Binomial theorem

The formula

The core statement, for a whole-number exponent n:

(a + b)ⁿ = Σ (k=0 to n) C(n,k) · a^(n-k) · b^k

Reading: multiply out (a + b) times itself n times, and the term with b^k in it shows up exactly C(n,k) times — no more, no less.

The coefficient itself, "n choose k":

C(n,k) = n! / ( k! · (n-k)! )

Reading: the number of distinct ways to pick k items from a set of n when order does not matter. It is a plain count — always a whole number, never a fraction, even though it is written as a ratio of factorials.

Pascal's rule, the recurrence that builds the triangle:

C(n,k) = C(n-1, k-1) + C(n-1, k)

Reading: every entry is the sum of the two entries directly above it. This is why you can generate the whole coefficient table with addition alone and never touch a factorial.

Two facts that fall straight out and matter on the bench:

Σ (k=0 to n) C(n,k) = 2ⁿ          # every row of the triangle sums to 2ⁿ
C(n,k) = C(n, n-k)                 # the row is symmetric

The row sum says there are 2ⁿ total possible combinations of n yes/no items — the size of the whole outcome space you are slicing up.

Where you meet it

Reliability review, k-of-n redundancy. You have three identical sensors feeding a 2-of-3 voter. Each sensor is up with probability R = 0.9. The chance the system stays up — at least 2 of 3 working — is a binomial sum:

P(≥2 of 3) = C(3,2)·0.9²·0.1 + C(3,3)·0.9³ = 0.243 + 0.729 = 0.972

Three sensors bought you 0.972 out of a 0.90 part. That single line is the entire justification for the voter on the review board's slide.

Acceptance sampling on the test floor. You pull 20 units from a lot, and you want the odds of seeing 0 or 1 defect if the true defect rate is 5%. Each term C(20,k)·0.05^k·0.95^(20-k) is one bar of a binomial distribution. The coefficients are the binomial theorem's coefficients — the distribution is (p + q)ⁿ expanded term by term, with p + q = 1.

Fast approximations in your head. For small x, the first two terms of the expansion are all you need. (1 + x)ⁿ ≈ 1 + n·x. A 2% overspeed on a fan raised to the 5th power in a power-law fit is roughly 1 + 5·0.02 = 1.10, a 10% rise. The exact 1.02⁵ = 1.1041, so the one-line estimate is off by four parts in a thousand — fine for a sanity check before you open the spreadsheet.

How it works

The reason C(n,k) shows up as the coefficient is not magic. When you multiply (a + b) by itself n times, every term in the product is built by choosing either a or b from each of the n factors. A term ends up as a^(n-k)·b^k exactly when you chose b from k of the factors. The number of ways to make that choice is the number of ways to pick which k factors gave up their b — which is C(n,k). The counting problem and the algebra problem are the same problem wearing two hats.

The gotcha people hit first: the binomial distribution assumes the trials are independent and share one fixed probability. The 2-of-3 sensor calc above is only valid if the three sensors fail independently. If all three sit on the same power rail, share a connector, or see the same vibration mode, one root cause takes out two or three at once and the tidy binomial number is optimistic — sometimes wildly so. Common-cause failure is exactly what the binomial math cannot see. That is why reliability engineers add a beta-factor term instead of trusting 0.972 on its own.

The second mistake: reaching for factorials when you should reach for Pascal's rule. C(52,5) is fine, but 52! overflows a 64-bit integer and anything past 170! overflows a double. Compute the coefficient as a running product — C(n,k) = C(n,k-1)·(n-k+1)/k — or build it additively from the triangle. You never need the giant factorials themselves; they cancel.

Limit of validity: the clean finite sum above is for non-negative integer n only. Newton's generalized version lets n be any real number — fractional or negative — but then the expansion becomes an infinite series:

(1 + x)^α = 1 + α·x + α(α-1)/2! · x² + α(α-1)(α-2)/3! · x³ + …

That series only converges for |x| < 1. Engineers use it constantly without naming it — every √(1 + x) ≈ 1 + x/2 and 1/(1 + x) ≈ 1 - x small-signal linearization is the first two terms of a generalized binomial expansion. Push x past 1 and the series diverges; the approximation is not just inaccurate, it is meaningless.

History

The pattern is far older than the name. Around the 3rd or 2nd century BC, the Indian prosodist Pingala worked out an arrangement equivalent to the coefficient triangle while counting the ways to combine long and short syllables in verse ^[1]^[2]. The Persian mathematician al-Karajī (c. 953–1029) is credited with the first real algebraic treatment — a now-lost work that laid out the coefficient table and the expansion, and used an early form of proof by induction to justify it, showing a result for one case and stepping it up to the next ^[3]^[4]. His successor al-Samawʾal wrote it down explicitly in the following century, tables and recurrence and all, crediting al-Karajī ^[2]. Independently, Chinese mathematicians Jia Xian in the 11th century and later Yang Hui documented the same triangle ^[1]^[2].

In Europe the triangle carries Blaise Pascal's name because of his 1654 Traité du triangle arithmétique, where he treated it rigorously and proved the coefficient identity by complete induction ^[1]^[2]. He was not first, but he was thorough, and the name stuck.

The leap that engineers actually lean on came from Isaac Newton. Around 1665, at twenty-two, he found that the expansion works for fractional and negative exponents if you let it run as an infinite series ^[1]^[2]. He never formally published a proof, but he stated it and worked examples in a letter — the epistola prior of 13 June 1676 — sent through Henry Oldenburg of the Royal Society to reach Leibniz ^[1]^[5]. Newton stated the general form without proof in those letters; a proof was not published until John Colson's in 1736 ^[6]. The question of exactly when the infinite series converges was not settled with rigor until Niels Henrik Abel took it up in 1826 ^[7].

Related tools

/tools/battery-config — series/parallel cell arrangements, the same k-of-n counting logic
/tools/led-array — array combinations and how many strings a supply can drive
/tools/parallel-resistors — combining identical elements, a combinatorics neighbor
/tools/resistor-e-series — picking values from a fixed set, a counting problem at heart

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →