The Laplace Transform

The Laplace transform is built for initial-value problems and causal systems — the tool a physicist reaches for the moment a differential equation comes bundled with the words "at $t=0$ ". Given a function $f(t)$ defined for $t\ge 0$ , it produces $F(s)=\int_0^\infty f(t)e^{-st}\,\mathrm{d}t$ , where $s$ is a complex frequency. Because the integration starts at the origin and the kernel $e^{-st}$ decays whenever $\mathrm{Re}\,s$ is large, the transform exists for any $f$ growing no faster than some exponential $Me^{\sigma_0 t}$ ; the region $\mathrm{Re}\,s>\sigma_0$ is the half-plane of convergence and $\sigma_0$ its abscissa. That single feature — a real, adjustable damping in the kernel — is why the Laplace transform tames functions the Fourier transform rejects. Conceptually it is the Fourier transform with the frequency allowed to go complex: folding the factor $e^{-\sigma t}$ into the signal recovers the Fourier transform of $f(t)e^{-\sigma t}\theta(t)$ — same machinery, extra convergence, at the cost of restricting to $t\ge 0$ and losing the symmetric unitary structure of Fourier theory. What you gain is the operational dictionary that makes the transform indispensable: differentiation in $t$ becomes multiplication by $s$ plus initial data, integration becomes division by $s$ , and convolution becomes ordinary multiplication. A linear constant-coefficient ODE therefore collapses to an algebraic equation for $F(s)$ whose inverse automatically satisfies the prescribed initial conditions — no constants of integration to fit afterward. Recovering $f(t)$ is the Bromwich integral up a vertical line to the right of every singularity, in practice done by partial fractions and residues, and the poles of $F(s)$ read off the dynamics directly: left half-plane means decay, the imaginary axis oscillation, the right half-plane instability.

Key takeaways

The Laplace transform $F(s)=\int_0^\infty f(t)\,e^{-st}\,\mathrm{d}t$ maps a function of time $t\ge 0$ to a function of complex frequency $s=\sigma+i\omega$ , converging in a half-plane $\mathrm{Re}\,s>\sigma_0$ .
Its defining superpower is the derivative rule $\mathcal{L}\{f'\}=sF(s)-f(0)$ , which turns a linear ODE with initial data into an algebraic equation and bakes the initial conditions straight into the answer.
Unlike the Fourier transform, the kernel $e^{-st}$ supplies exponential damping, so it handles functions that grow polynomially or exponentially and is the natural language of causal, one-sided signals.
Convolution in time becomes multiplication, $\mathcal{L}\{f*g\}=F(s)G(s)$ — the engine behind transfer functions, impulse responses, and causal Green's functions.
Inversion runs up the Bromwich contour $f(t)=\frac{1}{2\pi i}\int_{\gamma-i\infty}^{\gamma+i\infty}F(s)e^{st}\,\mathrm{d}s$ , done in practice by partial fractions and residues; the poles of $F(s)$ encode decay, oscillation, and stability.
Key idea: replace the oscillatory Fourier kernel $e^{-i\omega t}$ by the damped one-sided kernel $e^{-st}$ with $s$ complex — differentiation becomes multiplication by $s$ (with initial data as boundary terms), turning calculus into algebra in a half-plane of convergence.

The core insight

Replace the oscillatory Fourier kernel $e^{-i\omega t}$ by the damped one-sided kernel $e^{-st}$ with $s$ complex: this converts differentiation into multiplication by $s$ (with initial data appearing as boundary terms), turning calculus problems into algebra in a half-plane of convergence.

Where this lives today

The Laplace transform is the working language of everything that responds to an input. Classical control theory — the mathematics that flies aircraft, stabilises rockets, and tunes the feedback loop of a switching power supply — is built on the transfer function $H(s)$ and its pole-zero map; engineers place poles in the left half-plane to guarantee stability and read damping, bandwidth, and resonance straight off the $s$ -plane. Every circuit simulator in the SPICE lineage solves networks in the $s$ -domain, a capacitor becoming impedance $1/(sC)$ . In physics the same machinery builds retarded Green's functions and propagators wherever causality demands analyticity in a half-plane, and it fits relaxation data — NMR and MRI $T_1/T_2$ decays, dielectric and viscoelastic spectra, fluorescence lifetimes — by inverse Laplace transform. That last use exposes a genuinely open problem: numerical inversion is famously ill-posed, because very different $f(t)$ can produce almost identical $F(s)$ on the real axis, so recovering a relaxation-time distribution from noisy data is an active research area (Talbot, Stehfest, Gaver, and CME methods all trade stability against resolution, none universally reliable). The transform that makes forward problems trivial makes the inverse one of the hardest routine tasks in scientific computing.

