Why this matters for simulation: Transforms convert a differential problem into an algebraic one — Fourier transforms do the same for PDEs in space.
By the end of this lesson you can…
Write the Laplace transform $F(s) = \int_0^\infty f(t)e^{-st}\,dt$ as a limit of finite integrals, and say what the number $s$ does to the integrand.
State when the integral converges (exponential order, $\operatorname{Re}s \gt a$), and show numerically why $e^{t^2}$ has no transform.
Derive the transforms of $1$, $t^n$, $e^{at}$, $\sin\omega t$, $\cos\omega t$, $\sinh\omega t$ and $\cosh\omega t$ by integrating, and check each against quadrature.
Use linearity and the first shifting theorem to transform products like $e^{at}\cos\omega t$ without a new integral.
Derive $\mathcal L\{y'\} = sY - y(0)$ and $\mathcal L\{y''\} = s^2Y - sy(0) - y'(0)$ by parts, and explain why they turn a linear ODE into algebra.
Invert a rational $F(s)$ by partial fractions (distinct, repeated and complex roots) and tie its poles to the characteristic polynomial.
Every method so far attacks an ODE directly: guess ert, match coefficients, integrate, or step forward with RK4. This lesson changes the question. Instead of solving for y(t), we trade the whole function for a single new function Y(s) in which differentiation has become multiplication. An ODE becomes a polynomial equation, we solve it with algebra, and we translate back. The trade works because of one fact about ert that 5.2 Constant Coefficients & the Characteristic Equation relied on: differentiating it only multiplies it by r. Here the exponential becomes a measuring probe instead of a guess. This lesson builds the transform, its table and its derivative rule; 6.2 Solving Initial Value Problems uses them to solve initial value problems, and later lessons add switching forcing, impulses and convolution.
▶ VideoThe probe e^(-st) multiplying f, the area as s varies, exponential order and divergence, the table, the derivative rule, and inversion by partial fractionscaptions only
Transcript
Every method so far attacks the equation directly. Here we trade y of t for a function Y of s, where differentiation becomes multiplication. The ODE becomes algebra: transform, solve, invert. It all rests on one probe, e to the minus s t.
The transform multiplies f of t by the decaying weight e to the minus s t, and takes the area under the product from zero to infinity. One number for each s. The curve through them is F of s.
Take f equal to one. A large s kills the weight quickly, so F sees only the start; a small s lets it linger. The area tends to one over s, 0.5 at s equals 2; recording s point by point builds the curve. For s of zero or less, no limit.
Now f equals e to the t, a equals 1. At s equals 2 the area is 1.000, one over s minus a. At s equals 1.5, cut at T equals 12, it is 1.995 against 2; the tail is 0.005. Swept over s, the points follow the table right of s equals a.
Below a it is worse. At s equals 1 the area is exactly T. At s equals 0.5 it is 804.9 at T equals 12, 3.255 times 10 to the 5 at 24, and 9.703 times 10 to the 8 at 40. The formula's minus 2 is algebra, not the integral.
Now e to the t squared. At s equals 5 the area to T equals 4 is 0.2309, which looks converged. To 6 it is 60.6, to 8 it is 2.45 times 10 to the 9. The exponent t squared minus 5 t dips at 2.5, then climbs for every s: no transform.
Now the table. For t to the n, parts leaves n over s times the transform of t to the n minus 1. Unwinding gives n factorial over s to the n plus 1: 0.25, 0.25 and 0.375 at s equals 2. For e to the a t, one over s minus a.
For sine and cosine, transform e to the i omega t: s plus i omega over s squared plus omega squared. The real part is cosine, the imaginary part sine; at omega 2 and s 1, 0.2 and 0.4. Sinh and cosh follow from exponentials: 0.4 and 0.6 at s equals 3.
The integral is linear, so sums split. A factor e to the a t slides the transform by a, since the integrand becomes f times e to the minus s minus a t. So e to the t cosine 2 t at s equals 2 has area 0.2, like cosine 2 t at s equals 1.
Integrate by parts. The boundary term at infinity vanishes; the one at zero leaves minus y of 0. So y prime becomes s Y minus y of 0, and y double prime s squared Y minus s y of 0 minus y prime of 0. The probe is an eigenfunction of d dt, so a derivative costs a factor s. Check: minus 0.8 both ways.
Take y double prime plus 3 y prime plus 2 y equals zero, y of 0 equal 1, y prime of 0 zero. Transforming and adding gives s squared plus 3 s plus 2, times Y, equals s plus 3. That multiplier is the characteristic polynomial, r replaced by s; solving for Y is division.
Cover up the other factor: A at s equals minus 1 is 2, B at s equals minus 2 is minus 1. So Y is 2 over s plus 1 minus 1 over s plus 2, and the table gives 2 e to the minus t minus e to the minus 2 t. Each pole is one exponential.
