero — the analytic origin of causality. And the kind of singularity dictates the kind of answer: isolated poles give exponential modes $e^{s_k t}$ , while branch points such as $\sqrt{s}$ give power-law, diffusive tails like $t^{-1/2}$ .

Key takeaways

The Laplace transform is inverted by the Bromwich (Mellin–Fourier) integral $f(t)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}F(s)e^{st}\,\mathrm{d}s$ along a vertical line to the right of every singularity of $F$ .
The formula is the inverse Fourier transform of the damped causal function $e^{-ct}f(t)\theta(t)$ — the Bromwich line is the Fourier axis shifted into the region of convergence.
For $t>0$ close the contour left; Jordan's lemma kills the arc and $f(t)$ is the sum of residues of $F(s)e^{st}$ . For $t<0$ close right, enclose nothing, get $f(t)=0$ — automatic causality.
Simple poles give exponential modes $r_k e^{s_k t}$ ; higher-order poles give polynomial-times-exponential terms; branch points (e.g. $\sqrt{s}$ ) require a keyhole contour and give $t^{-1/2}$ tails. The rightmost singularity controls the long-time behaviour (the dominant-singularity/Tauberian principle).
Key idea: The inverse Laplace transform is a Fourier transform analytically continued off the real axis, $f(t)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}F(s)e^{st}\,\mathrm{d}s$ . Close the contour left for $t>0$ and sum residues; close right for $t<0$ and get zero. The poles and branch cuts of $F(s)$ encode the modes and tails of $f(t)$ .

Where this lives today

Laplace inversion is a daily instrument in computational physics and engineering. Whenever a linear ODE or PDE is solved by transform methods — circuit and control transfer functions, viscoelastic relaxation, heat conduction, subsurface flow in well-test analysis — the final step is an inverse transform, done by residues when the poles are known and numerically otherwise. That numerical branch is a research field of its own: the Talbot method deforms the Bromwich contour onto a steepest-descent path, the Gaver–Stehfest method uses weighted real-axis samples, and the Weeks method expands in Laguerre functions — workhorses in computational rheology, option pricing in quantitative finance, and reservoir simulation. The deepest modern twist is that inversion is ill-posed: recovering $f(t)$ from noisy samples of $F(s)$ is exponentially sensitive to error, because very different time signals can share nearly identical transforms. The same pathology appears in lattice quantum chromodynamics, where reconstructing a spectral function from a Euclidean correlator is an inverse-Laplace problem attacked with Backus–Gilbert, Tikhonov, and Bayesian maximum-entropy methods. The innocent-looking contour integral sits at the frontier of both practical simulation and ill-conditioned inverse problems.

The physical picture

Read $F(s)$ as a catalogue of the natural frequencies and decay rates hidden inside $f(t)$ . A pole at $s=s_k$ says "the signal contains a mode $e^{s_k t}$ ", and the residue there is that mode's amplitude. Inversion just reads the catalogue back out: one exponential per pole, weighted by its residue. The geometry that turns this slogan into a theorem is the Bromwich contour — a vertical line in the complex $s$ -plane that we bulge into a closed loop and evaluate by residues.

Figure

The Bromwich contour in the complex s-plane. The vertical line at Re s = c sits to the right of every singularity — here a conjugate pole pair at s = −3 ± 4i, so σ₀ = −3 and any c > −3 works. For t > 0 the line closes with a large left semicircle whose arc vanishes by Jordan's lemma, so f(t) equals the sum of residues of F(s)eˢᵗ at the enclosed poles.

The shift by $c$ is a convergence trick, nothing more. Real Fourier analysis fails on a function growing like $e^{at}$ , but multiply by $e^{-ct}$ with $c>a$ and it is pulled back into $L^1$ — the Bromwich line is the Fourier axis relocated to where convergence holds. Causality is then a matter of which way the contour must bulge: for $t<0$ , $e^{st}$ decays to the right, forcing a rightward closure — but $F$ is analytic there, no poles are trapped, and the answer is zero. The transform "remembers" that the signal switched on at $t=0$ . The most important qualitative dial is whether the nearest singularity is a pole or a branch cut, deciding whether $f(t)$ is a sum of clean exponentials or a slow algebraic tail.