The physical picture

Think of $s$ as a two-part probe. Its real part $\sigma=\mathrm{Re}\,s$ sets an exponential weighting — how aggressively you discount the future — while its imaginary part $\omega=\mathrm{Im}\,s$ probes oscillation, exactly as a Fourier frequency does. Dial $\sigma$ up high enough and you suppress any exponentially bounded growth, which is why the Laplace transform exists for signals the Fourier integral cannot swallow. The transform then acts as a spectral microscope for one-sided signals: the places where $F(s)$ blows up — its poles — are the natural modes of the system, their real parts setting decay or growth rates and their imaginary parts oscillation frequencies. Reading $f(t)$ off $F(s)$ is reading the dynamics off the geometry of the $s$ -plane.

Figure

The s-plane pole map of the damped oscillator x'' + 6x' + 25x = 0: the conjugate poles sit at s = −3 ± 4i, whose real part −3 fixes the decay rate e^(−3t) and whose imaginary part 4 fixes the oscillation frequency ω_d = 4 rad/s

Because differentiation becomes multiplication by $s$ with the initial value subtracted, the transform never forgets where the system started — an ODE plus its initial data is a complete physical question, answered in one algebraic stroke. Convolution becoming multiplication turns a tangled cause-and-effect integral into a plain product of input spectrum and transfer function, the bedrock of linear-systems theory. None of this is free: the integral only converges in a half-plane, and where it begins — the abscissa $\sigma_0$ — is set by the fastest growth in $f$ . The second figure shows that convergence geometry and the vertical Bromwich contour used to climb back to $f(t)$ .

Figure

Half-plane of convergence for f(t) = e^(2t): the defining integral converges only where Re s > σ₀ = 2, giving F(s) = 1/(s − 2) with a single pole at s = 2, and the Bromwich inversion contour runs up the vertical line Re s = γ chosen anywhere with γ > 2

The core equations

The starting point is the definition itself: a one-sided integral against the damped kernel, defined only where it converges. Everything operational follows from differentiating or shifting inside this one formula.

F(s)=\mathcal{L}\{f\}(s)=\int_0^{\infty} f(t)\,e^{-st}\,\mathrm{d}t,\qquad \mathrm{Re}\,s>\sigma_0 \qquad (1)

where: $f(t)$ — original function, defined for $t\ge 0$ , of exponential order; $F(s)$ — the transform, analytic for $\mathrm{Re}\,s>\sigma_0$ ; $s=\sigma+i\omega$ — complex frequency variable; $\sigma_0$ — abscissa of convergence, the smallest $\sigma$ for which the integral converges. Reading: $F(s)$ is a weighted census of $f$ , each instant $t$ weighted by the decaying exponential $e^{-st}$ . The derivative rule is the engine of the whole subject. Integrating by parts trades a time derivative for a factor of $s$ and drops the initial value out as a boundary term — which is why transformed ODEs remember their initial data.

\mathcal{L}\{f'\}=sF(s)-f(0^+),\qquad \mathcal{L}\{f''\}=s^2F(s)-sf(0^+)-f'(0^+) \qquad (2)

where: $f(0^+)$ — right-hand initial value of $f$ ; $f'(0^+)$ — right-hand initial value of $f'$ . Reading: each derivative brings down one power of $s$ and peels off one piece of initial data; an $n$ -th order ODE needs exactly $n$ initial values, and the rule supplies slots for all of them. Integration is the inverse operation and does the inverse thing to $s$ ; a time shift and a frequency shift are dual to one another, each trading a multiplicative exponential for a translation. These three rules cover most hand calculations.