Check: y of 1 is 0.6004236 by formula and by RK45 on the ODE. Poles are modes: distinct real poles give e to the r t, repeated ones t e to the r t, complex pairs decaying oscillations, by the table and the shift. Next: initial value problems.
The definition: a probe with a decay rate
For a function f(t) on t≥0, the Laplace transform is
Read it as a numerical recipe, the way you would code a quadrature. Pick a number s, multiply f by the decaying weight e−st, and compute the area under the product. That area is one number, F(s). Different s give different numbers, and the curve through them is the transform. A large s makes the weight die quickly, so F mostly sees the start of f. A small s lets the weight linger, so F sees the long-time behavior. That is the sense in which s is a probe: it chooses how far into the future we look. Nothing here is mysterious: it is a weighted average of f over all future time, computed once per value of s, and the transform is the whole family of those averages. Because we get one number per s, a function of t has become a function of s.
The simplest case is f=1. Then ∫0Te−stdt=(1−e−sT)/s, which tends to 1/s as T→∞ when s>0. At s=2 the area is 0.5. When s≤0, the weight does not decay, the area grows with T, and no limit exists. Whether the limit exists is the first real question.
When the integral exists: exponential order
Say f has exponential ordera if ∣f(t)∣≤Meat for some constants M and all large t. Then for real s,
f(t)e−st≤Me−(s−a)t,
and the integral converges absolutely whenever s>a, by comparison with the integral of e−(s−a)t. For complex s=σ+iτ, ∣e−st∣=e−σt, so the condition is Res=σ>a: the region of convergence is a right half-plane. Everything in this lesson is stated for real s and holds for complex s in that half-plane. If f is also piecewise continuous, F(s) exists there.
Functions of exponential order cover the usual suspects. A polynomial tn has order ε for every ε>0, since tn grows more slowly than any exponential. sinωt and cosωt have order 0 (they are bounded). eat has order a. A function like et2 has no exponential order: it eventually beats eat for every a, and the integral diverges for every s.
The lab below shows both outcomes. The shaded region is the area under f(t)e−st on [0,T], and the lower plot builds F(s) from recorded areas.
◆ InteractiveSlide s and watch the shaded area; press Sweep to build F(s) point by point
Loading simulation…
Take f=et (so a=1) with T=20. At s=2 the area is 1.000: it matches 1/(s−a)=1. At s=1.5, T=12 the area is 1.995 against the table value 2, and the readout’s “tail” of 0.005 is what the window leaves out. At s=1=a the integrand is exactly 1 and the area equals T (it is 20 at T=20 and 40 at T=40): no limit. Below a it is worse. At s=0.5 the area is 804.9 for T=12, 3.255×105 for T=24 and 9.703×108 for T=40, each extra ΔT of window multiplying it by about e(a−s)ΔT (the step from T=12 to 24 is e6≈403). Press Sweep: the recorded points follow 1/(s−1) to the right of the dashed line s=a and run off the top of the plot to its left.
Now pick f=et2. At s=5 the area up to T=4 is 0.2309, which looks like a converged answer. The same integral to T=6 is 60.6 and to T=8 is 2.45×109. The reason is visible in the exponent t2−5t: it dips to its minimum at t=2.5, then climbs without bound. A numerical check that stops early can be fooled, because the divergence lies beyond where you looked. The proof that no s works is the exponent: t2−st→∞ for every fixed s.
Deriving the table
We integrate each transform by hand for real s large enough. Each uses the same step: an antiderivative of e−st times something simple, with the boundary term at ∞ vanishing because s>a.
One.L{1}=1/s for s>0 (above).
Powers. Integrate by parts with u=tn and dv=e−stdt:
∫0∞tne−stdt=[−stne−st]0∞+sn∫0∞tn−1e−stdt.
The boundary term is 0 at both ends (tne−st→0 as t→∞, and tn=0 at t=0 for n≥1). So L{tn}=snL{tn−1}, and unwinding to L{1}=1/s gives
L{tn}=sn+1n!,s>0.
At s=2: L{t}=1/4, L{t2}=2/8=1/4 and L{t3}=6/16=0.375.
Exponential.∫0∞eate−stdt=∫0∞e−(s−a)tdt, so L{eat}=s−a1 for s>a. At a=1, s=3: 0.5.
Sine and cosine. Use eiωt=cosωt+isinωt and the exponential rule with a=iω (legitimate because ∣eiωt∣=1, so Res>0 suffices):
L{eiωt}=s−iω1=s2+ω2s+iω.
The real part is L{cosωt} and the imaginary part is L{sinωt}:
L{cosωt}=s2+ω2s,L{sinωt}=s2+ω2ω.
