Applied Mathematics · Section 2
Binomial Series
From a schoolboy's expansion to a universal approximation — and how to catch it in a lie.
Here's an idea so ordinary you met it in school and never gave it a second thought: how do you multiply out $(a+b)$ times itself, a few times over? You learned the answer for whole-number powers — $(a+b)^2$, $(a+b)^3$ — and the coefficients line up into Pascal's pretty triangle. Fine. But then a young Isaac Newton asked the question nobody had dared to: what if the power isn't a whole number at all? What's $(1+x)^{1/2}$, or $(1+x)^{-1}$, written out the same way? The triangle runs off the edge of the page — the sum never stops — and out of that one reckless move tumbles the most-used approximation in all of physics and engineering. It's the reason $\frac12 mv^2$ hides inside Einstein's energy, the reason your graphics card can fake a square root in a heartbeat, the reason a 1% slip in a measurement becomes a 3% slip in a volume. One little expansion, near the number 1, wearing a great many costumes. The last chapter names the pattern; the one before it shows you exactly where it starts to lie.
Chapter One — Theory
1. The Finite Binomial Theorem
Let's start where everybody starts — with an honest whole-number power. You want to multiply $(a+b)$ by itself $n$ times. Now, you could grind it out, but watch what's really going on. When you expand $(a+b)(a+b)\cdots(a+b)$ — $n$ identical brackets — every term in the answer is built by reaching into each bracket and grabbing either an $a$ or a $b$. So a term like $a^{n-k}b^{k}$ shows up exactly as many times as there are ways to choose which $k$ of the $n$ brackets hand over their $b$.
And that count has a name — the binomial coefficient, $\binom{n}{k}$, read "$n$ choose $k$." So the whole expansion is just a bookkeeping of choices:
That's the binomial theorem, and notice the thing that matters most for what's coming: the sum stops. It runs $k=0$ up to $k=n$ and then quits, because you can't choose more $b$'s than you have brackets. A finite power gives a finite expansion — clean, exact, no tail.
Pascal's triangle — the coefficients, for free
You don't even have to compute those factorials. Line the coefficients up row by row and a beautiful little rule appears: each number is the sum of the two sitting just above it.
Why should that be true? Because to choose $k$ brackets out of $n$, fix your eye on the last bracket: either you took its $b$ (now choose $k-1$ from the remaining $n-1$) or you didn't (choose all $k$ from the remaining $n-1$). Add the two cases — and there's Pascal's triangle, every row built from the one above it, no multiplication required. Set $a=1$ and rename $b$ as $x$, and we get the form we'll live in for the rest of the section:
So far, so schoolroom. But that little restriction — "$n$ a whole number" — is about to get torn up, and that's where the whole thing comes alive.
Chapter Two — Theory
2. Newton's Leap — Any Exponent at All
Here's the question that ought to feel illegal. The formula $(1+x)^n=\sum\binom{n}{k}x^k$ leans on $n$ being a counting number — you can't "choose $k$ things from half a thing." But look closer at the coefficient and write it out the long way:
Now that expression doesn't care what $n$ is. It's just $k$ factors on top, counting down, over $k!$. Feed it $n=\tfrac12$, or $n=-1$, or $n=\pi$ — it still hands you a perfectly good number. So Newton made the leap: define the generalized binomial coefficient for any real exponent $\alpha$,
and write down the same series. There's just one catch, and it's the whole story: when $\alpha$ isn't a non-negative integer, those factors on top never hit zero, so the sum never stops. The finite expansion becomes an infinite one:
And here's the price of admission, the condition you must never forget: because it's now an endless sum, it only converges when $|x| < 1$. Push $x$ to the edge and the tail stops dying away. (Contrast the whole-number case, which converges everywhere — precisely because it terminates, so there's no tail to misbehave.)
Two examples worth knowing by heart
Let's actually turn the crank. Take the square root, $\alpha=\tfrac12$. The coefficients are $\binom{1/2}{0}=1$, $\binom{1/2}{1}=\tfrac12$, $\binom{1/2}{2}=\tfrac{(1/2)(-1/2)}{2}=-\tfrac18$, $\binom{1/2}{3}=+\tfrac{1}{16}$, and so on:
Now take $\alpha=-1$ — the reciprocal. The coefficients are $\binom{-1}{k}=\frac{(-1)(-2)\cdots(-k)}{k!}=(-1)^k$, so:
And stop right there — do you recognize it? That's the geometric series from Section 1, the one summing to $\frac{1}{1-r}$, just with $r=-x$. The binomial series with $\alpha=-1$ is the geometric series. It was a special case of this all along. That's not a coincidence; it's the first hint that all these series in the book are relatives, and Section 3 is going to introduce the parent that begets the lot of them.