\mathcal{L}\Big\{\int_0^t f(\tau)\,\mathrm{d}\tau\Big\}=\frac{F(s)}{s},\qquad \mathcal{L}\{e^{at}f(t)\}=F(s-a),\qquad \mathcal{L}\{f(t-a)\theta(t-a)\}=e^{-as}F(s) \qquad (3)

where: $a$ — real shift constant ( $a>0$ for the causal delay); $\theta$ — Heaviside step. Reading: dividing by $s$ integrates, multiplying the signal by $e^{at}$ slides $F$ sideways by $a$ , and delaying by $a$ multiplies $F$ by $e^{-as}$ . The convolution theorem is what makes linear-systems theory algebraic: the messy overlap integral of input and impulse response is simply a product in the $s$ -domain.

\mathcal{L}\{(f*g)(t)\}=F(s)\,G(s),\qquad (f*g)(t)=\int_0^t f(\tau)\,g(t-\tau)\,\mathrm{d}\tau \qquad (4)

where: $f*g$ — causal (one-sided) convolution; $G(s)$ — transform of $g$ . Reading: a linear system's response is the input's transform times the transfer function, inverted — no integration until the very end. Recovering $f(t)$ from $F(s)$ is the Bromwich inversion integral, a contour running up a vertical line to the right of every singularity of $F$ ; closing it to the left and summing residues is the practical route.

f(t)=\frac{1}{2\pi i}\int_{\gamma-i\infty}^{\gamma+i\infty} F(s)\,e^{st}\,\mathrm{d}s,\qquad \gamma>\sigma_0 \qquad (5)

where: $\gamma$ — real abscissa of the vertical contour, right of all singularities of $F$ . Reading: each pole contributes a residue $\propto e^{s_k t}$ , so the time answer is a sum of exponentials whose rates are the pole positions. In practice one rarely evaluates (5) from scratch. A short table of elementary transforms plus partial fractions covers the rational $F(s)$ from constant-coefficient ODEs, and the two distributional entries close the loop with the delta and step.

\mathcal{L}\{t^n\}=\frac{n!}{s^{n+1}},\quad \mathcal{L}\{e^{at}\}=\frac{1}{s-a},\quad \mathcal{L}\{\sin\omega t\}=\frac{\omega}{s^2+\omega^2},\quad \mathcal{L}\{\cos\omega t\}=\frac{s}{s^2+\omega^2},\quad \mathcal{L}\{\delta(t)\}=1,\quad \mathcal{L}\{\theta(t)\}=\frac{1}{s} \qquad (6)

where: $n$ — non-negative integer, generalising to $\Gamma(\nu+1)/s^{\nu+1}$ for $\mathrm{Re}\,\nu>-1$ ; $a$ — constant, convergence for $\mathrm{Re}\,s>\mathrm{Re}\,a$ ; $\delta$ — Dirac impulse; $\theta$ — unit step. Reading: powers give poles at the origin, exponentials give a simple pole at $s=a$ , sinusoids give a conjugate pole pair on the imaginary axis, and the impulse transforms to the constant $1$ .

Derivation AThe derivative rule from integration by parts

The goal is to establish the single most important operational identity of the whole subject — that a time derivative $f'$ transforms to $sF(s)-f(0^+)$ — because it is exactly this rule that converts a linear differential equation into an algebraic one and threads the initial conditions through as boundary terms rather than as constants to be fixed afterward. We start from the definition applied to $f'$ , integrate by parts once, and show the boundary term at infinity dies inside the half-plane of convergence while the boundary term at zero delivers the initial value.

Step 1. Write the transform of the derivative straight from the definition, $\mathcal{L}\{f'\}=\int_0^\infty f'(t)\,e^{-st}\,\mathrm{d}t$ . Why: everything must trace back to equation (1); we are simply asking what the definition says when its argument is $f'$ .

Step 2. Integrate by parts with $u=e^{-st}$ and $\mathrm{d}v=f'\,\mathrm{d}t$ , so $\mathcal{L}\{f'\}=\big[f(t)e^{-st}\big]_0^\infty+s\int_0^\infty f(t)e^{-st}\,\mathrm{d}t$ . Why: integration by parts is the one move that trades the derivative on $f$ for a derivative on the kernel, and differentiating $e^{-st}$ just brings down a factor $-s$ , converting calculus into multiplication.