Figure

The singularity dictionary: the analytic structure of F(s) fixes the shape of f(t). A simple real pole at s = −γ gives a pure decay e⁻ᵞᵗ; a conjugate pair at −γ ± iω gives a damped oscillation e⁻ᵞᵗcos ωt; a branch cut from √s gives the diffusive power law 1/√(πt), which decays far more slowly than any exponential.

The core equations

The whole subject rests on one formula and a handful of consequences of the residue theorem. The master statement is the Bromwich integral — the exact inverse of $\mathcal L$ , valid for any $F$ analytic in a right half-plane. Everything below is a special case of it or a shortcut for evaluating it.

f(t)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}F(s)\,e^{st}\,\mathrm{d}s,\qquad c>\sigma_0 \qquad (1)

where: $F(s)$ — the Laplace transform of $f$ , analytic for $\operatorname{Re}s>\sigma_0$ ; $c$ — abscissa of the vertical Bromwich contour, right of all singularities; $\sigma_0$ — abscissa of convergence, the real part of the rightmost singularity; $s=c+i\omega$ — complex frequency. Reading it: integrate $F(s)e^{st}$ up an infinite vertical line; the result is independent of $c$ provided $c>\sigma_0$ . When $F$ has only isolated poles and decays on large arcs, the line integral collapses to a finite residue sum — the workhorse formula that makes inversion mechanical.

f(t)=\sum_k \operatorname{Res}_{s=s_k}\!\big[F(s)\,e^{st}\big]\qquad(t>0) \qquad (2)

where: $s_k$ — the poles of $F(s)$ , all at $\operatorname{Re}s_k<c$ ; $\operatorname{Res}$ — the residue at $s_k$ . Reading it: each pole donates one term — a simple pole gives $r_k e^{s_k t}$ , an $n$ -th order pole a degree- $(n-1)$ polynomial in $t$ times $e^{s_k t}$ . The mirror image for negative time closes the same contour to the right, where $F$ is analytic and encloses nothing.

f(t)=0\qquad(t<0) \qquad (3)

where: $t<0$ — instants before the source is switched on. Reading it: causality is built in automatically — the inverse is supported on $t\ge 0$ with no extra assumption. For a rational $F=P/Q$ with distinct roots of $Q$ , the residue sum has a closed form — the Heaviside expansion theorem, which sidesteps partial fractions.

\mathcal L^{-1}\!\Big\{\frac{P(s)}{Q(s)}\Big\}=\sum_k \frac{P(s_k)}{Q'(s_k)}\,e^{s_k t} \qquad (4)

where: $P,Q$ — polynomials with $\deg P<\deg Q$ ; $s_k$ — simple roots of $Q(s)=0$ ; $Q'$ — derivative of $Q$ . Reading it: evaluate $P/Q'$ at each root, hang $e^{s_k t}$ on it — inversion in one line. Products of transforms invert to convolutions, letting you build a complicated inverse from two simple ones without factoring the product.

\mathcal L^{-1}\{F(s)G(s)\}(t)=\int_0^t f(\tau)\,g(t-\tau)\,\mathrm{d}\tau \qquad (5)

where: $f,g$ — the inverse transforms of $F,G$ ; the integral is the causal convolution $f*g$ . Reading it: a transfer function times an input transform inverts to the input smeared against the impulse response. Finally, when $F$ carries a branch point rather than poles — the diffusion signature $\sqrt{s}$ — the inversion is a cut integral yielding the non-exponential kernels below.

\mathcal L^{-1}\!\Big\{\frac{1}{\sqrt{s}}\Big\}=\frac{1}{\sqrt{\pi t}},\qquad \mathcal L^{-1}\{e^{-a\sqrt{s}}\}=\frac{a}{2\sqrt{\pi}\,t^{3/2}}\,e^{-a^2/4t},\qquad \mathcal L^{-1}\!\Big\{\frac{1}{s^n}\Big\}=\frac{t^{n-1}}{(n-1)!} \qquad (6)

where: $a>0$ — depth/distance parameter ( $a=x/\sqrt{D}$ in diffusion); $n$ — positive integer (order of the pole at $s=0$ ); the $\sqrt{s}$ branch is cut on the negative real axis. Reading it: an order- $n$ pole at $s=0$ gives a polynomial $t^{n-1}$ , but a square-root branch point gives a fractional power — the fingerprint of a continuous spectrum.