Chapter Three — Theory
3. The Linear Approximation — the Workhorse
Now look — for an endless series, most of the time you don't want the endless thing at all. You want the first couple of terms, because when $x$ is small the rest of the tail is negligible. And the first two terms of the binomial series are so useful, so absolutely everywhere, that if you remember one formula from this entire section, make it this one:
Here's what it's saying, and why it's so powerful. Take any power law you like — a square root, a reciprocal, a cube, anything — and zoom way in on the point $x=0$ (that is, near input $1$). Up close, the curve is indistinguishable from a straight line of slope $\alpha$. Every power law, near 1, is just a line. That's the whole secret: it turns a nasty nonlinear thing into arithmetic you can do in your head.
And when you need to know how the curve bends away from that line, you keep one more term:
The sign of that second coefficient, $\frac{\alpha(\alpha-1)}{2}$, tells you which way the curve cups. For $\alpha=\tfrac12$ it's negative — the square root bends below its tangent. For $\alpha=-1$ it's positive — the reciprocal bends above. Hold onto that; in the finance application it's the difference between a bond trader who gets blindsided and one who doesn't.
Why "small $x$" is the secret ingredient
The reason this is so honest is that the error you make is governed by the first term you threw away, the $x^2$ one. Halve $x$ and you quarter that error. So if $x$ is, say, $0.01$, the linear approximation is good to about a part in ten thousand — you'd need a very fussy instrument to even notice the lie. That single fact, "the dropped term scales like $x^2$," is what makes the whole edifice of physics-by-approximation stand up. Now let's watch it work.
Chapter Four — Theory
4. The Toolkit
Here's everything on one card. Each row is the same idea — Newton's series — wearing a different exponent. Keep it beside you for the applications; it's all you'll need.
| Form | Series / closed form | Condition |
|---|---|---|
| Finite theorem | $\sum_{k=0}^{n}\binom{n}{k}x^k$ | $n\in\{0,1,2,\dots\}$ |
| Generalized series | $\sum_{k=0}^{\infty}\binom{\alpha}{k}x^k$ | $|x|<1$ |
| Linear approx. | $1+\alpha x$ | $|x|\ll1$ |
| Quadratic approx. | $1+\alpha x+\tfrac{\alpha(\alpha-1)}{2}x^2$ | $|x|\ll1$ |
| Square root | $1+\tfrac{x}{2}-\tfrac{x^2}{8}+\cdots$ | $|x|<1$ |
| Inverse root | $1-\tfrac{x}{2}+\tfrac{3x^2}{8}-\cdots$ | $|x|<1$ |
| Reciprocal | $1-x+x^2-\cdots$ (geometric!) | $|x|<1$ |
Chapter Five — Applications
5. Real-World Applications
For each of these, watch for the same fingerprint: a power law, evaluated near 1, replaced by its first one or two terms. That's the move every single time. The hard part is never the algebra — it's spotting that something in disguise is really $(1+\text{small})^{\alpha}$.
Where does $\frac12 mv^2$ hide inside Einstein's energy?
The problem. Einstein says a moving object's total energy is $E=\gamma mc^2$, where $\gamma=\left(1-\frac{v^2}{c^2}\right)^{-1/2}$. That's a strange-looking beast. But for two centuries before him, Newton had told us a moving object carries kinetic energy $\frac12 mv^2$, and that worked beautifully for cannonballs and planets. If Einstein is right, Newton had better fall out of him as a special case — at ordinary speeds. Does it?
The move. Let $x=-\frac{v^2}{c^2}$, which is tiny for anything slower than a rocket, and apply the binomial approximation with $\alpha=-\tfrac12$:
Now multiply back by $mc^2$:
The punchline. Look at what just happened. The first binomial term gives $mc^2$ — Einstein's famous rest energy, the energy locked in mass itself. The second term is $\frac12 mv^2$ — Newton's kinetic energy, exactly, not a penny off. Two hundred years of mechanics was the second term of a binomial expansion the whole time. Newton wasn't wrong; he was reading the first correction and mistaking it for the whole book. And the next term, the $\frac{3}{8}x^2$ one, is the relativistic correction that starts to matter only when $v$ creeps toward $c$. The series literally tells you where Newton's physics ends and Einstein's begins.