Step 3. Evaluate the boundary term. At the upper limit $f(t)e^{-st}\to 0$ because $f$ is of exponential order and $\mathrm{Re}\,s>\sigma_0$ makes $e^{-st}$ win; at the lower limit it gives $-f(0^+)$ . The surviving integral is $sF(s)$ , so $\mathcal{L}\{f'\}=sF(s)-f(0^+)$ . Why: the vanishing of the upper boundary term is exactly what the abscissa of convergence guarantees, and it is the only place the half-plane restriction is used.

Step 4. Iterate. Applying the same rule to $f''=(f')'$ gives $\mathcal{L}\{f''\}=s\,\mathcal{L}\{f'\}-f'(0^+)=s^2F(s)-sf(0^+)-f'(0^+)$ , and by induction the $n$ -th derivative carries $n$ initial values. Why: this is why an $n$ -th order ODE with $n$ prescribed initial values transforms to a single algebraic equation with no undetermined constants left over. *As a quick check: at $n=1$ the boundary term is $-f(0^+)$ , and direct substitution of $f=e^{at}$ gives $\mathcal{L}\{ae^{at}\}=a/(s-a)=s\cdot\frac{1}{s-a}-1$ , consistent since $f(0^+)=1$ .*

Go deeper · Derivation B · 1 minSolving the damped harmonic oscillator by Laplace transform

Derivation BSolving the damped harmonic oscillator by Laplace transform

The goal is to run the whole method end to end on a physical system — the underdamped oscillator $x''+2\beta x'+\omega_0^2 x=0$ with $x(0)=x_0$ and $x'(0)=v_0$ — and show that transforming turns the differential equation into simple algebra, that the initial data enter automatically through the derivative rule, and that inverting by completing the square hands back the familiar damped sinusoid without ever fitting a constant of integration. This is the template for every constant-coefficient initial-value problem in mechanics and circuits.

Step 1. Transform each term with the derivative rule, giving $\big(s^2X-sx_0-v_0\big)+2\beta\big(sX-x_0\big)+\omega_0^2X=0$ . Why: the second-derivative rule injects both $x_0$ and $v_0$ , and the first-derivative rule injects $x_0$ again, so all the initial information is now sitting in the algebraic equation.

Step 2. Collect $X$ and move the data to the right, $X(s)\big(s^2+2\beta s+\omega_0^2\big)=sx_0+v_0+2\beta x_0$ . Why: the differential operator has become the polynomial $s^2+2\beta s+\omega_0^2$ , the characteristic polynomial of the ODE, and the forcing plus initial data become its right-hand side.

Step 3. Solve and complete the square, $X(s)=\dfrac{(s+2\beta)x_0+v_0}{(s+\beta)^2+\omega_d^2}$ with $\omega_d=\sqrt{\omega_0^2-\beta^2}$ . Why: writing $s^2+2\beta s+\omega_0^2=(s+\beta)^2+\omega_d^2$ exposes a shifted sine/cosine denominator, and the frequency shift $s\to s+\beta$ is exactly what will produce the decay envelope.

Step 4. Invert using the shift theorem together with the sine and cosine entries, $x(t)=e^{-\beta t}\Big(x_0\cos\omega_d t+\dfrac{v_0+\beta x_0}{\omega_d}\sin\omega_d t\Big)$ . Why: the $s\to s+\beta$ shift supplies the $e^{-\beta t}$ factor and splitting the numerator into an $s$ -piece and a constant piece yields the cosine and sine; the poles sit at $s=-\beta\pm i\omega_d$ , whose real part is the decay rate and imaginary part the ringing frequency. *As a check: at $t=0$ , $x(0)=x_0$ and differentiating gives $x'(0)=v_0$ directly; the poles $s=-\beta\pm i\omega_d$ correctly encode a damped oscillation.*

Go deeper · Derivation C · 1 minThe causal Green's function as an inverse transform

Derivation CThe causal Green's function as an inverse transform

The goal is to show that the impulse response of any linear constant-coefficient operator is nothing but the inverse Laplace transform of the reciprocal of its characteristic polynomial, and that this response is automatically causal because the one-sided transform builds causality in from the start. This single result underlies transfer functions in control theory, retarded Green's functions in field theory, and response functions throughout many-body physics — all a direct consequence of the operational rules already in hand.