Derivation AThe Bromwich integral as an analytically continued Fourier transform

The goal is to show that the Bromwich formula is not an independent postulate but a direct consequence of Fourier inversion applied to a damped, causal function. We damp $f$ by $e^{-ct}$ with $c>\sigma_0$ to force it into $L^1$ , recognise its Fourier transform as $F$ on a vertical line, invert, and change variables to $s=c+i\omega$ . The payoff is conceptual: it explains why the contour is vertical, why it sits at $c>\sigma_0$ , and why the whole apparatus is Fourier analysis in disguise.

Step 1. Define the damped, one-sided function $g(t)=e^{-ct}f(t)\,\theta(t)$ . Since $f$ grows at most like $e^{\sigma_0 t}$ , choosing $c>\sigma_0$ makes $|g(t)|\le C e^{-(c-\sigma_0)t}$ integrable, so $g\in L^1(\mathbb R)$ and its Fourier transform exists. Why: real Fourier analysis needs an absolutely integrable function, and the exponential damping is exactly the surgery that supplies one for a growing $f$ .

Step 2. Compute its Fourier transform: $\hat g(\omega)=\int_{-\infty}^{\infty} g(t)e^{-i\omega t}\,\mathrm{d}t=\int_0^\infty f(t)e^{-(c+i\omega)t}\,\mathrm{d}t=F(c+i\omega)$ — the Laplace transform sampled along the vertical line $\operatorname{Re}s=c$ . Why: this is the hinge — it identifies the values of $F$ on the Bromwich line with a genuine Fourier transform, so Fourier inversion becomes available.

Step 3. Invert the Fourier transform, $g(t)=\frac{1}{2\pi}\int_{-\infty}^{\infty}F(c+i\omega)e^{i\omega t}\,\mathrm{d}\omega$ , then undo the damping via $e^{ct}$ : $f(t)=\frac{1}{2\pi}\int_{-\infty}^{\infty}F(c+i\omega)e^{(c+i\omega)t}\,\mathrm{d}\omega$ . Why: recovering $f$ is two moves — invert the Fourier transform to get $g$ , then strip off the $e^{-ct}$ introduced by hand.

Step 4. Change variables to $s=c+i\omega$ , so $\mathrm{d}s=i\,\mathrm{d}\omega$ and the real line maps to the vertical line from $c-i\infty$ to $c+i\infty$ ; the prefactor $\frac{1}{2\pi}\mathrm{d}\omega$ becomes $\frac{1}{2\pi i}\mathrm{d}s$ , yielding equation (1). Why: this change of variable promotes a horizontal real-frequency integral into a vertical contour — the Bromwich line is the Fourier axis translated into the region of convergence. *Verified: setting $F(c+i\omega)=\hat g(\omega)$ and reversing the steps recovers Fourier inversion; independence of $c$ for $c>\sigma_0$ follows because shifting $c$ deforms the contour through a region where $F e^{st}$ is analytic.*

Go deeper · Derivation B · 2 minClosing the contour: residues for t greater than zero, zero for t less than zero

Derivation BClosing the contour: residues for t greater than zero, zero for t less than zero

The aim is to convert the infinite line integral (1) into something finite and to extract causality as a bonus. We complete the Bromwich line into a closed loop with a large arc, use Jordan's lemma to kill the arc, and invoke the residue theorem. The direction we may close — left or right — is dictated by the sign of $t$ in $e^{st}$ , and it is precisely this forced choice that produces $f(t)=0$ for $t<0$ . The result is equations (2) and (3) together: the inversion recipe and the causality theorem in one construction.

Step 1. On a semicircular arc $s=c+Re^{i\theta}$ of radius $R$ , the exponential is $|e^{st}|=e^{t(c+R\cos\theta)}$ , which decays as $R\to\infty$ only where $t\cos\theta<0$ . For $t>0$ that is $\cos\theta<0$ , the left half; for $t<0$ the right half. Why: the sign of $t$ chooses which semicircle can be added for free, and this asymmetry is the seed of causality.