How does a chip take a square root without really taking one?
The problem. Square roots are expensive for a processor — far slower than a multiply. Yet graphics, physics engines, and signal code need them constantly, especially $\frac{1}{\sqrt{x}}$ to normalize a vector to unit length. How do you fake one fast?
The move. The binomial series hands you cheap, arithmetic-only stand-ins right out of the toolkit:
Why it's load-bearing. Once you've got the number written as $m^2(1+x)$ with $x$ small — and you always can, by pulling out the nearest perfect square — the root is just $m(1+\tfrac{x}{2})$, a multiply and an add. The legendary "fast inverse square root" hack used a clever bit-trick to get a rough first guess, but the polish step that made it accurate was exactly this binomial linearization. The same two-term idea hides in lens-maker formulas, in the small-angle correction to a pendulum's period, in the "thin-lens" and "paraxial" approximations all of optics is built on. Whenever an engineer writes "for small $x$," there's a binomial series doing the quiet work underneath.
How much does a bond's price move when rates wobble?
The problem. A bond's price is $P=\frac{F}{(1+y)^n}$ — face value discounted over $n$ years at yield $y$. When the yield jumps by a little $\Delta y$, every trader on Earth needs the new price now, and re-computing that power for a whole portfolio, for every tick, is too slow.
The move. Write the new price as the old one times $\left(\frac{1+y}{1+y+\Delta y}\right)^{n}=\left(1+\frac{\Delta y}{1+y}\right)^{-n}$, and binomial-expand in the small quantity $\frac{\Delta y}{1+y}$:
The punchline. The first binomial term is what traders call duration — the bond's sensitivity to rates, the straight-line estimate. The second term is convexity — the curvature correction. These aren't separate inventions of finance; they are literally the first and second terms of a binomial expansion, given fancy names. And because the price curve bends upward ($\alpha=-n$ makes that second coefficient positive), duration alone always understates how much a bond gains when rates fall and overstates how much it loses when they rise — convexity is a gift to the bondholder. A trader who keeps only the first term is flying with half the instrument panel.
A 1% error in a radius — how big an error in the volume?
The problem. You measure something — a ball bearing's radius, a voltage, a length — and your measurement is off by some small fraction. That quantity then gets raised to a power to compute what you actually care about: a volume goes as $r^3$, a kinetic energy as $v^2$, a period as $\sqrt{L}$. How does the error grow on the way through?
The move. Say the true value is $x_0$ and you're off by a relative amount $\varepsilon$, so you used $x_0(1+\varepsilon)$. Then for $y=x^\alpha$:
The punchline. Relative errors get multiplied by the exponent. A 1% error in a radius becomes a 3% error in a sphere's volume, because $\alpha=3$. A 2% error in a pendulum's length is only a 1% error in its period, because $\alpha=\tfrac12$. This is the entire foundation of error propagation in every lab and every engineering spec sheet — and it's nothing but the binomial linearization, $(1+\varepsilon)^\alpha\approx1+\alpha\varepsilon$, read as a rule about percentages.
Why is it called the binomial distribution?
The problem. Flip a biased coin $n$ times, heads with probability $p$. What's the chance of getting exactly $k$ heads? And why does that bell-ish shape carry the name "binomial"?
The move. The answer is $P(k)=\binom{n}{k}p^k(1-p)^{n-k}$ — and the name is not decoration. Watch what happens if you expand $(\,p + (1-p)\,)^n$ by the finite binomial theorem:
The punchline. Every individual probability is one term of a binomial expansion, and the fact that they sum to 1 is just the statement that $(p+(1-p))^n=1^n=1$. The distribution is the binomial theorem, sorted by how many heads you got. That same trick — packing a sequence of outcomes into the coefficients of a power — is the seed of generating functions, the technique that turns hard counting problems across all of combinatorics and probability into algebra you can differentiate.
How do iterative algorithms get their first foothold?
The problem. Methods like Newton–Raphson polish a guess into a precise answer — but they need a decent starting guess to converge fast, and they need cheap estimates of how a function behaves near a point.
The move. The binomial linearization is the universal "near this point, the function looks like this line" estimate. Need a seed for $\sqrt{S}$? Pull out a power of the base and use $m(1+\tfrac{x}{2})$. Need to renormalize a nearly-unit vector after a small rotation, where its length is $\sqrt{1+\delta}$ with $\delta$ tiny? Use $1+\tfrac{\delta}{2}$ and skip the real root entirely.