Step 1. Pose the operator driven by a unit impulse from rest, $\big(\sum_{k=0}^n a_k \tfrac{\mathrm{d}^k}{\mathrm{d}t^k}\big)G(t)=\delta(t)$ with $G,G',\dots$ all zero at $t=0^-$ . Why: quiescent initial conditions before the impulse are the physical statement of causality — nothing happens until the kick arrives — and they make every initial-data boundary term vanish.

Step 2. Transform. With zero initial data the derivative rule collapses to $s^k\tilde G(s)$ for each term, so $P(s)\tilde G(s)=\mathcal{L}\{\delta\}=1$ where $P(s)=\sum_k a_k s^k$ . Why: the impulse transforms to the constant $1$ (equation 6), and the characteristic polynomial $P(s)$ is precisely the differential operator with $\mathrm{d}/\mathrm{d}t$ replaced by $s$ .

Step 3. Solve in frequency, $\tilde G(s)=\dfrac{1}{P(s)}$ , the transfer function of the system. Why: the reciprocal of the characteristic polynomial is the system's fingerprint — its poles are the natural modes, and their location in the $s$ -plane decides stability.

Step 4. Invert and convolve for arbitrary forcing, $x(t)=(G*f)(t)=\int_0^t G(t-\tau)f(\tau)\,\mathrm{d}\tau$ . Why: the convolution theorem run backwards says the response to any input is the input convolved with the impulse response, and because $G(t)=0$ for $t<0$ the integral only reaches back over the past — a manifestly causal Green's function. *This is consistent with the convolution theorem, and the inversion of $1/P(s)$ is given explicitly by the residue sum $\sum_k \mathrm{Res}_{s_k}\,e^{s_k t}/P'(s_k)$ over the roots $s_k$ of $P$ .*

Limiting cases

How the Laplace transform connects to its neighbours and behaves in special regimes:

Case	Result	Meaning
Fourier limit	$s=i\omega,\ \sigma\to 0^+$	For a signal with abscissa $\sigma_0\le 0$ , putting $\sigma=0$ recovers the one-sided Fourier transform $F(i\omega)=\int_0^\infty f(t)e^{-i\omega t}\,\mathrm{d}t$
Initial-value theorem	$f(0^+)=\lim_{s\to\infty} sF(s)$	The large- $s$ tail of $F$ encodes the short-time behaviour of $f$ ; useful for reading off jumps
Final-value theorem	$\lim_{t\to\infty} f(t)=\lim_{s\to 0^+} sF(s)$	Valid only when every pole of $sF(s)$ lies strictly in the left half-plane; fails for oscillatory or growing $f$
Pole location	$\mathrm{Re}(s_k)<0\Rightarrow$ decay	A simple pole at $s_k$ contributes $\sim e^{s_k t}$ : left half-plane decays, imaginary axis oscillates, right half-plane blows up
Pure delay	$\mathcal{L}\{f(t-a)\theta(t-a)\}=e^{-as}F(s)$	A time delay multiplies by the non-rational factor $e^{-as}$ — a transcendental transfer function, no finite pole set

Common trapThe two traps: convergence domain and initial-value bookkeeping

Two mistakes dominate. First, the abscissa of convergence: $F(s)$ is only defined by its integral for $\mathrm{Re}\,s>\sigma_0$ , and you must not read formulae outside that domain as if they were the integral — the closed form $1/(s-2)$ is an analytic continuation of the transform of $e^{2t}$ , but the defining integral converges only for $\mathrm{Re}\,s>2$ , and the pole at $s=2$ is exactly the abscissa. Second, initial-value bookkeeping: the derivative rule uses the right-hand limit $f(0^+)$ , but for impulsive inputs the $0^-$ convention (which captures the jump the impulse produces) is often the physically correct one. Mixing conventions corrupts every transient term, so decide up front whether your $\delta(t)$ lives just before or just after $t=0$ and use the matching limit throughout.