Step 2. For $t>0$ , close left. Provided $F(s)\to 0$ uniformly on the arc (true for every proper rational function and all $s^{-p}$ , $p>0$ ), Jordan's lemma gives $\big|\int_{C_R}F(s)e^{st}\,\mathrm{d}s\big|\to 0$ . The closed contour — vertical line plus left arc — then encircles every pole of $F$ , all at $\operatorname{Re}s<c$ . Why: killing the arc is what lets the closed-loop residue sum equal the Bromwich line integral; without $F\to 0$ the shortcut is invalid.

Step 3. Apply the residue theorem to the counterclockwise loop: $\frac{1}{2\pi i}\oint F e^{st}\,\mathrm{d}s=\sum_k \operatorname{Res}_{s_k}[F e^{st}]$ . Since the arc vanishes, the left side is the Bromwich integral, giving equation (2) for $t>0$ . Why: this is the entire content of inversion for pole-only transforms — locate poles, take residues, add.

Step 4. For $t<0$ , close right instead. The enclosed region $\operatorname{Re}s>c>\sigma_0$ is where $F$ is analytic, so the loop traps no singularity: $\frac{1}{2\pi i}\oint F e^{st}\,\mathrm{d}s=0$ , hence $f(t)=0$ . Why: right-half-plane analyticity plus the rightward closure forced by $t<0$ delivers strict causality — the inverse cannot leak to negative times. *Verified: for $F=1/(s-a)$ with pole at $s=a<c$ , closing left gives $\operatorname{Res}=e^{at}$ so $f(t)=e^{at}\theta(t)$ , and closing right for $t<0$ gives $0$ — reproducing the elementary pair $\mathcal L\{e^{at}\theta(t)\}=1/(s-a)$ .*

Go deeper · Derivation C · 1 minBranch-cut inversion of one over root s by a keyhole contour

Derivation CBranch-cut inversion of one over root s by a keyhole contour

This derivation handles the case that breaks the residue recipe: a transform with a branch point instead of poles. We invert $F(s)=1/\sqrt{s}$ , whose square root branches at the origin. With no residues to collect, the Bromwich line is deformed into a keyhole wrapping the cut along the negative real axis, and the answer comes from the discontinuity of $\sqrt{s}$ across it. The result $f(t)=1/\sqrt{\pi t}$ is the prototype of every diffusive tail — a power law, not an exponential, the fingerprint of a continuous spectrum. The same machinery on $e^{-a\sqrt{s}}$ gives the Gaussian heat kernel used below.

Step 1. Put the cut of $\sqrt{s}$ along the negative real axis and deform the Bromwich line onto a keyhole: two segments hugging the cut from $-\infty$ to $0$ and back, joined by a vanishing circle at the origin. Since $1/\sqrt{s}$ has no poles, the keyhole integral equals the Bromwich integral, and the small circle contributes nothing ( $\int r^{-1/2}\,r\,\mathrm{d}\theta\to 0$ ). Why: with no residues available, the whole value comes from the mismatch of $\sqrt{s}$ on the two lips of the cut.

Step 2. Parametrise the lips as $s=re^{+i\pi}$ (above) and $s=re^{-i\pi}$ (below), $r>0$ . Then $\sqrt{s}=\pm i\sqrt{r}$ , and on both lips $e^{st}=e^{-rt}$ since $\operatorname{Re}s=-r$ . Why: the sign flip of $\sqrt{s}$ between the lips is the discontinuity that survives; the exponential is common to both and controls convergence.

Step 3. Combine the two lips with their opposite orientations. Accounting for $\mathrm{d}s=-\mathrm{d}r$ on the return leg, the upper ( $1/i\sqrt{r}$ ) and lower ( $1/(-i\sqrt{r})$ ) contributions add rather than cancel: $f(t)=\frac{1}{2\pi i}\int_0^\infty\Big(\frac{1}{-i\sqrt{r}}-\frac{1}{i\sqrt{r}}\Big)e^{-rt}\,(-\mathrm{d}r)=\frac{1}{\pi}\int_0^\infty \frac{e^{-rt}}{\sqrt{r}}\,\mathrm{d}r$ . Why: the two imaginary factors of opposite sign combine into a single real, convergent integral — the branch cut turns a complex contour into a real Laplace-type integral.