Why it matters. In tight loops — shaders running millions of times a frame, embedded controllers with no spare cycles — you cannot afford the exact transcendental function. The binomial first guess, sometimes followed by a single Newton step, gets you to full precision for a fraction of the cost. The series turns "an expensive exact answer" into "a cheap good-enough one," which is the daily bread of numerical computing.
Interactive — Practice
▸ Interactive Lab
Your turn. The two knobs that run the show here are the exponent $\alpha$ — the power you're raising things to — and how many terms of the series you keep. Nudge them and watch the approximation reach out, hug the true curve near $x=0$, and then peel away. Prefer the keyboard? Focus a slider and use the arrow keys; Home and End jump to the extremes.
Each problem from the book, made live. Drag its slider; the dashed line is the exact answer, and watch how well one or two binomial terms chase it.
Chapter Six — Practice
6. Word Problems
None of these will whisper the words "binomial" or "expansion" — that's your job to spot. Look for the thing that's secretly $(1+\text{small})^{\alpha}$, decide how many terms you actually need, and reach for the toolkit.
A proton in an accelerator moves at $v=0.10c$. (a) Using two binomial terms, by what fraction does its kinetic energy exceed Newton's $\tfrac12 mv^2$? (b) Roughly how fast must it go before that discrepancy reaches 10%?
A pendulum's period is $T=2\pi\sqrt{L/g}$. A clockmaker's pendulum is meant to be $1.000$ m long, but it's manufactured $0.4\%$ too long.
(a) By what percentage is the period off? (b) Over a 24-hour day, how many seconds does the clock gain or lose?
A zero-coupon bond matures in $n=10$ years at a yield of $y=5\%$. Rates rise by $\Delta y = 0.5\%$ (50 basis points).
(a) Estimate the percentage price change using duration alone. (b) Add the convexity term — does it help or hurt the holder, and by how much?
A machinist measures a ball bearing's radius and is confident to within $\pm0.8\%$. The bearings are sold by volume (a sphere, $V=\tfrac43\pi r^3$).
(a) What is the resulting relative uncertainty in the volume? (b) If instead the bearings were sold by surface area ($\propto r^2$), how would that uncertainty change?
Estimate $\sqrt{50}$ by hand, with no calculator, using the binomial series around the nearest perfect square.
(a) Give a one-term and a two-term estimate. (b) The true value is $7.0711\ldots$ — how many decimal places does the two-term answer get right?
A factory line produces parts that are defective $2\%$ of the time, independently. An inspector pulls a sample of $n=5$.
(a) What is the probability that exactly one part is defective? (b) Using the linear approximation $(1-p)^n\approx1-np$, estimate the chance that none are defective, and compare to the exact value.
Solutions
Problem 1 — Proton. With $\beta=0.1$, $\gamma=(1-\beta^2)^{-1/2}\approx 1+\tfrac12\beta^2+\tfrac38\beta^4$. The kinetic energy is $(\gamma-1)mc^2\approx\big(\tfrac12\beta^2+\tfrac38\beta^4\big)mc^2$, while Newton keeps only $\tfrac12\beta^2 mc^2$. (a) The fractional excess is $\frac{3/8\,\beta^4}{1/2\,\beta^2}=\tfrac34\beta^2=\tfrac34(0.01)=0.75\%$. (b) Set $\tfrac34\beta^2=0.10$ → $\beta\approx0.37$, so around $0.37c$.
What breaks this? Past $\beta\approx0.4$ the two-term expansion itself starts to creak — you've left "small $x$" behind, and you need the full $\gamma$, not its first corrections. The series tells you its own expiry date: it's good exactly while the next term is small.
Problem 2 — Pendulum. $T\propto L^{1/2}$, so $\frac{\Delta T}{T}\approx\tfrac12\frac{\Delta L}{L}=\tfrac12(0.4\%)=0.2\%$. (a) The period is $0.2\%$ too long — it runs slow. (b) A day is $86{,}400$ s; $0.2\%$ of that is $\approx 173$ s, so the clock loses about 2.9 minutes a day. A tiny length error, a not-tiny clock error.
What breaks this? This leans on the error being small enough that $\tfrac12$ is the only coefficient that matters. A $0.4\%$ error is firmly in that zone. A $40\%$ error would not be — you'd need the $-\tfrac18 x^2$ term, and eventually the exact root.