Worked examples

1 · The damped oscillator, fully numerical

A mass on a spring obeys $x''+6x'+25x=0$ with $x(0)=1$ and $x'(0)=0$ (units of rad/s). Find $x(t)$ , locate the poles of $X(s)$ , and read off the decay rate and ringing frequency.

Step 1. Match to the standard form $x''+2\beta x'+\omega_0^2 x=0$ : here $2\beta=6$ so $\beta=3\ \mathrm{s}^{-1}$ , and $\omega_0^2=25$ so $\omega_0=5\ \mathrm{rad/s}$ . The system is underdamped because $\omega_0>\beta$ , with $\omega_d=\sqrt{\omega_0^2-\beta^2}=\sqrt{25-9}=\mathbf{4\ rad/s}$ .

Step 2. Transform and solve. With $x_0=1,\ v_0=0$ the oscillator formula gives $X(s)=\dfrac{(s+2\beta)x_0+v_0}{(s+\beta)^2+\omega_d^2}=\dfrac{s+6}{(s+3)^2+16}$ . The poles are the zeros of the denominator, $s=-3\pm 4i$ .

Step 3. Invert with the shift and sine/cosine rules, splitting the numerator as $s+6=(s+3)+3$ : $x(t)=e^{-3t}\Big(\cos 4t+\tfrac{3}{4}\sin 4t\Big)$ . The $e^{-3t}$ factor is the frequency shift $s\to s+3$ ; the constant $3$ over $\omega_d=4$ builds the sine coefficient $0.75$ .

Step 4. Read the physics off the poles: the real part $-3\ \mathrm{s}^{-1}$ is the decay rate (the envelope falls by $1/e$ every $1/3\ \mathrm{s}$ ) and the imaginary part $\pm 4$ is the ringing frequency. Check the initial conditions: $x(0)=1$ , and $x'(0)=-3+3=0$ , both satisfied automatically. Poles at s = −3 ± 4i give x(t) = e^(−3t)(cos 4t + 0.75 sin 4t): decay rate 3 s⁻¹, ring frequency ω_d = 4 rad/s.

2 · Inverting a rational transform by partial fractions

Solve the overdamped equation $y''+3y'+2y=0$ with $y(0)=1$ , $y'(0)=0$ by Laplace transform, using partial fractions to invert.

Step 1. Transform. Using $\mathcal{L}\{y''\}=s^2Y-sy(0)-y'(0)$ and $\mathcal{L}\{y'\}=sY-y(0)$ : $\big(s^2Y-s\big)+3\big(sY-1\big)+2Y=0$ , so $Y(s^2+3s+2)=s+3$ and $Y(s)=\dfrac{s+3}{(s+1)(s+2)}$ , factoring the characteristic polynomial $s^2+3s+2$ .

Step 2. Partial fractions. Write $\dfrac{s+3}{(s+1)(s+2)}=\dfrac{A}{s+1}+\dfrac{B}{s+2}$ . Cover-up at $s=-1$ : $A=\dfrac{s+3}{s+2}\big|_{s=-1}=\dfrac{2}{1}=\mathbf{2}$ . Cover-up at $s=-2$ : $B=\dfrac{s+3}{s+1}\big|_{s=-2}=\dfrac{1}{-1}=\mathbf{-1}$ .

Step 3. Invert term by term with $\mathcal{L}^{-1}\{1/(s-a)\}=e^{at}$ : $y(t)=2e^{-t}-e^{-2t}$ . Both poles are real and negative ( $s=-1,-2$ ), so the motion is a pure decaying sum with no oscillation — the signature of overdamping.

Step 4. Verify. $y(0)=2-1=\mathbf{1}$ ; differentiating, $y'(t)=-2e^{-t}+2e^{-2t}$ so $y'(0)=-2+2=\mathbf{0}$ . Both initial conditions hold, and the initial-value theorem confirms $f(0^+)=\lim_{s\to\infty} sY(s)=\lim_{s\to\infty}\dfrac{s^2+3s}{s^2+3s+2}=1$ . y(t) = 2e^(−t) − e^(−2t), a sum of two decaying modes at rates 1 and 2 s⁻¹.

