Applied Mathematics · Section 4
Exponential Series
From a banker's compound interest to the one function that is its own rate of change.
Here's a function so special that nature reaches for it again and again, and once you see why, you can't unsee it. Imagine something whose rate of growth, at every instant, is set by how much of it there already is — money earning interest on its interest, a colony of cells each splitting on its own clock, a lump of uranium where every atom decays at its own steady chance. Ask what curve that traces and you get one answer, always the same: the exponential. And the reason it's the same answer is a kind of miracle — $e^x$ is the one function that is its own derivative, the only curve whose steepness at every point equals its height. Feed it a real number and it grows or fades; feed it an imaginary one and it stops growing and begins to spin, which is how the same little series that dates a fossil also runs every AC circuit and every radio on Earth. One function, its own shadow, wearing all of those costumes.
Chapter One — Theory
1. Where e Comes From
Let's not start with a mysterious constant handed down from on high. Let's earn it, the way a moneylender did. Put one pound in a bank that pays 100% interest a year. After a year you have £2 — fine. But what if they compound it twice a year, 50% each half? Then you get $(1+\tfrac12)^2 = 2.25$. Quarterly? $(1+\tfrac14)^4 \approx 2.44$. The more often they compound, the more you make — but watch, it doesn't run away. It creeps up to a wall:
That ceiling — the most you can squeeze out of continuous compounding — is the number $e$. It's not arbitrary; it's the natural rate of growth, what you get when growth is perfectly, instantaneously reinvested in itself.
The series falls right out
And here's the lovely part — we already know how to crack open $(1+\tfrac{x}{n})^n$, because that's exactly the binomial theorem from Section 2. Expand it for a general $x$ and let $n$ grow without bound. The $k$-th term is $\binom{n}{k}\frac{x^k}{n^k} = \frac{n(n-1)\cdots(n-k+1)}{n^k}\cdot\frac{x^k}{k!}$, and as $n\to\infty$ that messy fraction of $n$'s marches to exactly $1$. Every term settles onto $\frac{x^k}{k!}$, and the binomial expansion melts into a clean infinite series:
And notice the condition, or rather the lack of one: this converges everywhere. There's no $|x|<1$ wall here, no edge to fall off. Why? Because the denominators are factorials, and $n!$ eventually outruns any power $x^n$ you throw at it, no matter how big $x$ is. The terms are guaranteed to die. That fearlessness about the size of $x$ is what sets this series apart from the geometric and binomial ones — and it's the first sign we're dealing with something unusually robust.
Chapter Two — Theory
2. The Miracle — It Is Its Own Derivative
Now watch what happens when we take the slope of that series, term by term. This is the whole ballgame, so go slowly. Differentiate $e^x = 1 + x + \frac{x^2}{2!} + \frac{x^3}{3!} + \cdots$ with respect to $x$:
Look hard at the right-hand side. The $\frac{2x}{2!}$ became $\frac{x}{1!}$; the $\frac{3x^2}{3!}$ became $\frac{x^2}{2!}$. Every term slid down one slot and landed exactly on the term below it. The series came back to itself, unchanged:
That is the most important sentence in this section. $e^x$ is the function whose height equals its slope at every single point. Where it's tall, it's climbing steeply; where it's small, it barely moves. And that's not a quirky property — it's the reason the exponential shows up wherever a thing grows in proportion to itself. We'll cash that in next chapter.
The other gift: turning sums into products
One more property falls out of the series, and you'll use it constantly. Multiply $e^x$ by $e^y$, gather the terms, and the binomial theorem reassembles them into $e^{x+y}$:
So the exponential turns addition in the exponent into multiplication of values — which is exactly the property the logarithm (Section 5) will run backwards to turn multiplication into addition. They're two sides of one coin, and we're holding the heads side now.
Chapter Three — Theory
3. Growth, Decay, and the Most Common Law in Nature
Here's the payoff for that "own derivative" miracle. Suppose you've got some quantity $y$ whose rate of change is proportional to how much of it there is right now — more of it means faster change. In symbols, $\frac{dy}{dt} = k\,y$. This little equation is, no exaggeration, the single most common differential equation in all of science. And its solution is forced to be an exponential:
Why must it be $e^{kt}$? Because we just proved $e^{kt}$ is the one function that, when you take its slope, hands back a copy of itself times $k$ — which is precisely what the equation demands. Nothing else will do.