Problem 3 — Bond. Modified duration $=\frac{10}{1.05}=9.524$; convexity $=\frac{10\cdot11}{1.05^2}=99.77$. With $\Delta y=0.005$: (a) duration alone gives $-9.524(0.005)=-4.76\%$. (b) convexity adds $+\tfrac12(99.77)(0.005)^2=+0.12\%$, for $-4.64\%$. The convexity term is positive — it softens the loss. (Check: the exact change is $(1.05/1.055)^{10}-1=-4.63\%$, so the two terms nail it.)
What breaks this? For a big rate move — a full percentage point or more — even convexity isn't enough, and the dropped cubic term starts to matter. Two terms are a local truth, not a global one.
Problem 4 — Bearing. (a) Volume $\propto r^3$, so $\frac{\Delta V}{V}\approx3(0.8\%)=2.4\%$. (b) Area $\propto r^2$, so $\frac{\Delta A}{A}\approx2(0.8\%)=1.6\%$. The higher the power, the more your measurement error is amplified — which is why volume specs are always the tightest.
What breaks this? Only the assumption that $0.8\%$ is "small," which it surely is. If your measurement were off by $30\%$, the multiply-by-the-exponent rule would understate the true volume error, because you'd have wandered out of the linear zone.
Problem 5 — Square root of 50. Write $50=49(1+\tfrac{1}{49})$, so $\sqrt{50}=7\sqrt{1+x}$ with $x=\tfrac{1}{49}=0.020408$. (a) One term: $7(1+\tfrac{x}{2})=7(1.010204)=7.07143$. Two terms: $7(1+\tfrac{x}{2}-\tfrac{x^2}{8})=7(1.010152)=7.07106$. (b) The truth is $7.071068$; the two-term answer $7.07106$ is right to five decimal places — by hand. That's the power of making $x$ small first.
What breaks this? If you'd lazily expanded around $36$ instead ($x=\tfrac{14}{36}=0.39$), $x$ wouldn't be small and two terms would be poor. The technique is only as good as how close you start to a perfect square.
Problem 6 — Defects. $p=0.02$, $n=5$. (a) Exactly one defective: $\binom{5}{1}(0.02)^1(0.98)^4=5(0.02)(0.9224)=0.0922$, about 9.2%. (b) None defective: exact $(0.98)^5=0.9039$; linear estimate $1-np=1-5(0.02)=0.90$. The approximation $0.90$ is within a fifth of a percent of the exact $0.9039$ — the dropped $\binom{5}{2}p^2$ term is only $\approx0.004$.
What breaks this? The linear estimate $1-np$ only works while $np$ is small. Push $p$ up or $n$ way up and $1-np$ can even go negative — a "probability" below zero, the model screaming that you've left its territory.
Chapter Six·c — Practice
6c. When the Model Breaks
That tidy approximation has been keeping three promises the whole time, and every clean answer above quietly relied on them. The real skill isn't plugging into $1+\alpha x$ — it's noticing the instant one of these gives way:
- $x$ is genuinely small — so the dropped $x^2$ term really is negligible.
- $|x| < 1$ — so the infinite series converges at all (for non-integer $\alpha$).
- It's really an addition — a power of one thing, not a disguised product or a place where the exact function has no expansion.
A student uses $\sqrt{1+x}\approx1+\tfrac{x}{2}$ to estimate $\sqrt{4}=\sqrt{1+3}$, getting $1+1.5=2.5$. (a) What went wrong? (b) What's the honest fix?
(a) Here $x=3$, which is not small and not even less than 1 — the student marched the linear approximation clean out of the country where it lives. The true answer is $2$; the estimate $2.5$ is off by 25%. The dropped term $-\tfrac18 x^2=-1.125$ is enormous, the same order as the answer itself. Assumption 1 (and 2) failed.
(b) Don't expand around the wrong place. Pull out the nearest perfect square: $\sqrt{4}$ is exact, but for $\sqrt{4.2}$ write $4.2=4(1+0.05)$, so $\sqrt{4.2}=2\sqrt{1+0.05}\approx2(1.025)=2.05$. Range reduction — factoring out the big part to force $x$ small — is the universal repair, and it's exactly what every library $\sqrt{}$ does internally.
To find $\sqrt{30}$, a student writes $30=1+29$ and expands $\sqrt{1+29}$ term by term, expecting it to converge. (a) What happens? (b) Why does the whole-number case never have this trouble?