At ω=2, s=1 these are 1/5=0.2 and 2/5=0.4.
Hyperbolic.sinhωt=(eωt−e−ωt)/2, so by the exponential rule and linearity (next section) L{sinhωt}=s2−ω2ω and L{coshωt}=s2−ω2s for s>ω. At ω=2, s=3: 2/5=0.4 and 3/5=0.6.
f(t)
F(s)
valid for
1
1/s
s>0
tn
n!/sn+1
s>0
eat
1/(s−a)
s>a
sinωt
ω/(s2+ω2)
s>0
cosωt
s/(s2+ω2)
s>0
sinhωt
ω/(s2−ω2)
s>ω
coshωt
s/(s2−ω2)
s>ω
In the lab, choose f=t2 and s=2 (area 0.25), f=sinωt with ω=2, s=1 (area 0.4, from a window of T=40) and f=cosωt (area 0.2). The cell further down checks the other rows by quadrature.
Linearity and the first shifting theorem
The integral is linear, so for constants c1,c2,
L{c1f1+c2f2}=c1F1(s)+c2F2(s),
on the common region where both converge. That was already used for sinh and cosh.
Multiplying f by eat has a clean effect. Since eatf(t)e−st=f(t)e−(s−a)t,
L{eatf(t)}(s)=F(s−a).
This is the first shifting theorem: a factor eat in time slides the transform by a in s. It gives L{eatcosωt}=(s−a)2+ω2s−a and L{teat}=(s−a)21 with no new integral. The lab has both. Choose f=eatcos(ωt), a=1, ω=2, set s=2: the area is 0.2, exactly the area of plain cos2t at s=1. The picture matches the formula: the shaded region is that of the plain cosine, with the weight’s decay rate reduced by a.
Derivatives turn into multiplication
This is the property that makes the whole method work. Let Y(s)=L{y} and integrate by parts with u=e−st, dv=y′dt:
The boundary term at ∞ vanishes because y has exponential order a and we take s>a; the term at 0 is where the initial value enters. Apply the rule twice, once to y′ and once to y:
Why does this turn calculus into algebra? Because the probe e−st is an eigenfunction of d/dt with eigenvalue −s: dtde−st=−se−st. Integration by parts moves the derivative off y and onto the probe, where it costs only a factor s. All that is left of y‘s calculus is the boundary term, and that is exactly the initial data. This is why the initial conditions arrive automatically, no integration constants to fix afterwards.
For a quick numerical sanity check, take y=cos2t, so y′=−2sin2t and y(0)=1. At s=1, L{y′}=−2⋅0.4=−0.8, and sY−y(0)=1⋅0.2−1=−0.8. The two sides agree; the cell below prints them. A second check with no numerics: for y=eat the rule says L{aeat}=s⋅s−a1−1=s−aa, which is a times the table entry, as it must be since y′=ay. The rule is only bookkeeping for what integration by parts already did, so any y of exponential order can be used, including one we know only numerically.
Now put it to work on a constant-coefficient ODE ay′′+by′+cy=g(t) with L{g}=G(s). Transform the whole equation:
(as2+bs+c)Y(s)=G(s)+(as+b)y(0)+ay′(0).
The left coefficient is exactly the characteristic polynomial from 5.2 Constant Coefficients & the Characteristic Equation with r replaced by s. Solving for Y is division. Take y′′+3y′+2y=0, y(0)=1, y′(0)=0. Then (s2+3s+2)Y=s+3, so
Y(s)=s2+3s+2s+3=(s+1)(s+2)s+3.
The denominator vanishes at s=−1 and s=−2, the same roots r as the characteristic equation. The next section turns that observation into the solution.
Inverse transforms by partial fractions
Going back means reading the table from right to left. If F(s) is a sum of table entries, f(t) is the sum of the matching functions. (A transform determines a continuous f uniquely, a standard fact stated here without proof.) For a rational F, partial fractions rewrites it as such a sum.
Distinct real roots. For the example above, write (s+1)(s+2)s+3=s+1A+s+2B. The cover-up rule gives A by evaluating the other factor at s=−1: A=(−1+3)/(−1+2)=2, and B=(−2+3)/(−2+1)=−1. So Y=s+12−s+21 and
y(t)=2e−t−e−2t.
It satisfies y(0)=1 and y′(0)=−2+2=0. Each pole contributes one exponential, with rate equal to the pole’s location.
A repeated root.(s+1)2s+3=s+11+(s+1)22 (write s+3=(s+1)+2). The table with the first shift gives L−1{1/(s+1)2}=te−t, so f(t)=e−t+2te−t. A double pole produces the extra factor t, matching the repeated-root rule of the characteristic equation.
Complex roots.s2+2s+5s+1 has no real roots. Complete the square: s2+2s+5=(s+1)2+4. Then (s+1)2+22s+1 is the cosine entry shifted by a=−1, so f(t)=e−tcos2t. The complex pair s=−1±2i gives a decaying oscillation, once more the same story as the complex characteristic roots.
Computing it
🐍 PythonLab: quadrature against the table, sympy, the e^(t^2) trap, the derivative rule, and the ODE (blocks A to E)
⌘/Ctrl + Enter
Block A reproduces every table entry, block B shows sympy agreeing and giving up on et2, block C is the trap from the lab, block D is the derivative rule (−0.8 on both sides) and block E checks the partial-fraction answer against RK45: y(1)=0.6004236 both ways.
Now build the numerical transform yourself. The exercise uses the trapezoid rule on a finite window, so it also shows what a finite T does to a divergent integral.
🧪 Coding exerciseExercise: a numerical Laplace transform and the derivative rule✓ passed
⌘/Ctrl + Enter
💡 With t = np.linspace(0, T, n) and y = f(t) * np.exp(-s * t), the trapezoid sum is (t[1] - t[0]) * (y.sum() - 0.5 * (y[0] + y[-1])).
With the cell completed, L{sin2t}(1) comes out as 0.4, and et at s=0.5 with T=60 gives 2.137×1013, because 2(e30−1)=2.137×1013.
Concept checks
✅ Concept check
What is the value of $\mathcal L\{te^{2t}\}$ at $s = 4$? (Give a number.)
Why: By the first shifting theorem, L{t} = 1/s^2 becomes 1/(s - 2)^2. At s = 4 that is 1/4.
✅ Concept check
Which of these has no Laplace transform for any $s$?
Why: Exponential order is the condition; e^(t^2) outgrows every e^(at).
✅ Concept check
Why does the transform of $y''$ contain the terms $-sy(0) - y'(0)$?
Why: Each by-parts step contributes one boundary term at t = 0.
Problems
✏️ Pen & paperA table calculation●○○
Find L{3t2−5e−2t+sin4t} and state where it is valid.
Show worked solution
By linearity and the table, s33⋅2−s+25+s2+164. The first term needs s>0, the second s>−2 and the third s>0, so the sum is valid for s>0. At s=1 it is 6−5/3+4/17=4.5686.
✏️ Pen & paperSine from the derivative rule●●○
Let y=sinωt. Use only L{y′′}=s2Y−sy(0)−y′(0) and the fact that y′′=−ω2y to find Y(s).
Show worked solution
y(0)=0 and y′(0)=ω. Transforming y′′=−ω2y gives s2Y−ω=−ω2Y, so (s2+ω2)Y=ω and Y=ω/(s2+ω2). That matches the table entry we derived through complex exponentials, here with no integral at all.
✏️ Pen & paperAn inverse with a complex pair●●●
Find f with F(s)=(s−1)(s2+4)3s+1.
Show worked solution
Write F=s−1A+s2+4Bs+C. Cover-up at s=1: A=(3+1)/(1+4)=4/5. Matching 3s+1=A(s2+4)+(Bs+C)(s−1): the s2 coefficients give A+B=0, so B=−4/5; the constants give 4A−C=1, so C=11/5. Check the s coefficient: C−B=11/5+4/5=3. Then
Check: f(0)=4/5−4/5=0 and f′(0)=4/5+11/5=3, and F≈3/s2 for large s agrees with f(0)=0, f′(0)=3 (the transform of 3t).
Key Takeaways
F(s)=∫0∞f(t)e−stdt is the area under f weighted by a decaying probe; s sets how fast the weight dies.
The integral converges for Res>a when ∣f∣≤Meat (exponential order). Below a the area grows without bound, and the closed-form formula does not equal the integral there.
Functions like et2 have no exponential order and no transform, and a truncated numerical integral can fail to reveal it.
The table comes from integration: 1/s, n!/sn+1, 1/(s−a), and sine and cosine through eiωt; linearity and the first shift F(s−a) extend it.
Integration by parts gives L{y′}=sY−y(0) and L{y′′}=s2Y−sy(0)−y′(0): derivatives become multiplication, and the initial data arrive automatically.
A constant-coefficient ODE becomes (as2+bs+c)Y=data: the characteristic polynomial multiplies Y.
Inverting by partial fractions turns each pole into a mode: distinct real poles to ert, repeated to tert, complex pairs to eσtcosωt and eσtsinωt.
Next, in 6.2 Solving Initial Value Problems, the whole pipeline (transform, solve, invert) is run on initial value problems, including systems.
Finished this lesson?
Mark it complete to track your progress (saved in this browser). Reading every section does this for you.