Everything hinges on the sign of that one constant $k$:
- If $k>0$, the thing grows — population booming, money compounding, a chain reaction running away. It doubles every $\frac{\ln 2}{k}$ of time, over and over.
- If $k<0$, it decays — a radioactive sample, a charged capacitor bleeding out, a hot cup of coffee cooling. It halves every half-life $t_{1/2} = \frac{\ln 2}{|k|}$.
And there's one number worth knowing in your bones: the time constant $\tau = \frac{1}{|k|}$. After one $\tau$, a decaying quantity falls to $\frac{1}{e}\approx 37\%$ of where it started; a growing one climbs to $e\approx 2.72$ times. That $\tau$ is the natural clock of the process — and notice the echo of Section 1's stability theme: $k<0$ is the regime where things settle down, the exponential's version of $|r|<1$.
Chapter Four — Theory
4. Euler's Bridge — When the Exponential Learns to Spin
Now for the part that, the first time you see it, genuinely doesn't seem allowed. We've been feeding the series real numbers. What if we feed it an imaginary one? Put $i\theta$ in for $x$, where $i=\sqrt{-1}$, and just turn the crank. The powers of $i$ cycle $1, i, -1, -i, 1, \ldots$, so the series splits into two interleaved piles — the real terms and the imaginary terms:
But those two piles are series you'll meet head-on in Section 3 — they're exactly the Taylor series for cosine and sine! So out drops the most celebrated identity in mathematics:
Sit with what this says. A real input makes the exponential grow along a line; an imaginary input makes it rotate around a circle. Same series, same function — the direction you feed it decides whether you get growth or spin. Set $\theta=\pi$ and you get $e^{i\pi} = -1$, the famous $e^{i\pi}+1=0$ tying together $e$, $i$, $\pi$, $1$, and $0$ in five symbols. And the practical payoff is enormous: anything that oscillates — an AC voltage, a radio wave, a vibrating string — can be written as a rotating $e^{i\omega t}$, which is far easier to push through the algebra of calculus than sines and cosines ever were.
The toolkit
Here's the whole section on one card — every face of the one function.
| Identity | Form | Note |
|---|---|---|
| The series | $\displaystyle\sum_{n=0}^{\infty}\frac{x^n}{n!}$ | all $x$ |
| Own derivative | $\frac{d}{dx}e^x = e^x$ | the defining trait |
| Growth / decay | $y_0\,e^{kt}$ | $\tau=1/|k|$ |
| Sum law | $e^{x+y}=e^x e^y$ | add → multiply |
| Euler | $e^{i\theta}=\cos\theta+i\sin\theta$ | rotation |
| Poisson sum | $\displaystyle\sum_{k=0}^{\infty}\frac{\lambda^k}{k!}=e^{\lambda}$ | normalizer |
Chapter Five — Applications
5. Real-World Applications
For every one of these, watch for the same fingerprint: a rate of change proportional to the amount present. The instant you spot that, you already know the answer is $e^{kt}$ — and the only real work left is reading off $k$ and its sign.
How do you read the age of something dead for millennia?
The problem. A bone, a scrap of charcoal, a linen wrapping — long dead, no label. How old? Living things take up carbon-14 from the air at a steady rate while alive, then stop the moment they die, and from then on the C-14 only decays. If you could read how much is left, you could read the time.
The move. Every C-14 atom has a fixed chance per unit time of decaying — rate proportional to the amount present — so the count follows $N(t) = N_0\,e^{-\lambda t}$. The decay rate is pinned by the half-life, $t_{1/2} = \frac{\ln 2}{\lambda} = 5730$ years. Measure the surviving fraction $f = N/N_0$, invert, and read the clock:
The punchline. A sample with a quarter of its original C-14 has been through two half-lives — $11{,}460$ years dead, no guesswork. The whole edifice of radiocarbon dating, the backbone of archaeology and a Nobel Prize, is one decay equation read backwards. The same arithmetic dates rocks (other isotopes, longer half-lives) and tells a pharmacologist when a drug has cleared a patient's blood.