Physics of the portrait

The Laplace transform is where a century and a half of applied mathematics converges: an integral Euler and Laplace wrote down for probability, a symbolic calculus Heaviside forced onto reluctant mathematicians to solve telegraph transients, and finally the rigorous contour that justified all of it.

Origins · 1 minLaplace transform and how this was discovered

Laplace transform

the operational calculus, 1780s–1930s · integral transforms and linear systems

The transform carries Pierre-Simon Laplace's name, but its history is a relay. Leonhard Euler used integrals of the form $\int e^{-st}f(t)\,\mathrm{d}t$ in the 1730s and 1740s to solve differential equations, and Laplace developed them systematically in his 1785 memoir on probability, where the transform of a density is essentially its generating function. For a century it stayed a mathematician's tool. The revolution came from an outsider: Oliver Heaviside, a self-taught telegraph engineer, who in the 1880s and 1890s built an "operational calculus" that treated $\mathrm{d}/\mathrm{d}t$ as an algebraic symbol and solved circuit transients by pure manipulation — getting answers that were unmistakably right even as mathematicians recoiled at the missing rigour ("Shall I refuse my dinner because I do not fully understand the process of digestion?"). That rigour arrived from Thomas John I'Anson Bromwich, who in 1916 showed Heaviside's operators were Laplace transforms and that inversion was a contour integral — equation (5). Gustav Doetsch's 1937 treatise fixed the modern definition and theory. The result is the standard grammar of linear systems: what Snel's law is to optics, the pole-zero map is to control and circuits.

History

1737

— Euler employs integrals of the form ∫ e^(−st) f(t) dt to solve differential equations

1785

— Laplace develops the transform in his memoir on probability generating functions

1893

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

1916

— Bromwich derives the complex-contour inversion integral that makes the method rigorous

1937

— Doetsch publishes the systematic modern theory of the Laplace transformation

Going further

Texts. Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 20 (integral transforms) · Stone & Goldbart, Mathematics for Physics, Ch. on integral transforms · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 13 · Hassani, Mathematical Physics, Ch. on integral transforms and complex analysis · Byron & Fuller, Mathematics of Classical and Quantum Physics, §5 · Churchill, Operational Mathematics, 3rd ed. (the classic dedicated treatment). Key results to keep in mind: the integration-by-parts derivation gives $\mathcal{L}\{f'\}=sF(s)-f(0^+)$ ; the elementary table includes $\mathcal{L}\{\sin\omega t\}=\omega/(s^2+\omega^2)$ (derivable from the exponential rule); the damped-oscillator solution $x(t)=e^{-3t}(\cos 4t+0.75\sin 4t)$ satisfies $x''+6x'+25x=0$ with $x(0)=1,\ x'(0)=0$ and has poles at $s=-3\pm 4i$ ; the partial-fraction inversion gives $\mathcal{L}^{-1}\{(s+3)/((s+1)(s+2))\}=2e^{-t}-e^{-2t}$ ; the Green's-function reciprocal is $\tilde G=1/P(s)$ ; and all factorial constants satisfy $\mathcal{L}\{t^n\}=n!/s^{n+1}$ .

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

Ask the tutor: Give me any ODE with its initial conditions and I will walk you from the $s$ -domain algebra to the inverted answer, pointing out where the poles put the physics.

What is the actual difference between the Laplace and Fourier transforms?

They share a kernel and differ in two ways. Fourier uses the purely oscillatory $e^{-i\omega t}$ over the whole real line; Laplace uses $e^{-st}$ with $s=\sigma+i\omega$ complex, integrated only from $t=0$ . The real part $\sigma$ supplies exponential damping, so Laplace converges for growing signals (like $e^{2t}$ ) that Fourier cannot handle, and the one-sided range makes it the natural tool for causal problems with a definite start. The price is asymmetry: no clean Plancherel/unitary structure, and inversion needs a contour. If a signal's abscissa is $\sigma_0\le 0$ , evaluating the Laplace transform on the imaginary axis $s=i\omega$ reproduces its one-sided Fourier transform — Fourier is the boundary case of Laplace.