Step 4. Substitute $r=u^2$ : $\int_0^\infty r^{-1/2}e^{-rt}\,\mathrm{d}r=2\int_0^\infty e^{-tu^2}\,\mathrm{d}u=\sqrt{\pi/t}$ , so $f(t)=\frac{1}{\pi}\sqrt{\pi/t}=\frac{1}{\sqrt{\pi t}}$ . Why: the Gaussian integral $\int_0^\infty e^{-tu^2}\mathrm{d}u=\tfrac12\sqrt{\pi/t}$ converts the cut integral into the diffusive $t^{-1/2}$ power law. *Verified: the forward check $\mathcal L\{t^{-1/2}/\sqrt{\pi}\}=\Gamma(\tfrac12)/(\sqrt{\pi}\,s^{1/2})=1/\sqrt{s}$ holds because $\Gamma(\tfrac12)=\sqrt{\pi}$ ; the power-law tail confirms that a branch point, unlike a pole, produces no exponential.*

Limiting cases

The organising principle is that the analytic structure of $F(s)$ dictates the functional form of $f(t)$ . The table reads that off case by case, from isolated poles to branch points, closing with the asymptotic rule that lets you skip inversion when you only want the long-time behaviour.

Case	Result	Meaning
Simple poles only	$f(t)=\sum_k r_k\,e^{s_k t}$ , $r_k=\operatorname{Res}_{s_k}F(s)e^{st}$	Each first-order pole at $s_k$ contributes a pure exponential mode; every proper rational $F=P/Q$ with distinct roots falls here.
Higher-order pole	$\operatorname{Res}_{s_0}=\lim_{s\to s_0}\frac{1}{(n-1)!}\frac{\mathrm{d}^{\,n-1}}{\mathrm{d}s^{\,n-1}}\big[(s-s_0)^n F e^{st}\big]$	An $n$ -th order pole gives terms $t^m e^{s_0 t}$ , $m=0,\dots,n-1$ ; e.g. $1/s^2\to t$ , $1/(s-a)^2\to t\,e^{at}$ (critical damping).
Branch point at $s=0$	$F=1/\sqrt{s}\ \Rightarrow\ f=1/\sqrt{\pi t}$	A cut replaces the residue sum with a keyhole integral, giving power-law (diffusive) decay instead of exponentials — a continuous spectrum.
Complex-conjugate pair	$s=-\gamma\pm i\omega_d\ \Rightarrow\ f\propto e^{-\gamma t}\cos(\omega_d t+\varphi)$	Poles off the real axis pair up into a damped (or, on the axis, undamped) oscillation — the ubiquitous underdamped response.
Large- $t$ asymptotics	$f(t)\sim r_{k^}\,e^{s_{k^}t}$ , $s_{k^*}=$ rightmost singularity	The singularity with the largest $\operatorname{Re}s$ dominates as $t\to\infty$ (dominant-singularity / Tauberian principle); a branch point at $0$ gives algebraic relaxation.

Common trapPlacing the Bromwich line to the left of a singularity, or closing the wrong way

The Bromwich contour must lie to the right of every singularity of $F(s)$ , i.e. $c>\sigma_0$ . Slide it left of even one pole and you silently drop that pole's residue, producing an answer that is not merely wrong but non-causal — it generally fails to vanish for $t<0$ . A related trap is the closure direction: for $t>0$ the factor $e^{st}$ decays only on the left semicircle, so you may close left and only left (the mirror rule holds for $t<0$ ). Finally, the residue shortcut is licensed by Jordan's lemma, which requires $F(s)\to 0$ uniformly on the arc — apply it to an $F$ that does not decay at infinity (an improper rational function with $\deg P\ge\deg Q$ , or a bare $e^{as}$ ) and the arc no longer vanishes, so the residue sum is incomplete. The cure is to divide out the polynomial part first, whose inverse is delta functions and their derivatives.

Worked examples

1 · Free response of an underdamped oscillator (fully worked)

A damped oscillator obeys $\ddot x+6\dot x+25x=0$ with $x(0)=1$ , $\dot x(0)=0$ ( $t$ in seconds). Solve it by Laplace inversion and read off the envelope and ringing frequency.