How long until it doubles? The rule of 70
The problem. Money at continuous interest, a bacterial culture, a country's GDP, the transistors on a chip — all grow at a rate proportional to their current size, $A = A_0\,e^{rt}$. People don't think in growth rates, though; they think in "when does it double?"
The move. Set $e^{rt}=2$ and solve: $rt = \ln 2 \approx 0.693$, so the doubling time is $t_2 = \frac{\ln 2}{r} \approx \frac{0.70}{r}$. Write $r$ as a percentage and it becomes the famous shortcut:
The punchline. 7% growth doubles in about 10 years; 2% takes 35. Bankers, demographers, and epidemiologists all carry this one number — 70 — in their heads, and it comes straight out of $\ln 2$. It's why "a couple of percent a year" quietly reshapes the world over a lifetime, and why compounding is the most underestimated force in finance.
How fast does a capacitor charge — and why 63%?
The problem. Flip a switch and a capacitor starts filling through a resistor. The fuller it gets, the smaller the voltage pushing more charge in, so the current tapers off. How does the voltage actually rise, and how long until it's "basically full"?
The move. The charging current is proportional to the remaining gap to full — a rate proportional to an amount — so the gap decays exponentially and the voltage climbs as
The punchline. After one time constant $\tau = RC$, the capacitor reaches $1 - e^{-1} \approx 63\%$ of full; after $5\tau$ it's past 99% — "done," for any engineer. That $\tau = RC$ sets the speed of every timer, filter, and debounce circuit ever built, and the 63% figure is so universal that oscilloscope users eyeball it to read $\tau$ straight off the screen. Discharge runs the mirror image: $e^{-t/\tau}$, down to 37% in one $\tau$.
When will the coffee be cool enough to drink?
The problem. A hot cup sits in a room. It cools fast at first, then slower and slower as it nears room temperature. When does it reach drinkable?
The move. Newton's law of cooling says the rate of heat loss is proportional to the temperature gap between object and room — a rate proportional to an amount again — so the gap decays exponentially:
The punchline. The coffee doesn't cool at a steady rate; it cools proportionally, which is why the last few degrees take forever. The same exponential approach-to-equilibrium governs a building's thermal mass, a CPU's heatsink, and a forensic examiner estimating time of death from body temperature — all the same gap-shrinking-in-proportion-to-itself story.
How many emails arrive in an hour — and how long until the next?
The problem. Random events drizzle in independently at some average rate $\lambda$ — emails, radioactive clicks, customers at a counter, requests at a server. How many do you get in a fixed window, and how long do you wait between them?
The move. The count follows the Poisson distribution, $P(k) = e^{-\lambda}\frac{\lambda^k}{k!}$ — and look where that $e^{-\lambda}$ out front comes from. The probabilities have to sum to 1, and $\sum_{k=0}^{\infty}\frac{\lambda^k}{k!} = e^{\lambda}$ by our series, so the normalizer is forced to be $e^{-\lambda}$. The waiting time between events is exponential, $e^{-\lambda t}$, and memoryless: the wait doesn't care how long you've already waited.
The punchline. The exponential series isn't just decorating this distribution — it's holding it together. Poisson and exponential waiting times are the bedrock of queueing theory, reliability engineering, insurance, and the load models behind every data center. The same $e^{-\lambda}$ that normalizes the count is the series of this whole section, doing structural work.
Why does every engineer write a sine wave as e^{iωt}?
The problem. AC power, radio, audio, light — the world hums with oscillation, and you have to push these signals through circuits and filters that add, scale, differentiate, and delay them. Doing that algebra with $\sin$ and $\cos$ and their trig identities is a slog.
The move. Euler's formula lets you write any oscillation as a rotating exponential, $e^{i\omega t} = \cos\omega t + i\sin\omega t$ — a point going around a circle at frequency $\omega$. Differentiating it just multiplies by $i\omega$; delaying it just multiplies by a phase. Hard calculus becomes easy arithmetic with complex numbers.