Why does the initial value f(0) appear in the derivative rule, and is it 0⁻ or 0⁺?

It appears as the boundary term when you integrate the definition by parts — the lower limit contributes $-f(0)$ . Which limit you use depends on where the integration starts. The standard one-sided transform integrates from just above zero, giving the right-hand value $f(0^+)$ , correct for smooth initial data. But for impulsive inputs — a $\delta(t)$ kick at the origin — the response can jump across $t=0$ , and then the $0^-$ convention (capturing the state before the impulse) is usually what the physics wants, because it lets the impulse itself produce the jump rather than hiding it in the initial condition. Pick one convention, and keep your $\delta(t)$ and your initial values on the same side of zero.

Do I ever actually have to compute the Bromwich contour integral?

Rarely, for the problems you meet by hand. Almost every $F(s)$ from a constant-coefficient ODE is a rational function, inverted by partial fractions plus a six-line table — no contour needed. The Bromwich integral matters conceptually (it is the inverse, and it explains why the answer is a sum of exponentials, one per pole) and practically when $F(s)$ is not rational: branch cuts from $\sqrt{s}$ in diffusion and heat-conduction problems, or the $e^{-as}$ delay factors of distributed systems, force you to deform the contour and integrate around the cuts. For everything polynomial-over-polynomial, partial fractions are the Bromwich integral in disguise.

What is the abscissa of convergence and why should I care?

The abscissa $\sigma_0$ is the vertical line $\mathrm{Re}\,s=\sigma_0$ to the right of which the defining integral converges; it equals the exponential growth rate of $f$ . For $f=e^{2t}$ it is $\sigma_0=2$ , because $\int_0^\infty e^{(2-s)t}\,\mathrm{d}t$ converges only when $\mathrm{Re}\,s>2$ . You care for two reasons. First, the Bromwich contour must lie to the right of $\sigma_0$ (hence right of every pole) for inversion to reconstruct $f$ . Second, the closed form $F(s)=1/(s-2)$ is an analytic continuation valid everywhere except the pole, but the integral only means anything in its half-plane — so convergence, stability, and the final-value theorem all hinge on where $\sigma_0$ sits relative to the imaginary axis.

Why does the final-value theorem sometimes give a wrong answer?

Because it has a hypothesis that is easy to forget. The theorem $\lim_{t\to\infty}f(t)=\lim_{s\to 0^+}sF(s)$ holds only when the left-hand limit exists, which requires every pole of $sF(s)$ to lie strictly in the open left half-plane. The classic trap is $f(t)=\sin\omega t$ : here $sF(s)=\omega s/(s^2+\omega^2)\to 0$ as $s\to 0$ , so the formula naively predicts $f\to 0$ — yet $\sin\omega t$ never settles. The catch is that $sF(s)$ has poles on the imaginary axis at $s=\pm i\omega$ , violating the hypothesis. Anything that oscillates forever or grows breaks the theorem the same way — always check the pole locations first.

What does a pole in the right half-plane mean physically?

Instability. By the inversion formula each simple pole at $s_k$ contributes a term $\propto e^{s_k t}$ , so a pole with positive real part gives a mode that grows exponentially — in a control system, the signature of a runaway feedback loop that amplifies its own error until it saturates or destroys the plant. This is why control design is, at bottom, about moving poles: you add feedback to drag every closed-loop pole into the left half-plane, where its mode decays. Poles exactly on the imaginary axis are the marginal case (undamped oscillation, an ideal LC circuit); poles in the left half-plane give the decaying transients of any real damped system.

Is the one-sided or the two-sided transform the "real" Laplace transform?

Both exist, but in physics the one-sided transform, integrating from $0$ to $\infty$ , is the default unless stated otherwise. Its restriction to $t\ge 0$ is a feature: it matches causal problems that start at a definite instant, and it is what makes the derivative rule carry initial data. The two-sided (bilateral) transform integrates over the whole real line and appears in probability as the moment generating function, but it converges only in a strip rather than a half-plane and cannot encode initial conditions, so it is far less useful for differential equations. When a physicist says "Laplace transform" unqualified, assume one-sided.