Step 1. Transform the ODE with $\mathcal L\{\ddot x\}=s^2X-sx(0)-\dot x(0)$ and $\mathcal L\{\dot x\}=sX-x(0)$ . With the given data, $(s^2X-s)+6(sX-1)+25X=0$ , so $X(s)=\dfrac{s+6}{s^2+6s+25}$ . Completing the square, $s^2+6s+25=(s+3)^2+16$ , exposing a conjugate pole pair at $s=-3\pm 4i$ — decay rate $\gamma=3\ \mathrm{s^{-1}}$ , ringing frequency $\omega_d=4\ \mathrm{rad/s}$ .

Step 2. Invert by residues (equation 2). At $s_+=-3+4i$ the residue of $X(s)e^{st}$ is $\dfrac{(s_++6)e^{s_+t}}{2(s_++3)}=\dfrac{3+4i}{8i}\,e^{s_+t}$ , and the coefficient simplifies to $\dfrac{(3+4i)(-i)}{8}=\dfrac{4-3i}{8}=0.5-0.375i$ . The pole $s_-=-3-4i$ gives the complex conjugate, so the two residues sum to twice the real part:

x(t)=2\,\operatorname{Re}\!\big[(0.5-0.375i)\,e^{(-3+4i)t}\big]=e^{-3t}\big(\cos 4t+0.75\sin 4t\big).

Step 3. Collapse the sinusoids into one amplitude-and-phase form. With $\cos 4t+0.75\sin 4t=R\cos(4t-\varphi)$ , the amplitude is $R=\sqrt{1^2+0.75^2}=\mathbf{1.25}$ and the phase $\varphi=\arctan(0.75)=\mathbf{36.87^\circ}$ , so

\boxed{\,x(t)=1.25\,e^{-3t}\cos\!\big(4t-36.87^\circ\big)\,}

Step 4. Check the initial conditions and the physics. At $t=0$ : $x(0)=1.25\cos(-36.87^\circ)=1.25\times 0.8=\mathbf{1}$ ✓, and $\dot x(0)=-3(1)+4(0.75)=\mathbf{0}$ ✓. The two poles sit in the left half-plane (stable); their imaginary parts give the $4\ \mathrm{rad/s}$ ring, their common real part $-3$ sets the $e^{-3t}$ envelope, so the amplitude falls to $1/e$ in $1/3\ \mathrm{s}$ . The oscillator rings at $\omega_d=4\ \mathrm{rad/s}$ inside a decaying envelope of amplitude $1.25\,e^{-3t}$ .

2 · Temperature in a semi-infinite solid (branch-cut inversion)

A large block of copper ( $D=1.11\times10^{-4}\ \mathrm{m^2/s}$ ) initially at $0$ has its face at $x=0$ raised to $T_0$ at $t=0$ and held there. Find $T(x,t)$ and estimate how deep the heat reaches after $1$ second.

Step 1. Laplace-transform $\partial_t T=D\,\partial_{xx}T$ in time. With $T(x,0)=0$ it becomes $s\bar T=D\,\bar T''$ , whose bounded ( $x\to\infty$ ) solution is $\bar T=A(s)\,e^{-x\sqrt{s/D}}$ . The face condition $T(0,t)=T_0$ transforms to $\bar T(0,s)=T_0/s$ , fixing $A=T_0/s$ , so $\bar T(x,s)=\dfrac{T_0}{s}\,e^{-x\sqrt{s/D}}$ .

Step 2. Invert. The transform carries a $\sqrt{s}$ branch point, so this is a keyhole (branch-cut) inversion, not a residue sum. The standard cut integral for $\mathcal L^{-1}\{e^{-a\sqrt{s}}/s\}$ with $a=x/\sqrt{D}$ gives the complementary error function,

T(x,t)=T_0\,\operatorname{erfc}\!\Big(\frac{x}{2\sqrt{Dt}}\Big),

the same branch-cut machinery as the $1/\sqrt{s}\to 1/\sqrt{\pi t}$ derivation, now producing the erfc diffusion profile rather than exponential modes.

Step 3. Read off the penetration depth. With the similarity variable $\eta=x/(2\sqrt{Dt})$ , the temperature reaches a fixed fraction of $T_0$ wherever $\eta$ is fixed; at $\eta=1$ , $\operatorname{erfc}(1)=0.157$ , so $x=2\sqrt{Dt}$ marks where the step has reached about $16\%$ of $T_0$ . For copper after $t=1\ \mathrm{s}$ , $x=2\sqrt{(1.11\times10^{-4})(1)}=\mathbf{2.1\ cm}$ . The thermal front advances as $\sqrt{Dt}$ , not linearly in $t$ — the diffusive, not wavelike, spread that the branch cut predicts.