The punchline. This "phasor" trick is how every electrical engineer on the planet analyzes AC circuits, and the same $e^{i\omega t}$ building block is the atom of the Fourier transform — the idea that any signal is a sum of pure rotations. And here the stability theme of Section 1 comes full circle: a system's response is built from terms $e^{(a+i\omega)t}$, and it settles iff every $a<0$. Decay is stability; the sign of the real part is the whole story, just as $|r|<1$ was for the geometric series.
Interactive — Practice
▸ Interactive Lab
Your turn. Three knobs: the input $x$, the number of terms $N$ you keep, and the rate $k$. Watch the compounding limit climb to $e^x$, watch a handful of terms snap onto the true curve, and watch the rate $k$ flip a process from runaway growth to gentle decay as it crosses zero. Keyboard works too — focus a slider and nudge with the arrow keys; Home and End jump to the extremes.
Each problem from the book, made live. Drag the slider; the dashed line is the value the exponential law predicts.
Chapter Six — Practice
6. Word Problems
Every one of these hands you a process whose rate rides on its own size — and none of them says "exponential." That's your job: spot the proportional rate, pin down $k$ and its sign, and reach for $e^{kt}$.
A wooden artifact is found with 25% of the carbon-14 a living tree would have. C-14 has a half-life of 5730 years.
(a) How old is the artifact? (b) Another sample reads 10% — older or younger, and roughly how old?
A fund grows continuously at 4% a year.
(a) How long until your money doubles? (b) At 7%, how much faster?
A capacitor charges through a resistor with time constant τ = 2 ms, toward 5 V.
(a) What voltage after 2 ms (one τ)? (b) How long to reach 99% of full?
Coffee at 90°C in a 20°C room cools to 60°C after 10 minutes.
(a) Find the cooling constant $k$. (b) When does it reach a drinkable 50°C?
A help desk gets on average λ = 3 calls per hour, Poisson-distributed.
(a) Probability of no calls in an hour? (b) Probability of exactly 2?
Estimate $\sqrt{e} = e^{0.5}$ by hand from the series.
(a) Use terms up to $x^3$. (b) How many decimals do you trust, and why does the series converge so fast here?
Solutions
Problem 1 — Artifact. (a) 25% is two halvings ($100\to50\to25$), so two half-lives: $2\times 5730 = \textbf{11{,}460}$ years. (b) 10% is more decay, so older: $t = 5730\,\log_2(10) \approx 5730\times 3.32 \approx \textbf{19{,}000}$ years.
What breaks this? The clean answer assumes the atmospheric C-14 level was constant over time and that no carbon leaked in or out of the sample. Both can drift — which is why real radiocarbon dates are calibrated against tree-ring records, and why the method runs out of usable signal past about 50,000 years (≈9 half-lives), where so little C-14 remains it's swamped by noise. (See Probe 2.)
Problem 2 — Doubling. (a) $t = \frac{\ln 2}{0.04} = \frac{0.693}{0.04} \approx \textbf{17.3 years}$ (rule of 70 says $70/4 = 17.5$ — close enough for the back of an envelope). (b) At 7%: $70/7 = \textbf{10 years}$, so nearly twice as fast for not-quite-twice the rate.
What breaks this? Pure exponential growth assumes the rate stays fixed and resources are unlimited. No real fund, population, or technology grows at a constant rate forever — sooner or later something saturates and the curve bends over into an S. (See Probe 1.)
Problem 3 — Capacitor. (a) At one τ, $V = 5(1 - e^{-1}) = 5(0.632) = \textbf{3.16 V}$. (b) Need $1 - e^{-t/\tau} = 0.99$, so $e^{-t/\tau} = 0.01$, $t/\tau = \ln 100 \approx 4.6$ — so about 4.6τ ≈ 9.2 ms (the "$5\tau$" rule of thumb, rounded up).
What breaks this? The tidy $e^{-t/\tau}$ assumes ideal components — a constant $R$ and $C$, no leakage. Real capacitors leak and have parasitic resistance, and at very short times wire inductance matters, so the curve isn't a perfect exponential at the extremes. Good enough for almost everything, but not for the last fraction of a percent.