(a) With $x=29$, every term is bigger than the last — $1, 14.5, -105, \ldots$ — the series diverges, spewing nonsense. The binomial series for non-integer $\alpha$ only converges for $|x|<1$, full stop. The fix is the same as before: $30=25(1+\tfrac{5}{25})=25(1+0.2)$, so $\sqrt{30}=5\sqrt{1.2}$, and now $x=0.2$ behaves.
(b) A whole-number power like $(1+x)^3$ terminates after four terms — there's no infinite tail, so there's nothing to diverge. That's the deep difference: integer exponents give exact finite polynomials valid for all $x$; fractional and negative ones give infinite series chained to $|x|<1$.
A student computes $(-8)^{1/3}=-2$ (true), then tries to reach it via $(1+x)^{1/3}$ and gets tangled. (a) Where's the snag? (b) What does the series actually require?
(a) The binomial series $(1+x)^{1/3}$ is built around a base near $+1$ — it assumes you're sitting just to one side of 1 on the positive number line. There's no small $x$ that makes $1+x$ equal $-8$ while keeping $|x|<1$; you'd need $x=-9$, far outside convergence. Real powers of negative numbers open the door to complex values and branch choices, which the plain real series simply isn't built for.
(b) The series wants a base close to a positive 1. For genuinely negative bases with non-integer powers, you've left the real binomial series behind and need complex analysis — a different machine entirely. (Whole-number powers, of course, are fine: $(-8)$ cubed is just $-512$.)
A risk manager linearizes a portfolio's value $V=(1+x)^{-2}$ (an inverse-square exposure) as $1-2x$ and budgets symmetrically for $x=\pm0.1$. (a) What does the dropped term do? (b) Why is ignoring it dangerous?
(a) The second coefficient is $\frac{(-2)(-3)}{2}=+3$, so $V\approx1-2x+3x^2$ — and that curvature is positive, bending the true value above the straight-line estimate on both sides. At $x=+0.1$ the line says $0.80$ but the truth is $0.826$; at $x=-0.1$ the line says $1.20$ but the truth is $1.235$. The linear model understates the value in both directions.
(b) Because the error isn't symmetric noise — it has a definite sign, set by $\alpha(\alpha-1)$. A trader who keeps only the line systematically misprices, always the same way, and a systematic bias is exactly the kind of error that compounds into a blow-up. This is the binomial cousin of Jensen's inequality: the curve and its tangent are not the same, and which side they sit on is information, not a rounding error.
The takeaway for modelers
| Ask… | If "no"… |
|---|---|
| Is $x$ truly small? | Keep the $x^2$ term, or range-reduce until it is (Probes 1, 4) |
| Is $|x|<1$ for a non-integer power? | The series diverges → factor out the big part first (Probe 2) |
| Is the base near a positive 1? | Negative base + fractional power → complex/branch territory (Probe 3) |
| Does the sign of the curvature matter here? | Linear model is biased one way → keep convexity (Probe 4) |
And here's the consolation: when the approximation breaks, it breaks loudly — a 25% miss, a diverging sum, a value that bends the wrong way. Each blunder is the math leaning over to tell you exactly which promise you broke. Keep one more term, or pull out the big part first, and you're back in business.
Chapter Seven — Close
7. The One Pattern to Remember
Whenever something is a power of (1 + a small quantity), it is, up close, just a straight line of slope $\alpha$: $(1+x)^\alpha\approx1+\alpha x$. Keep one term for the value, two for the curvature ($+\tfrac{\alpha(\alpha-1)}{2}x^2$), and the whole infinite series ($\sum\binom{\alpha}{k}x^k$, valid while $|x|<1$) when you want it exactly.
That's the whole story. The moment you spot a power law sitting near 1 — a relativistic factor, a discounted price, a measurement raised to an exponent, a root you need in a hurry — you don't have to wrestle with the power at all. You linearize, and the nonlinear world goes flat and friendly right under your hands. And that little condition $|x|<1$, far from being a mathematician's fine print, is once again doing double duty as a reality check: it's the line between "this approximation is telling you the truth" and "you've wandered too far from 1, and the model is about to lie." Stay near 1, keep enough terms, and a binomial series will carry you an astonishingly long way.
One exponent α, one small x — and every power law in the world flattens into a line you can hold in your head. Next, in Section 3, the master machine that builds all of these series at once.