Physics of the portrait

The inversion formula was built across a century — an engineer who used it before it was justified, and the mathematician who justified it.

Origins · 1 minBromwich integral and how this was discovered

Bromwich integral

the inversion contour f(t) = (1/2πi)∮ F(s)eˢᵗ ds · complex-variable theory of the Laplace transform

The formula carries the name of Thomas John I'Anson Bromwich (1875–1929), a Cambridge mathematician, but its story begins with an engineer's audacity. In the 1880s and 1890s Oliver Heaviside, solving transient problems in telegraph cables, invented an "operational calculus" that treated the differentiation symbol $p=\mathrm{d}/\mathrm{d}t$ as an ordinary algebraic quantity — writing solutions like $1/(p+a)$ and expanding them into series with cheerful disregard for rigour. His answers were right and his methods scandalous; challenged by mathematicians who could not derive them, Heaviside retorted that he would not refuse dinner because he did not fully understand digestion. The rehabilitation came in 1916, when Bromwich showed that Heaviside's operators are exactly evaluated by a contour integral in the complex plane — the vertical-line integral now bearing his name — grounding the operational calculus in Cauchy's residue theorem. The same object had appeared through the Fourier–Mellin inversion (Hjalmar Mellin's work around 1900), and Gustav Doetsch's 1930s treatises made the whole apparatus a standard tool. What Heaviside computed by daring, Bromwich explained by contour — a recurring pattern in mathematical physics, where a useful trick waits years for the complex plane to reveal why it works.

History

1785

— Laplace uses the integral ∫₀^∞ e⁻ˢᵗ f(t) dt in his work on probability

1893

— Heaviside's operational calculus solves transient circuits by treating d/dt as an algebraic symbol

1916

— Bromwich derives the contour-integral inversion, putting Heaviside's methods on a rigorous footing

1937

— Doetsch publishes Theorie und Anwendung der Laplace-Transformation, systematizing the theory

1979

— Talbot introduces the deformed-contour algorithm for accurate numerical inversion

Going further

Texts. Arfken & Weber, Mathematical Methods for Physicists, 7th ed., Ch. 15 (Laplace transforms and the Bromwich integral) · Hassani, Mathematical Physics, chapters on complex analysis and integral transforms · Stone & Goldbart, Mathematics for Physics, Ch. on Laplace transforms and contour methods · Byron & Fuller, Mathematics of Classical and Quantum Physics, §6 (transforms) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, Ch. 13 · Doetsch, Introduction to the Theory and Application of the Laplace Transformation (the classic monograph).

Provenance. The Bromwich formula, the residue/Heaviside expansion theorems, and the standard inverse pairs used here ( $1/s^n\to t^{n-1}/(n-1)!$ , $1/(s^2+\omega^2)\to\sin\omega t/\omega$ , $1/\sqrt{s}\to 1/\sqrt{\pi t}$ , $e^{-a\sqrt{s}}\to (a/2\sqrt{\pi}\,t^{3/2})e^{-a^2/4t}$ , and $e^{-a\sqrt{s}}/s\to\operatorname{erfc}(a/2\sqrt{t})$ ) are standard results in Arfken & Weber Ch. 15, Hassani, and Stone & Goldbart. A verifier should: (1) re-derive the Bromwich integral from Fourier inversion of $e^{-ct}f(t)\theta(t)$ ; (2) recompute the oscillator residues, confirming $x(t)=1.25\,e^{-3t}\cos(4t-36.87^\circ)$ with $x(0)=1$ , $\dot x(0)=0$ ; (3) verify the branch-cut integral $\tfrac1\pi\int_0^\infty r^{-1/2}e^{-rt}\,\mathrm{d}r=1/\sqrt{\pi t}$ using $\Gamma(\tfrac12)=\sqrt{\pi}$ ; (4) confirm the semi-infinite-solid result $T=T_0\operatorname{erfc}(x/2\sqrt{Dt})$ and the copper penetration $x=2\sqrt{Dt}=2.1\ \mathrm{cm}$ at $t=1\ \mathrm{s}$ ; (5) check each table entry against its forward transform.