Problem 4 — Coffee. The gap decays: $(T-20) = 70\,e^{-kt}$. (a) At $t=10$: $40 = 70\,e^{-10k}$, so $e^{-10k} = \tfrac47$, $k = \frac{\ln(7/4)}{10} = \frac{0.560}{10} \approx \textbf{0.056 per min}$. (b) For $T=50$: gap $=30$, so $30 = 70\,e^{-kt}$, $t = \frac{\ln(70/30)}{0.056} = \frac{0.847}{0.056} \approx \textbf{15.1 minutes}$.
What breaks this? Newton's cooling assumes the room stays at a fixed temperature and the object cools uniformly. A big pot with a hot core and cool surface, or a room that warms up, breaks the single-$k$ story — you'd need more than one exponential, or a full heat-equation model.
Problem 5 — Help desk. (a) $P(0) = e^{-3} \approx \textbf{0.0498}$ — about a 5% chance of a quiet hour. (b) $P(2) = e^{-3}\frac{3^2}{2!} = 0.0498\times 4.5 \approx \textbf{0.224}$.
What breaks this? Poisson assumes calls are independent and arrive at a constant average rate. Real call volume spikes at 9 a.m. and dies at lunch, and one outage can trigger a flood of correlated calls — so a single fixed $\lambda$ understates the bursts. (See Probe 4.)
Problem 6 — √e. (a) $1 + 0.5 + 0.125 + 0.0208\overline{3} = \textbf{1.6458}$. (b) The true value is $1.64872$, so you're good to about 3 decimals with four terms. The next term $\frac{0.5^4}{24} \approx 0.0026$ would push it to 1.6484 — the factorial denominator makes each term roughly an order of magnitude smaller, so accuracy piles up fast.
What breaks this? Nothing, here — small positive $x$ is the series' happy place. The trouble starts with large or negative $x$, where you either need many terms or run into cancellation. (See Probes 2 and 3.)
Chapter Six·c — Practice
6c. When the Model Breaks
The exponential is the most robust series in this book — it converges for every input — but the models built on it lean on promises that the world doesn't always keep. The skill is catching the moment one breaks:
- The rate $k$ is constant — the same proportional rate at every instant, forever.
- Growth is truly unbounded — no finite resource, population, or ceiling is being ignored.
- You can compute it cleanly — no overflow for huge inputs, no cancellation for very negative ones.
A startup's users double every month; a slide projects $e^{kt}$ out three years to a billion. (a) What's the unspoken assumption? (b) What model replaces it?
(a) Pure exponential growth assumes the rate $k$ never sags and the pool of new users is infinite. But there are only so many people on Earth — as you saturate the market, each new user gets harder to find, so the effective $k$ slides toward zero. Assumption 2 fails: a real ceiling was ignored.
(b) The fix is the logistic curve, which starts exponential and then bends over into an S as it nears the ceiling $L$: $\frac{dy}{dt} = k\,y\,(1 - y/L)$. Early on, $y/L$ is tiny and it looks exactly like $e^{kt}$; near the top, growth chokes off. This is the very same correction Section 1's outbreak probe needed — the exponential is the honest early story, and a lie about the late one.
A programmer computes $e^{-30}$ by plugging $x=-30$ straight into the series. The terms are huge and alternate in sign; the answer comes out wildly wrong. (a) Why? (b) The fix?
(a) With $x=-30$, individual terms balloon to about $30^{30}/30! \sim 10^{12}$ before the factorials finally win, and they alternate sign. Adding gigantic numbers of opposite sign to land on a tiny answer ($e^{-30}\approx 9\times10^{-14}$) means the digits you care about are drowned by rounding error in the big ones — catastrophic cancellation. The series converges mathematically but betrays you numerically.
(b) Never sum the series at a large negative argument. Compute $e^{+30}$ — all-positive terms, no cancellation — and then take the reciprocal: $e^{-30} = 1/e^{30}$. Same value, clean arithmetic. The math is sound; only the order of operations was foolish. A convergent series is not automatically a stable computation.
A simulation needs $e^{1000}$. The program returns inf. (a) What happened? (b) How do practitioners cope?
(a) $e^{1000} \approx 10^{434}$, which sails past the largest number a standard floating-point format can store (about $10^{308}$). It's not a math error — the value is simply too large to represent, so it overflows to infinity, and any later arithmetic is poisoned.
(b) Work in the log domain (Section 5, right next door). Statisticians and ML engineers almost never multiply probabilities or exponentiate big numbers directly; they add logs instead, and use tricks like the "log-sum-exp" that subtract off the largest exponent first. You carry $\ln(\text{value})$ around and only exponentiate at the very end, once the number is back in a safe range. The exponential and the logarithm are partners precisely so you can dodge each other's overflow.
An analyst fits $e^{kt}$ to last quarter's traffic and extrapolates, but the real rate $k$ was quietly drifting the whole time. (a) Which assumption fails? (b) Is the model worthless?
(a) Assumption 1 — a constant $k$. A single exponential bakes in one fixed proportional rate, but interest rates move, decay environments change, viral coefficients fade as novelty wears off. Fitting one $k$ to a drifting process gives a curve that matches the past and misses the future.
(b) Not worthless — but you must let $k$ breathe. Use a local rate (fit over short windows), or a model where $k$ itself evolves, or simply quote the exponential as a short-horizon approximation with honest error bars. The danger is the quiet kind: the fit looks great in-sample and fails out-of-sample, with no alarm. Question whether $k$ is a constant or the average of a trend before you trust the extrapolation.
A spreadsheet uses $e^{rt}$ for an account that actually pays interest monthly, not continuously. (a) How wrong is it? (b) When is the swap safe?
(a) Continuous compounding $e^{rt}$ is the $n\to\infty$ limit of monthly compounding $(1+r/12)^{12t}$. For a 5% rate the two differ by only a hair — $e^{0.05}\approx 1.05127$ versus $(1+0.05/12)^{12}\approx 1.05116$, about a hundredth of a percent. For modest rates the continuous form is a clean, convenient stand-in.
(b) The swap is safe when $r$ is small and you don't need exactness to the penny. But for a high rate, a short horizon, or a contract where the compounding convention is legally specified, use the actual discrete formula — the limit is an idealization, and the gap, though small, is real and grows with $r$. Know which world your number lives in.
The takeaway for modelers
| Ask… | If "no"… |
|---|---|
| Is the rate $k$ really constant? | Drifting rate → local fit or evolving-$k$ model (Probe 4) |
| Is growth truly unbounded? | A ceiling exists → logistic / saturating model (Probe 1) |
| Is the argument modest in size? | Huge → log domain; very negative → use $1/e^{|x|}$ (Probes 2, 3) |
| Is continuous the right convention? | Discrete compounding specified → use the exact form (Probe 5) |
And the consoling part: an inf, a garbage sum, an S-curve that refuses to keep climbing — each is the model telling you which promise broke. The exponential is so willing to grow that it's your job, not its, to remember where the world says stop.
Chapter Seven — Close
7. The One Pattern to Remember
When a thing's rate of change is proportional to how much of it there is — $\frac{dy}{dt}=k\,y$ — it grows or fades as $y_0\,e^{kt}$, and there is no other possibility, because $e^x$ is the one function equal to its own derivative. Feed it a real exponent and it grows ($k>0$) or settles ($k<0$); feed it an imaginary one and it stops growing and starts to spin, $e^{i\theta}=\cos\theta+i\sin\theta$.
That's the whole story. The moment you spot a rate that rides on its own size — interest, decay, cooling, a chain reaction, a queue filling — you already know the curve is an exponential and the only question is the sign and size of $k$. And that sign carries the same meaning it did for the geometric series: $k<0$ is decay, stability, a thing that settles; $k>0$ is runaway. The series converges everywhere, fearlessly, because the factorials always win — and the same little sum that dates a fossil, fed an imaginary number, turns into the rotation behind every wave in the world.
One rate k — and the same function dates the dead, charges your phone, and, fed an i, spins a radio wave into being.