Prerequisites. The Laplace Transform · The Fourier Transform · Properties of the Fourier Transform · The Convolution Theorem · The Dirac Delta Function. Leads to. Solving Differential Equations by Transform Methods · Green's Functions via Integral Transforms · Tempered Distributions and Fourier Transforms.

Ask the tutor: Give me any $F(s)$ and I will locate its poles or branch cuts, place the Bromwich contour, and walk you through the residues or keyhole integral step by step.

Do I always need the Bromwich integral, or can I just use a table?

For everyday transforms a table plus partial fractions is faster, and that is what most calculations use. But the Bromwich integral is the definition guaranteeing the table exists and is unique, and it is indispensable in three cases: when $F$ has a branch point (tables of exponentials will not help — you need the keyhole cut integral); when you want long-time asymptotics from the rightmost singularity without fully inverting; and when $F$ is known only numerically. Think of the contour integral as the master formula and the table as a cache of its common outputs.

Does the answer depend on which value of c I pick for the contour?

No, as long as $c$ lies to the right of every singularity ( $c>\sigma_0$ ). Shifting the line from $c_1$ to $c_2$ sweeps the contour through a strip where $F(s)e^{st}$ is analytic, so by Cauchy's theorem the integral is unchanged. What you may not do is move $c$ across a singularity: crossing a pole changes the enclosed residues and gives a different, non-causal function. So $c$ is free within the half-plane of analyticity and forbidden outside it.

Why does closing the contour left versus right encode causality?

Because the integrand's growth is set by $|e^{st}|=e^{t\,\operatorname{Re}s}$ . For $t>0$ this decays as $\operatorname{Re}s\to-\infty$ , so only a left-closing arc vanishes — and the left half-plane is where all the poles live, so you collect them. For $t<0$ it decays as $\operatorname{Re}s\to+\infty$ , forcing a right-closing arc; but $F$ is analytic there, so the loop encloses nothing and $f(t)=0$ . The sign of $t$ dictates the closure, the closure dictates which singularities are seen, and right-half-plane analyticity guarantees none are seen for negative time. Causality is not assumed but derived.

What is the practical difference between a pole and a branch cut in the answer?

A pole is an isolated singularity contributing a single sharp mode: a simple pole gives $e^{s_k t}$ , a double pole $t\,e^{s_k t}$ , and so on — a finite sum of modes with definite decay rates. A branch cut is a continuous locus of non-analyticity that no finite residue sum can capture; a keyhole contour turns the inversion into an integral over a continuum of exponentials, which superpose into a non-exponential shape like $t^{-1/2}$ or an error function. Physically, poles signal discrete oscillators (circuits, normal modes), branch cuts a continuous spectrum (diffusion, radiation into open media).

Why is numerical Laplace inversion notoriously hard?

Because the problem is ill-posed: the forward transform smooths aggressively — fast-varying features of $f(t)$ are exponentially damped in $F(s)$ — so undoing it must exponentially amplify small errors. Two very different signals can produce transforms agreeing to many digits, so tiny noise in sampled $F(s)$ can swing the reconstructed $f(t)$ wildly. Good algorithms (Talbot, Gaver–Stehfest, Weeks) tame this with steepest-descent contours, high-precision arithmetic, and regularization, but there is no general cure. The same pathology appears in reconstructing spectral functions from lattice-QCD correlators, stabilized by Bayesian and maximum-entropy methods.

How does the convolution theorem help me invert a product of two transforms?

When $F=GH$ factors into transforms you already know how to invert, you can skip partial fractions: the inverse of the product is the causal convolution $\mathcal L^{-1}\{GH\}=\int_0^t g(\tau)h(t-\tau)\,\mathrm{d}\tau$ . This is exactly how a system's output is its impulse response $h$ convolved against an input $g$ , since the transfer function multiplies the input transform in the $s$ -domain. It is most valuable when partial-fractioning the product would be painful but the two factors are elementary — e.g. inverting $1/[(s^2+\omega^2)(s+a)]$ as a convolution of $\sin\omega t/\omega$ with $e^{-at}$ .